After failing to respond to any of the ways I had pointed out in my last post that he was misrepresenting every study he flashed on the screen, Benn Jordan shared a blog post someone else had written called Andy Masley doesn’t understand how sound works. The conclusion of the post is that I don’t actually understand the science of sound, my criticism falls apart, and I’ve become ideologically motivated to defend data centers over seeking the truth.

However, the criticisms are all wrong, confused, and seem to be coming from someone who’s lurching to assume I have bad motives. And Jordan obviously knows that it’s all mistaken, because he works on and thinks about sound way more than me, and I can run circles around it. He’s once again tricking his audience. I’m going to use this post to respond to them individually.

I wasn’t originally going to respond to the blog post, because it’s pretty bad throughout, but since Jordan shared it I do feel compelled to explain why it’s wrong.

Criticism 1: Decibels: intensity vs pressure

The first section is called “He does not seem to know what a decibel is measuring”

Masley claims that “Every 10dB increase represents a 10x increase in sound intensity.” This is technically true, but it is the wrong quantity for the argument he is making. Intensity describes the power per unit area carried by the sound wave through the air. What matters for the biological mechanisms Jordan is describing (pressure on the eardrum, stimulation of vestibular structures, outer hair cell response) is not intensity but pressure. And while a 10dB increase represents a 10× increase in intensity, it represents only a ~3.16× increase in pressure. Reporting the intensity ratio rather than the pressure ratio inflates the apparent safety margin by roughly a factor of 300. By the quantity that actually maps to the effects Jordan is proposing, the gap between residential data center levels and occupational exposure limits is one to two orders of magnitude in pressure, not the five orders of magnitude that Masley’s framing implies. It matters significantly less what the power of sound is at the source here than the pressure change at the point of “impact” for the listener.

You can translate a sound’s decibel level into an intensity or a pressure ratio. They both describe the same thing. If residential data center infrasound is 50 dB below the limits where it would need to be to harm you, that gap is both a ~100,000× intensity ratio and a ~316× pressure ratio. Which one we should look at depends on the situation we’re trying to understand.

It’s true that sound pressure describes the potential for physical harm once the sound reaches your ear. But to understand the amount of energy you need to create that dangerous pressure in the first place, you need to know the intensity.

Intensity matters here because it tells you roughly how much more energy the data center would need to contribute to produce enough sound to be harmful to us. The intensity being 100,000x smaller than it needs to be implies that the data center would need to convert 100,000x as much of the energy flowing through it into sound as it currently does for it to harm us. This is obviously ridiculous!

This is a useful comparison to make to understand how reasonable the situation is. The critic just lurches to assume this doesn’t matter at all:

It matters significantly less what the power of sound is at the source here than the pressure change at the point of “impact” for the listener.

I think for understanding how realistic the chance of a data center harming us with infrasound is, it does in fact help to know that the data center would need to be 100,000x as large to produce adequate levels of infrasound to harm us. The critic assumes that this means I’m acting in bad faith. I disagree! The whole point of my post is to understand whether the source of the sound can produce enough sound to harm us.

Masley conveniently only cited the broad occupational guidance for 8-hour workdays while omitting the exposure limit for workplaces requiring maintained mental concentration, which is a much lower 86dB. OSHA also identifies 85dBA as the occupational limit for sound exposure over an 8-hour period. Using this number for comparison, data centers are only ~4x quieter than occupational standards allow for certain roles.

These figures are for eight hours of exposure. The dose makes the poison. It’s quite absurd to say “this is under the threshold for eight hours of exposure, so it’s surely under the threshold for 24/7 exposure!”

The study notes that the impacts of infrasound may pose greater risk to “pregnant women and adolescents,” possibly justifying an even lower rating of dBG.

I will concede a point here that occupational limits are far above residential limits. However, both remain far above what a data center can produce. The reason I hadn’t gone into detail about this is that as noted in the post itself, these occupational levels (including the advise for pregnant women and adolescents) are from old government advice that traces back to papers that don’t replicate, and that infrasound experts have commented on as not being reliable. I’ve already addressed this criticism, but the author is writing this as if I’m trying to hide it. Again, it’s hard not to read this as a bad faith response hoping that the reader doesn’t actually go through my original post.

Masley handwaves away this last point, but studies have shown that there is a potential link between chronic sound exposure during pregnancy and negative health outcomes. The CDC even tells pregnant women:

Avoid low frequency sounds (noises you feel as a rumble or vibration). Low-frequency noise travels through your body more easily than high-frequency noise. Low-frequency noise can cause changes that could affect your developing baby.

This page never mentions infrasound and only lists volumes of low frequency sound that I’ve agreed could be harmful at infrasonic frequencies. And the harms listed are fetal hearing development, not any of the symptoms Jordan lists. It’s also precautionary and doesn’t cite any evidence. Again, I’m not ignoring anything here.

Criticism 2: Perception thresholds/does infrasound become audible?

The second section is called “He does not understand what a perception threshold is.”

Masley’s central rhetorical move is structurally simple: the perception threshold at 10 Hz is 97 dB; residential infrasound levels are below that; therefore, there is no effect.

This is a non-sequitur.

A perception threshold is the level at which a conscious, reportable sensation occurs. It is the floor below which you cannot perceive something. It is not the level below which biological effects cease. These are different things, and conflating them is not a caution any serious researcher in this field would permit themselves.

The author is correct that these are different. I made that clear in the post itself.

I never once said that “Because infrasound isn’t audible, it can’t affect you.” I’m basing all this on what looks to me like a pretty solid separate scientific consensus that inaudible infrasound doesn’t seem to harm us, so therefore it only becomes harmful when it behaves like normal noise pollution (when we can hear it) or when it’s extremely extremely loud and physically harms us.

My argument isn’t a logical deduction from my blind belief that “you can’t hear it” to “it can’t affect you.” It’s an empirical summary of what the literature on infrasound health effects has found. For about sixty years now, researchers have gone looking for physiological and clinical effects of infrasound at sub-perception levels. Properly blinded exposure studies don’t find them. The studies that claim to find them are generally unblinded, badly controlled, or fail to replicate. The nocebo studies reliably produces the symptoms people attribute to infrasound just by telling subjects that infrasound is present, with no actual infrasound required. That’s the pattern I’m drawing on. “Sub-perception stimuli can in principle have effects” is true in the abstract. When you go looking in this specific place though, you don’t find the effects being claimed.

The author is describing my argument as a logical syllogism and then saying this is a non-sequitur. But this isn’t a syllogism at all, it’s a summary of all the good scientific literature on the problem. If the author (and Jordan) think the literature is wrong, they need to engage with it directly.

In fact, Jordan makes this point, comparing infrasound potentially to ultraviolet light: despite being imperceptible, he argues, these both can cause harm in humans. Masley dismisses this out of whole cloth simply because “this blurs two opposite ends of the audible spectrum.” Because one is infra- and the other is ultra-, it’s apparently not even worth considering the broader point that imperceptible waves can be harmful in some scenarios despite their imperceptibility.

No that’s not what I’m saying. I’m saying that Jordan is suspiciously lurching to the extremely high energy end of the light spectrum when we know that the low energy end (comparable to infrasound) doesn’t have negative impacts on us if we can’t detect its presence.

Maybe all comparisons to light are bad. I was mainly suspicious that Jordan lurched from an invisible harm we obviously know is real (UV damaging us) to infrasound. Maybe this was unfair? Idk.

Here is the extraordinary part. Masley himself cites this directly. He writes: “research suggests that while inner hair cells don’t respond to infrasound, outer hair cells of the cochlea can respond to very low-frequency sound at levels below what we consciously detect, and some studies have reported cochlear or brain responses under near-threshold or below-threshold conditions.” This implies that sub-threshold sound can have an impact on humans.

He then, in the very next sentence, dismisses the implication entirely: “But no one has shown this mechanism produces the long list of real-world symptoms people often blame on environmental infrasound.” He is now using his belief as argument in support of the same belief.

The author has a pattern of accusing me of circular reasoning when we actually just disagree about empirical claims.

• Claim 1: there exists a measurable physiological response to sub-threshold infrasound somewhere in the body.

  • This is true. Outer hair cells respond, as I acknowledge in the post.

• Claim 2: this response produces the specific symptoms Jordan is attributing to data center infrasound. Things like vertigo, heart palpitations, cortisol dysregulation, insomnia, cognitive impairment, and the rest.

  • This is not established at all, and all studies trying to find these symptoms from infrasound have come up with nothing.

Therefore, I conclude that because 2 seems to be wrong, our best bet right now is to assume the outer hair cells that do respond to infrasound don’t have an effect. If they did, why can’t we find it anywhere?

My argument isn’t “ignore the mechanism research.” It’s “the mechanism research does not license the leap from ‘outer hair cells respond’ to ‘residential data center infrasound is making people sick.’ That’s a separate claim that needs its own evidence, and the studies built to test it keep coming up empty.”

I’m not using my conclusion to support my conclusion.

Sub-threshold stimuli produce measurable physiological effects in virtually every sensory system we have studied. This is not controversial.

This sentence is doing a lot of work it hasn’t earned.

Yes, you can find detectable physiological responses to sub-threshold stimuli in lots of sensory systems. The magnitudes of these effects are typically tiny, measured in milliseconds of reaction time or microvolts of neural activity, and detectable only under tightly controlled lab conditions.

The author is using this to imply that sub-threshold infrasound could plausibly cause the symptoms Jordan’s describing. But given the state of the evidence, this is a huge unsupported leap. Making this leap is how you get Wind Turbine Syndrome.

There is a deeper problem here, too. Jordan’s entire argument is about sub-threshold effects. His claim is not “people can hear the data center” or “people can consciously detect the data center.” His claim is “the pressure waves affect biology below the level of conscious perception.” Masley’s rebuttal is “the signal is below the threshold for conscious perception.” Yes, obviously, that is the point.

Masley has spent hours attacking Jordan for claims Jordan never made, because Masley apparently believes “sub-audible” might as well mean “nonexistent.”

Jordan’s first video is literally called What You Can’t Hear Can Hurt You. The whole post is a response to that argument.

My actual claim is that when researchers design experiments to test whether sub-perception infrasound causes the health effects Jordan attributes to it, it doesn’t. That’s a direct response to Jordan. The critic has read “the studies testing Jordan’s claim consistently find no effect” and somehow heard “Jordan isn’t making that claim.” I don’t know what to do with that.

Criticism 3: Confusing “perceptible” with “audible”

The threshold error described above is quantitative: Masley treats the audibility threshold as the floor below which all biological response stops. What I want to flag now is a separate, conceptual error that runs beneath it – one he commits so consistently that I am not sure he realizes he is committing it.

Audibility and perceptibility are not the same thing.

Not only do I realize that, I make it clear in the post!

Throughout Masley’s piece, “inaudible” does the work of “undetectable,” “imperceptible,” and “biologically inert” – three claims that are not equivalent and that require separate evidence. He never makes this distinction. He appears not to know it exists despite citing research which clearly states:

Pure tones become gradually less continuous, the tonal sensation ceases around 20 Hz, and below 10 Hz it is possible to perceive the single cycles of the sound. A sensation of pressure at the eardrums also occurs. The dynamic range of the auditory system decreases with decreasing frequency.

I do make this distinction, I just say they seem to overlap despite being distinct. In my opening part on the science of infrasound, I explain that our experiments seem to find that 1) inaudible infrasound doesn’t have an effect, so 2) the only effects happen when it becomes audible. I also clarify that we can sometimes perceive infrasound as physical pressure on our ears and chest before hearing it, and at that point it becomes dangerous to us.

Consider the vestibular system. The semicircular canals of the inner ear respond to pressure changes and acceleration. They do not produce conscious auditory perception – they produce balance, orientation, and spatial awareness. Low-frequency pressure waves can stimulate vestibular structures at levels far below what the cochlea registers as sound. The sensation, if any, is not heard. It is felt as dizziness, unsteadiness, or nausea. This is not speculative. It is the operating basis of clinical tests like vestibular-evoked myogenic potentials, a standard tool used by ear, nose, and throat doctors.

The study is on very loud stimuli (in the 95-105 dB SPL range) which reliably evoke responses, delivered directly into the ear canal via insert earphones. Below about 75-85 dB SPL you stop getting reliable results, which is why the test uses the levels it does.

Studies have shown a signal can be inaudible and still make you dizzy above the occupationally safe decibel range.

This review doesn’t say what the author is claiming.

The sentence the author seems to be referencing describes "isolated reports" of vestibular symptoms following exposure to wind turbines and rocket engines, phrasing that, in a clinical review, means "case reports exist, no established syndrome." The same paper explicitly states that the phenomenon "is not fully understood" and that "ongoing research continues to study" whether low-level infrasound produces vestibular effects. And the research has since been done, and I cite it throughout my post!

The most telling moment in the piece is one we have already visited. Masley cites Salt & Hullar – a study specifically about a mechanism by which the inner ear detects infrasound without rendering it as conscious sound. The outer hair cells respond; the signal is processed; no auditory percept is generated. This is, precisely, detection without audibility. Masley cites it. Then he just handwaves away the implication because there hasn’t been enough follow-on research on the subject. This would be fine if the headline of his article was “I’m not entirely convinced that infrasound is a problem.” But no, the idea of infrasound being potentially problematic is fake, pseudoscientific, conspiracy theory.

I hand wave it away because it identifies a mechanism that could create an effect, but we’ve tested for that effect over and over and haven’t found it. Anyone claiming that science implies the effect is actually there is in fact doing pseudoscience.

The cash-out is this. Jordan’s entire claim is about perceptible but inaudible effects – pressure waves that the body registers through pathways other than conscious hearing. Masley’s entire rebuttal is that the signal is inaudible. These two writers are not disagreeing about evidence. They are using different definitions of what counts as the phenomenon being studied. And Masley does not appear to realize this, because he appears to believe that “you cannot hear it” and “nothing is happening” are the same sentence despite repeatedly citing studies which state the contrary.

This is so bizarre that I don’t even know how to respond to it. Reading it, I kind of just feel deeply tired. How can you come away from my post and honestly write this?

Criticism 4: Nocebos are unfalsifiable

Masley relies heavily on the work of Crichton & Petrie, a body of research on the nocebo effect: show one group a video emphasizing harm from wind turbine infrasound, show another group a neutral video, expose both to real and sham infrasound. The high-expectancy group reports more symptoms, including during sham exposure. This is a real finding. It belongs in any honest treatment of this research area.

But here is what the Crichton & Petrie studies actually show: high-expectancy groups report more symptoms than low-expectancy groups. They do not show that low-expectancy groups report zero symptoms. They do not show that infrasound has no physical effect at all. They show that expectation modulates symptom reporting.

A study showing that expectation amplifies a signal is not the same as a study showing there is no signal. Masley treats “nocebo effects exist” as synonymous with “nocebo effects mean your research is bunk.” This is not how that works.

This isn’t how I treat the nocebo effect in the post. I very clearly separate studies that try to test for infrasound harms from studies that test for a nocebo effect. I first try to show that studies testing for harms of infrasound below perceptible levels consistently don’t find any, and then talk about how studies show that you can get a pretty decent nocebo effect on infrasound and generate similar symptoms. I don’t say anywhere that the nocebo studies alone prove that infrasound is from the nocebo effect. The nocebo effect would be clearly falsified if people exposed to infrasound who weren’t primed to feel bad had any measurable negative reactions at all. Best I can tell from every serious inquiry that’s looked into this, they don’t. I use the nocebo effect as “this seems very likely to be what’s causing the symptoms once we know they’re not caused by infrasound” rather than “this proves that the symptoms are caused by the nocebo effect.” Again I think this is just a bad faith reading of my post.

Consider what this means for Jordan’s double-blind experiment. Jordan ran a study with a “haunted painting” cover story, specifically to equalize expectation across groups. Masley’s response is to dismiss the entire design by leading with the nocebo literature – in effect assuming a nocebo explanation regardless of what the blinding protocol was. This is unfalsifiable reasoning. If nocebo explains symptoms when the design is bad and nocebo explains symptoms when the design is good, then crying “nocebo effect” is more rhetorical shield than scientific analysis.

I addressed six separate problems with Jordan’s study, only one of them was that he was actively priming each group to feel bad (which does in fact seem relevant!). Again this just flatly misrepresents my post that anyone can read.

Furthermore, Crichton & Petrie show that expectation modulates symptom reporting. They do not show that infrasound has no effect. This is not the “exact opposite” of what Jordan claims. It is, at most, an important qualification to it.

I didn’t use Crichton & Petrie to argue that infrasound has no effect, I used them to say that the nocebo effect seems real, and since I had already used other studies to argue that infrasound has no effect, this looks like the next best explanation.

Criticism 5: The evolutionary argument

Natural infrasound – from thunderstorms, ocean microbaroms, atmospheric gravity waves, wind over terrain – is intermittent, spectrally broad, and non-stationary. It varies by season, geography, weather. It is not a constant feature of any individual’s sensory environment.

Industrial infrasound from a data center HVAC plant is continuous, tonally pure at specific frequencies, and temporally consistent.

It’s correct that natural infrasound is intermittent and broadband while modern industrial infrasound is continuous and concentrated at specific frequencies. But the point about evolution I’m making doesn’t depend on the sounds being identical, just whether human bodies have any reason to have evolved a generalized harm response to inaudible low frequency pressure waves. The biology just doesn’t seem to be there. If we evolved this way, people would mostly not live near the ocean and would be regularly sick if they did. We don’t actually observe any evidence that humans or other animals living near the ocean were evolutionarily selected against. And again, we’re just not getting any evidence that we evolved this way from any of the many good studies done on infrasound.

Arguing that humans must be adapted to industrial continuous-tone infrasound because we evolved alongside transient broadband natural infrasound is like arguing that humans must be adapted to staring at a 500-nit monitor six inches from our faces for twelve hours a day because the sun has always existed.

The sun has always been bright. This does not make a welder’s arc safe to look at without goggles.

The difference is that with these is that we evolved to be able to see the light a monitor gives off, and we evolved to see the light a welder’s arc gives off. We did not evolve to see the infrared light either gives off, but we did evolve to physically feel that infrared light when it gets to be so much that it’s dangerous to us. Similarly, we can detect infrasound when it’s starting to become dangerous to us. Thanks evolution! Have you ever worried about the infrared light your monitor gives off? That seems like the comparable case here, not visible light.

Refrigerators: typically ranging 40-50 dB at around 400-500 Hz (not infrasound!) Washing machines: ... the sound they emit is between 100-1000Hz (again, not infrasound!) Trains: ... these emit sound all over the audible range.

Refrigerators, HVAC systems, washing machines, and most rotating mechanical equipment do produce infrasound alongside their audible sound?

We’re surrounded by mechanical sources producing low-frequency and infrasonic energy alongside audible sound, and the population is not experiencing the symptom cluster being attributed to data center infrasound. This is worth talking about and bringing up.

Criticism 6: Am I just a stooge for the data centers?

An honest skeptic could say many things about Jordan’s video. They could say his Finnish citation is misleading. They could say the heart-contraction paper he relies on doesn’t replicate. They could say his haunted painting experiment has design problems. They could say the audible noise pollution story is a well-established mechanism for the symptoms residents report, and that Jordan hasn’t demonstrated the effects of infrasound rise above this level. All of these would be defensible.

What an honest skeptic could not say is that Jordan’s sources, as a body, say the “exact opposite” of what Jordan claims. That is a specific claim about the literature. It is also false. The sources Jordan cites are, at worst, ambiguous or overreached-from. At best, they say roughly what Jordan says they say. They do not, as Masley has repeatedly asserted, diametrically contradict the video they are supposedly citations for.

I’m not convinced by any of the author’s prior points here, so I maintain that Jordan’s sources either say the opposite of what he claims (that infrasound doesn’t cause harm) or are completely unrelated to what he’s saying (that audible noise pollution causes harm, with no mention of infrasound). I would on Twitter sometimes use some hyperbole and blur these together into “all his sources say the opposite of what he claims” so I guess the critic’s right that when I do that, I’m wrong.

Masley has staked out a position that effectively equates to “data centers are an unmitigated good.” He writes in his response to Jordan:

✅ Data centers don’t waste water.

✅ AI artwork is without victims.

✅ Using ChatGPT is not bad for the environment.

His first statement is pretty clearly correct at this point. He’s generally right on item 3, although it is still difficult to find adequate clean power for data centers. But Masley has painted himself into such an ideological corner that he can’t even recognize the ethical issues posed by AI art and the artists whose work on which it is trained without compensation.

The iPhone caused a massive drop in the number of bank teller jobs in the US. I think it’s both true that this was an unfortunate consequence of the iPhone, but also that the bank tellers weren’t “victims of the iPhone.” Maybe this is a terminological difference, I don’t know, but I think it’s reasonable enough to not think it means I’ve been blinded by ideology. Similarly, as I say in the art piece, I think the correct opinion of AI art is that it’s as if someone went to a museum, looked at a lot of other people’s paintings, and then went and made paintings inspired by them but distinct enough that they don’t breach copyright. When I was young I used to draw a lot of Tolkien-like maps. They were different enough that if I sold them, I wouldn’t be violating copyright, even though I learned them basically exclusively from looking at maps at the beginning of Lord of the Rings books. I don’t think it’s a sign that I’ve painted myself into an ideological corner to say that I wasn’t harming the Tolkien estate here by doing this, and that AI doesn’t steal from artists by doing similar things.

Have I staked out a position that data centers are an unmitigated good? Also no. In the piece, I say xAI Colossus was bad and shouldn’t have been built, that the crypto mine data center in Texas was clearly really bad, and that noise pollution issues with data centers are very real. Leading up to this I was also posting a lot about how pollution and climate change are the other real issues I’m worried about. Again, this author doesn’t seem interested in what I actually wrote, he’s just lurching toward these big claims to denounce me.

Why did Benn Jordan share this?

I normally wouldn’t reply to criticism this bad, but I felt obligated to after Jordan elevated it in the discourse and his followers seemed to agree that it showed I was wrong. Again, he’s tricking his followers. Why, if this is so easy for me to dunk on despite not thinking about sound much day to day, did Jordan share it as if it’s all correct, despite working on sound all the time? I think it’s because he’s again being dishonest here and knows who’s right, and is just looking for another way to dodge the obvious fact that none of his sources said what he claimed.

Flashing my credentials

I don’t normally like to do this, but I’ll point out that of the three of us, I’m the one with the physics degree. The author seems to have a degree in architectural technology and design. I assume Jordan knows a lot more than me about audio engineering, but I’d kindly request that if you see a title that says Andy Masley doesn’t understand how sound works you treat it with some skepticism. I have a degree in this stuff and taught it for 7 years! If you look up “IB Physics” in YouTube, one of the first things you’ll see is a picture of my face. I claim that I do in fact understand how sound works.

A music rec

My favorite piece of low frequency music is Monoliths and Dimensions by Sunn O))). A top 10 album for me. I ended up listening to it a lot writing this and the last post. Check it out.

• Intro

• What is infrasound?

• What the science says about infrasound harms

  • The history of infrasound science

  • The consensus

• Benn Jordan’s infrasound videos are a complete disaster

  • First video - Infrasound: What You Can’t Hear CAN Hurt You

  • Second video - Datacenters Behaving Like Acoustic Weapons

• Responses to this piece from the authors of the studies Jordan cites

• Jordan’s response to this article was pretty goofy and reaffirmed that he can’t defend any of his misleading citations

  • Bluesky back-and-forth

  • Jordan’s blog post rebuttal of this post

  • Response to “Andy Masley doesn’t understand how sound works”

• There’s a lot of social permission to treat data centers as boogeymen right now

Intro

There is a fast-spreading new idea that data centers cause a unique harm: “infrasound.” Environmentalists may remember this as a pseudoscientific idea promoted by some fossil fuel interests to scare people about wind turbines in the 2010s. Infrasound is now the topic of what looks like the single most popular piece of media made about data centers in 2026 so far: the video Datacenters Behaving Like Accoustic Weapons by the popular YouTuber Benn Jordan. Jordan’s previous video on infrasound is one of the first things Gemini recommends when you search it. These two videos are each a complete moment-to-moment disaster, and have received no meaningful pushback, so I will be the first. Even if you have no interest in either data centers or infrasound, I think this is an amazing sociological example of how highbrow misinformation is developed and rapidly spread. I found the experience of unpacking it jaw-dropping. After writing this Jordan accused me of being a paid shill, but you don’t actually have to trust anything I’m saying. All you have to do is trust the authors of the studies Jordan himself is citing everywhere.

This threatens to become yet another bizarre folk theory about data centers that educated people believe and give you funny looks if you say it’s fake, like that they heat the areas around them by huge amounts, they use a whole bottle of water per prompt, they seriously harm water supplies, or they have an outsized effect on national electricity prices.

I’ll argue here that all our best current evidence about the harms of inaudible infrasound implies that they’re fake, including recent popular claims about data centers. To be clear, data center noise pollution issues are very real. But “infrasound” is something very different, and is being smuggled into the real noise issues by pseudoscience enthusiasts to attribute almost mystical powers to data centers: inaudible sounds they give off that you can’t even hear can physically harm you and make you sick and mentally unwell.

Jordan’s video is a perfect case study of highbrow misinformation. The host is cool. He opens with the nuance that his video is hosted in a data center, and then immediately jumps into a pretty extreme claim:

1:28 - To put it simply, regardless of where your ethical, environmental, political, or economic interests are, if a data center is being built nearby your home, you’re generally kind of f*****.

Some wild claims have already been smuggled in, but by adding some points about how he’s also using a data center to host the video, they also seem like well-considered takes by a guy who’s willing to acknowledge trade-offs and nuance.

The video looks rigorous and technical, and is presented with the professionalism and chill vibes of the kind of explainer you might see from Hank Green and other trustworthy sources. He surrounds himself with equipment and images that look scientific. He talks to people who have been harmed by regular noise pollution, and shares their stories. A lot of legit-seeming studies are flashed on the screen. A big powerful bad guy is being taken down by a chill local dude. Why think about the individual claims too much?

Despite all this, the video is actually promoting a completely unfounded conspiracy theory, and every single study he uses to justify his conclusions either explicitly says the opposite of what he claims or has such terrible methodology that it’s useless. This is effectively a high-status Alex Jones video, it’s really uniquely terrible, but it’s also flying under the radar of most educated people because it looks and feels reasonable.

I'll show thoroughly and conclusively that the video’s wrong. Before I do that, I need to explain what infrasound is and the state of the science.

What is infrasound?

Sound is vibrating air particles. A sound’s frequency is how frequently the air particles move back and forth. A frequency of 3 Hertz (Hz) means the air particles move back and forth 3 times per second. The particles in this gif complete 1/2 of a full back-and-forth oscillation per second, so their frequency is 0.5 Hz.

The higher the frequency of a sound wave, the shorter its wavelength. We hear higher frequency sounds as higher-pitched. For more on the basic physics of waves you can watch my old physics videos.

Frequency alone doesn’t tell you how much energy a sound wave carries. The energy of sound in air, which we experience as volume, depends mainly on the pressure amplitude of the sound wave. Pressure amplitude is like the physical height of a wave in the ocean. If you were standing in the ocean, waves hitting you very frequently could cause you to tip over, but they would have less impact if each were small, and way more if they were really large, so their height matters a lot too. Sound works similarly: frequency and amplitude are different variables, and the volume we hear is determined by the combination of the two.

Infrasound is sound at very low frequency, below about 20 Hz. Human hearing sensitivity drops sharply in that range, so very low-frequency sounds usually have to be very loud before we can hear or otherwise perceive them. Infrasound is not all sound that isn’t loud enough to regularly hear. There are also sounds with frequencies too high for us to hear: ultrasounds. The prefix infra means “below” (“infrastructure” is structure metaphorically or literally below society) and ultra means “beyond” or “above.” This is loosely like the light spectrum: infrared light is below the lowest frequencies we can see (red), and ultraviolet light is above the highest (violet).

Ultraviolet light is dangerous because it carries more energy than visible light, high enough for electrons in atoms in our DNA to absorb and get knocked loose, which is why UV from the sun can cause cancer.

Infrared light carries less energy per photon than visible light, not enough to eject electrons from our DNA the way UV can. But infrared light can still transfer heat. The warmth you feel standing near a fire is from infrared light, and too much can burn you.

Low-frequency infrasound generally diffracts around or passes through obstacles more easily than higher-frequency sound, which can make it harder to block completely. It can travel much farther and through much more material than normal sound.

At high enough intensity, we may detect infrasound in ways other than hearing it. It can create pressure, fullness in the ears, or a pulsing or rumbling sensation. Infrared at high intensity can create warmth. There isn’t a crisp moment where it suddenly becomes detectable. The boundaries are fuzzy. The lower the frequency, the louder the infrasound needs to be before we can detect it.

For ordinary environmental infrasound, current evidence does not support major direct health effects below where we can hear and perceive it.

Sound pressure levels are measured in decibels, or dB, which is a logarithmic scale. Every 10 dB increase represents a 10x increase in sound intensity, every 20 dB increase is a 100x increase, a 60 dB increase is a million, and so on. The gap between 50 dB and 60 dB is a factor of ten in intensity. The gap between 50 dB and 100 dB is a factor of a hundred thousand.

Sound is measured this way because the range of real-world sound intensities is so huge that a linear scale wouldn’t work. A jet engine at takeoff is a trillion times more intense than the quietest sound a human can detect. To avoid having to write huge numbers out every time we want to measure sound, we just use a log scale. This is important for understanding infrasound because differences in dB can seem small, but actually represent huge orders of magnitudes in difference. The difference between a loud data center and enough infrasound to physical hear (maybe 60 dB vs 110 dB) may not seem like that much. On the scale it’s just a doubling. But in reality every 10 additional dB multiplies the power required to produce the sound by 10. So a difference of 50 multiplies the power required by 10x10x10x10x10, = 100,000 times as much power. People sometimes try to imply that infrasound could be dangerous if it just got a little higher on the decibel scale, ignoring that this would require it to get at least thousands of times as much power from its source.

Infrasound at 20 Hz needs to be at roughly 79 dB for us to detect it, about the level of a vacuum cleaner running a few feet away. A frequency of 10 Hz needs to be at about 97 dB, like a motorcycle at close range or a subway passing on the platform. A frequency of 5 Hz needs to be at about 107 dB, like a chainsaw at arm’s length.

Given the logarithmic scale, these thresholds are nearly three orders of magnitude apart. 5 Hz sound needs about 600 times as much power as 20 Hz sound to become audible to us.

It’s important to not confuse the effects of infrasound with ultrasound. Sometimes people talk as if the harms of all inaudible sound are interchangeable, but this blurs two opposite ends of the audible spectrum. It’s as confused as mixing up infrared light with ultraviolet light. The first can heat you up near a fire, the second can knock out atoms from DNA and give you cancer.

What are the ways infrasound could plausibly harm us? There are at least three:

1. Direct mechanical damage at very high intensities. Extremely intense low-frequency sound can cause discomfort, and injury at high enough levels. A sense of pressure in the middle ear starts showing up at roughly 127 dB of infrasound, like a jackhammer at arm’s length. Chest wall vibration, a gagging sensation, and visual field vibration occur at 145 dB, approaching the sound of a shotgun fired a few feet from your head. Abdominal wall vibration happens at 150 dB, which is the sound of a jet taking off at 25 meters away. This is loud enough to rupture an eardrum. These are not levels that occur in any normal residential environment. And infrasound is not “invisible” at these extremes. In any study, participants can tell when infrasound at this volume is turned on and off. They can either hear it, or feel these physical effects in the pressure of the air around them and know that something’s wrong.

2. Low-frequency rumbles that we can physically feel causing annoyance and sleep disruption. When infrasound is loud enough to detect, or when the thing producing it is also producing audible sound, it can affect us in the same way ordinary noise pollution affects us. It can mess with our sleep and stress response. In this case, infrasound is not really “invisible” anymore.

3. Vestibular or chochlear detection The inner ear contains structures involved in hearing, motion, and orientation. Research suggests that while inner hair cells don’t respond to infrasound, outer hair cells of the cochlea can respond to very low-frequency sound at levels below what we consciously detect, and some studies have reported cochlear or brain responses under near-threshold or below-threshold conditions. But no one has shown this mechanism produces the long list of real-world symptoms people often blame on environmental infrasound.

One last thing to note is that many things in our lives generate infrasound, and infrasound can travel so far that our environment contains sound created by things far away. Refrigerators, HVAC systems, distant highway traffic, wind blowing past a building, ocean waves, thunderstorms, washing machines, trains, airplanes, and even the swaying of tall buildings can generate them. Cities are constantly full of low-frequency sound energy. But the levels vary enormously by source, distance, frequency, and how you measure them. Nature contains a lot of infrasound as well. It would be a little strange if we evolved to feel bad when infrasound is present, because the Earth has always had a lot of it, in the same way Earth has a lot of sources of infrared light.

What the science says about infrasound harms

The history of infrasound science

People have been interested in infrasound for a very long time, but the first really rigorous work by modern scientific standards came out of NASA and the Air Force in the 60s. They were trying to simulate rocket launch conditions, and exposed volunteers to very loud infrasound at around 140 dB. At this extreme volume, the participants experienced ear pain, headaches, and nausea.

Geoff Leventhall is one of the most-cited researchers in the field. He’s spent a lot of time correcting public misunderstandings. A big complaint of his is people often read the NASA-style high-exposure studies as if they applied to ordinary environmental levels of infrasound, but in reality the bad symptoms from this study only show up at extreme volumes around 140 dB. The level of infrasound in people’s homes, even in pretty loud areas, is about a million times less powerful than the rocket launch levels these studies measured.

There’s a huge literature of good studies and reviews on infrasound since then, continuing into the present.

Marshall et al. 2023 is one of the most important. They took 37 noise-sensitive adults and had them spend three 72-hour periods in a sleep lab set up like a studio apartment. They were exposed to simulated wind turbine infrasound, sham infrasound, or traffic noise. Researchers monitored sleep physiology like brain activity, breathing, heart rate, muscle movement, as well as a lot of psychological, cognitive, cardiovascular, and self-report measures. Infrasound did not change the physiological or psychological outcomes tested, while traffic noise in comparison disrupted sleep.

The role of expectation itself, the “nocebo effect,” has also been tested directly. The most famous work is by Fiona Crichton and Keith Petrie at the University of Auckland in the early 2010s. They showed one group a video emphasizing harm from wind-turbine infrasound and showed another group a neutral or reassuring video. Then they exposed participants to both real and sham infrasound. The high-expectancy group reported larger symptom increases, including during sham exposure. The lower-expectancy group reported fewer changes. Most studies on this seem to imply that infrasound effects are nocebos.

Lots of public health bodies have done their own reviews, including the Australian NHMRC, Health Canada, the Massachusetts Department of Public Health, the Netherlands’ RIVM, and others. They’ve all concluded there’s no convincing evidence for a distinct health problem caused by ordinary environmental infrasound near or below the perception threshold.

There was also one huge Health Canada study of 1,238 households living near roughly 400 wind turbines in southern Ontario and Prince Edward Island, which found no relationship between wind turbine noise exposure (including infrasound) and hypertension, migraines, tinnitus, sleep disorders, or chronic disease. People who could physically hear the turbines at audible levels were annoyed.

One influential mechanistic paper often cited by people who think infrasound causes harm is “Responses of the ear to low frequency sounds, infrasound and wind turbines.” The cochlea is the organ in your inner ear that detects sound. It has two types of hair cells: inner and outer. Inner hair cells detect most of what we hear, and they turn out to be pretty insensitive to infrasound. Outer hair cells have a more complicated job, mainly amplifying certain sounds. A review by Salt and Hullar argued that these outer cells can respond to infrasound at levels below where a person would consciously register the sound. So it looks like there’s a possible pathway for infrasound to influence us, but no one’s produced any convincing evidence that it actually causes physiological or psychological effects outside the nocebo effect.

One other key fact: a lot of the good scientific studies of infrasound were specifically designed to look at wind turbines. In the 2010s there was a panic about “Wind Turbine Syndrome,” a long list of symptoms purportedly caused by wind turbine noise, pushed by an unholy alliance of international cranks, local NIMBYs, and some fossil-fuel interests. The whole thing originated in a single incredibly stupid book, written and published by an anti-wind couple who among other things compared wind turbines to Jim Crow. A lot of scientists knew this was pseudoscience and wanted to formally debunk it. Simon Chapman is the great public debunker of Wind Turbine Syndrome, so I refer you to his free book if you’d like the full history.

The consensus

Contrary to Benn Jordan’s video, where he claims infrasound is “grossly understudied,” infrasound and low-frequency noise have been studied for decades by a ton of reputable scientists and major health organizations. Controlled studies and public health reviews haven’t found a real, distinct health condition caused by ordinary environmental infrasound at the levels we actually experience, even in noise polluted areas. When low-frequency sound becomes strong enough to be heard or otherwise felt, it can cause annoyance, discomfort, and sleep disruption like any other normal noise pollution. At extremely high levels, it can physically harm us in the way other sound can. We can always detect it when it’s at this level.

The reported symptoms of inaudible, undetectable infrasound appear to be a nocebo. If you tell people infrasound is going to harm them but don’t actually expose them to any, they develop symptoms. If you don’t tell them, and play infrasound without telling them, they don’t develop any measurable symptoms.

Benn Jordan’s infrasound videos are a complete disaster

Benn Jordan has made 3 videos on infrasound:

The top one is his most viewed video of all time. I’m going to skip this one, because its topic is a little different: audible low-frequency sounds (sometimes called “the hum”) that some people can hear and find disturbing in the way other unpleasant sounds are disturbing. That falls under traditional noise pollution. And while I seriously doubt these sounds cause “psychosis” or “death,” the effects of traditional noise pollution are outside the scope of this post.

I’ll first unpack his original video on infrasound that’s inaudible to everyone, and then talk about the data center video. I’ll use quotes and timestamps for both.

First video - Infrasound: What You Can’t Hear CAN Hurt You

I would recommend watching a bit of this video before we dive in. It’s very easy to watch without noticing anything wrong. It’s very well-produced, and I think if you have the experience of watching it before reading what follows, this post will be much more jarring in a fun way.

0:00-1:40 - Intro

The intro to this video is already silly. He records infrasound in his home and then compresses and speeds it up until he can hear it.

0:24 - Whoa. I’ve done a whole lot of investigating and I still have absolutely no idea what’s causing that sound and it’s still going on right now as I’m recording this.

Lots of things in our lives make infrasound. It’s easy to record any of it, compress it, speed it up until it’s audible, and it always sounds alien and weird when you do. You can do this with infrasound from a fridge or HVAC or distant traffic. If you compress a whole day of it into something that’s audible, you’re going to get eerie sounds from far away. This tells you nothing on its own.

There’s a lot of infrared light everywhere too. I could aim an infrared camera at a dog, and then say “Whoa… something’s wrong with my dog. He looks Evil”

and this would tell us about as much as making infrasound audible.

0:32 - And if you’re in a city or suburb, chances are that similar sounds are emanating just below your hearing range.

Yes, most things we do and use create infrasound. This is like starting a video with a scary infrared image and then saying “And if you’re in a city or suburb, chances are that similar light is emanating just below the range you can see it.”

0:39 - And I have some bad news for you. A meta-analysis suggests that it can actually be pretty harmful to your health.

He flashes three papers on the screen here. He doesn’t link any of them in the video description. None of them support what he’s saying at all, they all go completely against it. None of these are a meta-analysis showing infrasound is harmful.

The first is an occupational measurement paper that implies infrasound is safe at levels hundreds of thousands of times more intense than anything people regularly experience in their homes, even in loud places. The second is part of a 3-part research project where the final report is called “Infrasound Does Not Explain Symptoms Related to Wind Turbines.” The third is a NIOSH workplace report that explicitly concludes the sound levels it measured are not known to cause negative health effects.

The first paper paper is “Exposure to infrasonic noise in agriculture.”

It’s an occupational exposure assessment for Polish farm workers. The author took a sound meter to 118 pieces of farm machinery in Poland and measured the infrasound levels they produce. And that’s all.

It’s not a meta-analysis. The abstract has some speculative language about infrasound having “possible ergonomic and health consequences,” and this is probably what Jordan is pointing at when he flashes the cover. The paper also asserts that “a very important harmful factor is infrasound exposure for pregnant women and adolescents at workplaces in agriculture.” That claim cites older Eastern European occupational health literature that’s never been replicated in modern studies. That older literature comes up over and over in infrasound harm claims. It’s almost all small studies with bad controls, often decades old, that don’t replicate. Most of the serious modern work on infrasound, and the major public health reviews from the last fifteen years, doesn’t find its evidence compelling.

Geoff Leventhall has flagged this pattern in his reviews of the infrasound literature. Governments set a regulatory threshold for infrasound exposure based on cautious extrapolation from very limited data, and then new papers cite that threshold as if it were evidence of harm at that level. Later papers cite those papers. The chain of legit-seeming papers starts with original evidence that either doesn’t exist or doesn’t replicate. This paper is part of that pattern.

The paper measures infrasound an operator would receive directly on a tractor or combine, with the engine and mechanical systems inches away. The measured infrasound levels here are really high. This doesn’t tell us anything about what everyday people experience in their homes. No source of infrasound (just like no source of regular sound) is nearly as loud as sitting on a tractor.

Worse for Jordan’s later video, the paper compares its measurements to the existing Polish occupational exposure limit of 102 dBG for an 8-hour workday. Data center and wind turbine infrasound at residential distances is typically in the 50 to 75 dBG range. That might sound like a small gap, but remember that decibels are logarithmic. Every 10 dB increase represents a 10x increase in sound intensity, and every 20 dB increase represents a 100x increase.

So this is telling us that even the loudest data centers near people’s homes are several hundred to over a hundred thousand times quieter than what the occupational standard the Bilski paper is using considers safe for a full 8-hour workday of direct exposure. The paper Jordan is flashing as evidence of infrasound harm is, if you actually read it, implying that our regular daily infrasound exposure could be multiplied by 1000 and still be totally safe.

The second paper is “Symptoms intuitively associated with wind turbine infrasound” published in Environmental Research.

The paper concludes that infrasound had no detectable effect.

The paper is one of the main outputs of a single research project commissioned by the Finnish Government and conducted by the Technical Research Centre of Finland (VTT), the Finnish Institute of Occupational Health, the University of Helsinki, and the Finnish Institute for Health and Welfare from 2018 to 2020. Its goal was to figure out if wind turbine infrasound actually causes the symptoms some residents attribute to it.

The three outputs of the project are:

1. The paper that Jordan shows. This is only the questionnaire study. Within 2.5 km of a wind turbine, 15% of Finnish respondents said they associated various symptoms with wind turbine infrasound, and the symptoms spanned multiple organ systems. The paper documents what people report and what they believe is causing harm, not whether infrasound actually harmed people.

2. “Annoyance, perception, and physiological effects of wind turbine infrasound.” Here participants were divided into two groups: people who reported the types of symptoms they attributed to wind turbine infrasound, and people who didn’t. Both groups were exposed to real wind turbine sound recordings, with and without the infrasonic components. Neither group could reliably detect when the infrasound was present. It had no effect on reported annoyance, or measurable effect on people’s nervous system responses. There were no differences between the symptomatic and non-symptomatic groups.

3. The full final report of the project, published by the Finnish Government under the title “Infrasound Does Not Explain Symptoms Related to Wind Turbines.” Jordan conveniently does not mention this title.

The project found exactly what every other serious investigation of this question has found. Some residents report symptoms, they intuitively attribute those symptoms to infrasound. When you actually expose them to infrasound in controlled conditions, they can’t detect it, it doesn’t produce the symptoms, and it doesn’t produce any measurable physiological response. Looks like a nocebo.

Jordan only shows the first paper where people claim that infrasound caused their symptoms, and never clarifies that the conclusion of this study was that infrasound wasn’t to blame for whatever was happening.

The third paper is “Evaluation of Low-Frequency Noise, Infrasound, and Health Symptoms at an Administrative Building and Men’s Shelter: A Case Study.”

In 2019, a nonprofit serving homeless people had two incidents of loud noise and strong vibrations coming from an adjacent landfill. Employees reported feeling unwell during and after. The buildings were evacuated. NIOSH was called in to conduct a Health Hazard Evaluation. This is the paper documenting that investigation.

The paper is clear that HHEs are workplace investigations to help employers recognize and control hazards. They are not research studies. This one is a descriptive report published after the fact. Its findings don’t support what Jordan is implying at all.

NIOSH arrived four months after the incidents and took sound measurements during normal flare operation. The overall measured sound levels across frequencies up to 100 Hz were 64 to 73 dB. The American Conference of Governmental Industrial Hygienists threshold limit values for occupational exposure to infrasound and low-frequency noise are 145 dB at specific frequencies and 150 dB overall. The measured levels were roughly 77 to 86 dB below the threshold at which harm is known to occur, at least tens of millions of times less intense, and possibly hundreds of millions of times less intense at the quieter end.

The paper explicitly says “The frequency-specific or overall sound levels we measured are not known to cause adverse health effects.”

The paper proposes explanations for the symptoms employees reported, and none of them involves infrasound damaging the body:

1. People audibly heard and felt two specific bad events and felt bad.

2. The routine sound at the site had an unbalanced frequency profile: more energy at low frequencies than at higher ones. This kind of imbalance is known to cause annoyance complaints even when the overall sound level is low. The paper cites a 2021 review by van Kamp and van den Berg that found, where health effects are observed near low-frequency audible (not infrasound) noise sources, those effects track with how annoyed people are, not with direct exposure to the low-frequency sound itself.

3. A lot of the symptoms people reported (headaches, fatigue, trouble sleeping) are common and nonspecific. When NIOSH looked at the employees’ medical records, a meaningful fraction had plausible alternative explanations for their symptoms that had nothing to do with the flare.

4. Employees felt the organization and the county had mishandled the incidents. The paper frames this as a breach of “psychological contract,” the sense that your workplace has not kept faith with you. That breakdown, the paper argues, shaped both how employees experienced their symptoms and their unwillingness to return to the building.

The paper’s Limitations section acknowledges that no definitive conclusions about the cause of the symptoms can be drawn.

And in a final twist, the main wind turbine study this paper cites is Michaud et al. 2016, the Health Canada study, and summarizes its conclusion: the study authors did not find a relationship between wind turbine noise and reported health effects such as headaches, dizziness, sleep problems, and stress.

So just to wrap, Jordan’s third piece of evidence here is a paper that:

• Measured sound levels well below any known harm threshold.

• Explicitly states those levels are not known to cause adverse health effects.

• Attributes the reported symptoms to annoyance and a breach of psychological contract rather than direct infrasound harm.

• Explicitly refuses to draw causal conclusions.

• Cites Health Canada to say wind turbine noise doesn’t cause the symptoms people attribute to it.

The actual meta-analyses and review papers on infrasound are overwhelmingly consistent: there are no health effects unless the sound is loud enough to be perceived, and at that point it behaves like regular noise pollution. There’s no serious meta-analysis that reaches the conclusion Jordan’s implying.

Why did he just flash three papers that obviously don’t conclude what he says? It’s such a bizarre move that my trust in him as a narrator cratered.

0:46 - Have you ever been to a particular place or maybe walked into a building and immediately felt like something was wrong? Like a rather noticeable and intense feeling of discomfort or maybe even fear despite there not being anything notably wrong with the environment. And then the moment you leave that place, you start feeling better.

Your bullshit detectors should be going off. This is a very common part of the human experience. A lot of places just look ominous, and a lot of people have anxiety.

Look at the images he chooses as examples here:

It’s obvious that these are just generally ominous places! The second is a famous image from /r/liminalspaces, which collects pictures of places that feel impermanent, and where you feel transient and possibly uneasy.

0:59 - Well, it turns out that there’s a chance that it wasn’t all in your imagination or anything paranormal. It could have just been infrasound.

All the thorough research that’s been done strongly implies that infrasound doesn’t cause this at all.

1:25 - We’re going to take a close look at the infrasonic research and see how how some types of these sounds can make people enjoy music less or feel incredibly ill.

He flashes this on the screen as he says this:

This is Chapter 11 of the WHO’s “Compendium of WHO and other UN guidance on health and environment” with the WHO, UN Environment, and UNICEF logos displayed. Flashing this implies WHO-endorsed evidence that infrasound makes people enjoy music less or get incredibly ill.

The document says nothing like this.

This is a policy guidance document compiling recommendations across environmental health topics. Chapter 11 is on guidance for environmental noise, listing recommended levels for road, rail, aircraft, wind-turbine, and leisure noise. It doesn’t discuss infrasound as a distinct health harm.

There is no overlap at all between what Jordan is narrating and what the document on the screen actually contains. He’s just showing a completely unrelated government document with health agencies his viewers trust, while saying completely different false stuff about what the infrasound research implies.

1:31 - And according to one famous study, even see ghostly figures near them who aren’t actually there.

He’s referring to Vic Tandy’s “The Ghost in the Machine” article, which he’ll discuss more later. He shows images from various press articles summarizing Tandy’s work, which tend to describe it as more legitimate than it was.

I’ll save the full overview of this article for when he goes into more detail, but for now just be aware that this “famous study” is not a study at all. It’s a personal anecdote from a single guy who felt uneasy after being told the place he was working was haunted, took a single measurement, decided infrasound was the cause, and published the story in a parapsychology journal founded to investigate ghosts. The study never replicated, and was specifically tested and failed when a proper experiment was finally run. Jordan has once again misrepresented his source.

1:40–3:48 - Philosophy of time

This section has nothing to do with infrasound. It’s about Greek concepts of time, water clocks, and objective versus subjective perception.

I normally give YouTube essayists slack here. They’re making videos for people with short attention spans, and tangents are fine. But this video’s already been so deceptive that I can’t help thinking this section exists to prop up the others. It makes Jordan look worldly and well-rounded, and lends credibility to everything else he’s saying. Showing himself making a DIY sun clock is neat, but I’m worried this is a move to look scientific, like a more hip internet-intellectual way of wearing a lab coat in the video.

There’s another really bad move he’s making here. He’s implying our capitalist western culture has trained us to focus on easy-to-measure effects and to ignore our felt subjective sense of what’s going on. He seems to be using this to open the door to wild claims about things that are invisible to our daily experience but can still harm us. The things he’s about to talk about aren’t actually invisible to the measurement tools of science, just to our senses. His next move is to use this setup for a wild connection between two different invisible things: UV light and infrasound.

3:48–4:34 - UV and infrasound

4:02 - Only when the wavelength of light enters the human visible spectrum do we call it a color. We can’t observe higher wavelengths like ultraviolet, but we can observe the effects of it when it burns our skin or causes cancer. And if you keep cranking up the frequency of invisible light to X-ray and gamma, even lower levels of light can eject electrons from atoms... And for you or I, that means death. A lot of death. It goes without saying that when waves of pressure happen between 20 and 20,000 times per second and become audible to us, do we call it a sound. And while pressure waves above our range of hearing can give us headaches and hearing damage, it fortunately doesn’t do anything as dramatic as rearranging atoms. But medical research is starting to trickle out data about the sounds that are happening below our range of hearing. And things are looking quite a bit more sinister.

As a former physics teacher, this section was torture.

He’s setting up this framing about how things that are invisible to us can still harm us, and then very quickly moves from ‘ultraviolet light can obviously kill us’ to ‘and there’s new disturbing evidence that infrasound can harm us’ without acknowledging the obvious problem that infrared light does not really harm us at all unless it’s so powerful that it heats up our bodies in the way other normal heat sources do. This is the correct analogy to infrasound, not UV light. Just like infrared light can only harm us if it’s powerful enough that we physically feel it in the way we do other heat sources, infrasound can only harm us if it’s so powerful that it causes us to physically feel its presence in the way we do other very loud sources of sound. He’s not so subtly reaching way across this spectrum to say that because something that’s so high energy that we can’t see it can cause damage, maybe something that’s so low energy that we can’t hear it can cause damage too.

In your everyday life you probably don’t spend much time thinking about prices measured in fractions of a cent, or in tens of millions of dollars. Both amounts are “invisible” to your daily experience, but for opposite reasons. What Jordan’s doing is kind of like saying “Because you never think about tens of millions of dollars, other money amounts you don’t think about might also damage your wallet if you spend them, and there’s new evidence that fractions of a cent can harm your wallet too.”

“Things are looking quite a bit more sinister” is doing a lot of work. The medical research is not “trickling out” and is not at all sinister. As covered above, we’ve had decades of solid research, and the conclusion of all serious scientific work on this is incredibly consistent.

4:50–5:21 - The symptom list

4:50 - For example, research has strongly suggested that infrasound can cause headaches, fatigue, loss of concentration, mood changes, depression, sleeping disorders, panic disorders, nausea, dizziness.

The research he cites doesn’t say this at all, but first I need to note that a long list of basically anything and everything bad you might experience in your everyday life is a common signature of a nocebo. Think about how infrasound compares to any other exposure to something bad for us. Carbon monoxide or lead poisoning produces very specific symptoms. Even normal noise pollution has very specific symptoms like hearing loss and cardiovascular effects of chronic sleep disruption. They don’t have these incredibly general long lists of symptoms.

What are the studies he’s flashing on the screen here to support this?

Well, first he puts this map of infrasound science up on the screen:

This looks authoritative and implies a century-long scientific history of studying the harms he’s discussing.

The image is from the 2020 paper “Low-Frequency Noise and Its Main Effects on Human Health — A Review of the Literature between 2016 and 2019,” published in Applied Sciences by four researchers at the University of Minho in Portugal. The timeline is a history of the general low-frequency-noise research field.

But he’s tricked his viewers again. The paper is about low-frequency noise, not infrasound. Those are not the same thing here. Low-frequency noise typically refers to the audible range roughly between 20 and 200 Hz, infrasound is below about 20 Hz and is generally inaudible. The paper uses “low-frequency noise” throughout and only occasionally touches infrasound specifically. So the timeline isn’t even a history of the field Jordan is claiming support from.

The paper doesn’t conclude what he’s implying at all. It’s mainly analyzing audible noise, not infrasound. He’s mixing up audible low-frequency noise (roughly 20–200 Hz, which you can hear) with inaudible infrasound (below about 20 Hz, which you generally can’t). The paper finds plenty of health effects from regular audible noise pollution, like annoyance and sleep disorders, mostly from things like road traffic, railways, the audible part of wind turbine noise, oil and gas drilling, and compressor stations. The paper explicitly excludes inaudible infrasound from its in-depth analysis: in the methodology, “low-frequency sound/infrasound” is listed as one of nine categories in the 142-paper database, but it’s deliberately not among the three categories selected for the 39-paper in-depth review. The paper doesn’t comment on whether infrasound is harmful or not.

The next section is even worse. He flashes and says a ton of individual symptoms described by the paper:

Not one of them is attributed to infrasound by the paper he’s citing, they are literally all caused by audible noise:

• Headaches and nausea come from Blair’s study of oil and gas well construction and drilling in a residential area. The paper is explicit that this was from continuous noise measured at 51.5 to 80.0 dBC, with headaches and nausea reported when levels exceeded 60 dBC. The “C” in dBC refers to C-weighting, a measurement scale for audible sound that captures low-frequency components. 60 dBC is roughly the loudness of normal conversation. The study wasn’t measuring inaudible infrasound, it was measuring the audible din of heavy industrial drilling next to people’s homes.

• Fatigue, lack of concentration, negative mood, and dizziness all come from the same sentence in the Pohl, Gabriel, and Hübner wind turbine study. And here Jordan is doing something more deceptive. These aren’t the results of a study, they’re options from a survey checklist, the list of 12 symptoms the researchers asked residents about. Quoting the paper: “ascertaining the physical and psychological symptoms referenced by residents that participated in the study (general mental indisposition, performance and reduced work capacity, lack of concentration, fatigue, tension, nervousness, negative mood, dizziness, irritability, indisposition, reduced sleep quality, and annoyance).” This is just the questionnaire the surveyors handed out.
  And what did the study actually find when residents were asked? The exact opposite of the picture Jordan is painting. Pohl concluded “the annoyance experienced was very low.” Only 8.5% of residents in 2012 reported any feelings of pressure related to the wind farm, dropping to 6.8% by 2014. Only 6.1% reported felt vibrations in their body, dropping to 3.8%. Symptoms of dizziness, one of the words Jordan highlighted, “were not observed in this study.”

• Anxiety and depression come from Abbasi’s study of wind turbine maintenance workers, who stand on and close to them. The study’s exposure levels were 60, 66, and 83 dBA. This is clearly audible, occupational noise. The paper’s own summary of the findings emphasizes that “the worst health status is due to working conditions and chronic exposure to occupational risk factors, such as noise” and that the harmful effects were concentrated in the maintenance team who were “in the vicinity of wind turbines, due to the reception of very intense noise.” In other words, the study’s own framing is that these effects track with exposure to loud audible noise, not infrasound.

• Sleep disturbance is discussed across several studies the paper reviews, and again the attribution is to audible noise. The paper says directly: “Long-term exposure to low-frequency noise from wind energy is a major factor in sleep disturbances in residents who live near wind farms.” Note the framing: low-frequency noise, not infrasound.

When the study did look at infrasound specifically, it noted that “the noise level of the wind turbine measured in the lower frequency range is below the human sensory threshold” and the paper reports this without treating it as evidence of an inaudible-harm mechanism. The paper does not argue that inaudible infrasound causes the sleep disturbances. It consistently attributes them to audible low-frequency components.

Jordan has now completely misrepresented all 11 studies he’s mentioned in the first five minutes of the video. 7 imply the exact opposite of what he’s claiming they do, 3 are completely unrelated to what he’s saying, and the only one that agrees with him isn’t a study at all, it’s a one-off anecdote written in a publication that mainly studies ghosts. From what I can tell his follow-up to this video is the most popular new piece of media made about data centers so far this year.

5:04 - I am very aware of the placebo effect and of course disclude my personal experiences with generated infrasound here in my lab when making this video. But I can tell you subjectively that it very much sucked. And even after turning the infrasound off, it took a while for things like intense nausea and dizziness to subside.

He’s confusing the placebo with nocebo effect here, but either way this is silly. He put himself in a position to experience the nocebo effect, and then he experienced it, and said “but it very much sucked.” I’d like a little acknowledgement that literally all good science points to this being a nocebo.

5:21–5:51 - The heart-contraction study

5:21 - A few years ago, a large published and peer-reviewed medical study showed that 100 dB of infrasound, primarily around 10 hertz, caused a very negative effect on the heart’s ability to contract with a 9% decrease in contraction force for every 10 dB above that. Your heart, an organ that some people would argue is necessary for survival, has to work quite a bit harder when you’re exposed to high levels of infrasound. And this is on top of quite a bit of animal research suggesting negative effects on heart, liver, nervous system, and lungs.

Jordan flashes “Negative Effect of High-Level Infrasound on Human Myocardial Contractility: In-Vitro Controlled Experiment” as he’s saying this.

You might be surprised to learn the full context of this paper implies something different.

Researchers at the University Hospital of Mainz took small samples of heart muscle tissue from patients having cardiac surgery, mounted them on tweezers in a laboratory apparatus, stimulated them electrically to contract, and exposed them to infrasound at 100 to 120 dB at 16 Hz for an hour. They measured a decrease in contraction force. That is the entire study.

Jordan gets the decibels slightly wrong. He says “100 dB.” The actual experiment tested three levels, 100, 110, and 120 dB, and most of the dramatic contraction-force decrease happens at 120 dB. Because decibels are logarithmic, that 20 dB gap between what Jordan said and what the experiment actually tested is a factor of 100 in sound intensity. But that’s a tiny problem compared to his weirder move here. In the same journal, two peer-reviewed critical responses of this paper were published and linked directly from the top of the PubMed Central page he’s screenshotting. They completely demolish the original paper. He either didn’t click them or hoped the viewer wouldn’t.

Here are the links:

• Comments on the Article “Negative Effect of High-Level Infrasound on Human Myocardial Contractility: In-Vitro Controlled Experiment”

• Can Infrasound from Wind Turbines Affect Myocardial Contractility? A critical Review

Three big fatal problems with this paper:

1. The critical reviewers found that Chaban’s apparatus produced excessive air movement, which caused the tips of the tweezers holding the muscle samples to vibrate. The muscle samples weren’t behaving differently due to the infrasound. They were being mechanically shaken by vibrating tweezer tips in the near field of the sound source. The critical review concluded: “The results of the paper of Chaban et al. are invalidated by a physically inappropriate setup that produces excessive air movement, not present in a regular sound wave of the same sound pressure in the far field.” Their overall conclusion: “Chaban et al. identified an artificial problem that doesn’t exist in reality.”

2. The reviewers pointed out that the total acoustic dose, the amount of pressure energy delivered over time, that caused an 18% decrease in contraction force in Chaban’s experiment (one hour at 120 dB) is matched by a single cycle of a swimmer doing the breaststroke. Every time a swimmer pushes their arms forward and back, the pressure their chest experiences from pushing through water equals or exceeds the total acoustic energy Chaban’s lab spent an hour blasting at dead tissue samples. If Chaban’s finding mapped onto anything biologically real, competitive swimming would be incredibly dangerous.

3. Infrasound never reaches anywhere near Chaban’s experimental volume in people’s homes. Wind turbine infrasound at typical residential distance from large modern turbines is about 65 dB, measured in high wind conditions at 200 m distance. Data center infrasound at residential distances is in a similar range. That’s 55 to 65 dB below what Chaban tested, a factor of hundreds of thousands in acoustic intensity.

I’m starting to think Jordan isn’t the most reliable narrator.

He pairs this paper with a vague appeal to “quite a bit of animal research suggesting negative effects on heart, liver, nervous system, and lungs.” He does not name the research. What he is gesturing at is the body of work around “vibroacoustic disease” or VAD, a hypothetical condition developed in Portugal starting in the late 1980s by a surgical pathologist named Nuno Castelo Branco and an engineer named Mariana Alves-Pereira.

Their central claim is that long-term exposure to low-frequency noise causes a “whole-body systemic pathology.” Symptoms include thickening of heart tissue, respiratory damage, cognitive decline, epilepsy, increased rates of tumors, depression, and eventually death. It’s a lot of responsibility for a single environmental mechanism! And this is the basic source of every claim about animal research and infrasound damaging organs.

VAD has never been accepted by any mainstream medical body. It has never been independently replicated in nearly forty years since it was proposed. The UK Health Protection Agency reviewed the evidence in 2010 and concluded that the disease “has not gained clinical recognition” and that there is no evidence infrasound at levels normally encountered in the environment causes it. A 2013 analysis of the VAD literature in the journal Noise & Health found that Castelo Branco’s self-citation rate across the body of VAD papers is 69%, and Alves-Pereira’s is 36%, meaning the “field” is mainly two authors citing themselves to build the appearance of a body of evidence. Leventhall, reviewing the literature, quoted Henning von Gierke, former chief of the USAF biodynamics laboratory, calling VAD “an unproven theory belonging to a small group of authors.”

Their animal studies use exposure levels way above anything found in residential environments, the kind of occupational exposure you’d get next to a jet engine. None of those studies have been replicated by independent researchers outside the Portuguese group.

When Jordan says “animal research shows effects on heart, liver, nervous system, and lungs” this is what he means. One research group, publishing on itself, at exposure levels nobody and no animals ever encounter, with a disease category no medical authority recognizes, that has failed to replicate for four decades.

5:51–6:37 - ‘It’s hard to study’

5:51 - Another pretty scary thing is that this list is likely very incomplete due to the difficulty and lack of research on infrasound. It turns out that not many people want to volunteer to feel super uncomfortable while their organs may get damaged for some reason.

Infrasound is one of the most thoroughly studied topics in noise research. We have decades of papers, dozens of controlled trials, multiple large epidemiological studies, and national health body reviews from multiple continents. The Marshall et al. 2023 study alone had 37 participants doing 72-hour laboratory stays with tons of equipment measuring basically all possible physiological effects. The Health Canada study covered 1,238 households. People absolutely do volunteer for these studies, they’ve been done.

“We just don’t have enough research” is a standard move in pseudoscience. It lets you dismiss the actual negative findings in the literature by asserting the literature is thin. But the literature is not thin, and its findings are not ambiguous. The reason he can’t cite studies showing what he wants them to show is not that the studies haven’t been done. It’s that the studies have been done and they’ve implied he’s wrong.

6:07 - If you look into this, a lot of the research that goes into the health effects of infrasound is overwhelmingly concentrated on infrasound caused by wind turbines. And yes, wind farms do generate infrasonic noise pollution. And it turns out that a lot of this research is either partially or completely funded by the fossil fuel industry to try and find something nefarious or wrong with wind energy.

This is an incredibly weird way to frame the actual research. Most of the infrasound research has been done in response to fossil fuel industry claims that wind farms cause harm, and the good studies have all shown they don’t. What Jordan is not-so-subtly implying is that the infrasound research is invalid because fossil fuel money touched it. In reality, basically every good study on infrasound cuts against the fossil fuel industry’s claims.

He does to his credit at least flash the Wind Turbine Syndrome Wikipedia page, which is a good debunking.

6:31 - And this situation is a major problem in science and research. I did a video covering some of this here

The phrasing implies that Jordan has a separate video laying out how the fossil fuel industry has distorted infrasound or wind turbine research, and that the curious viewer can go watch it. The viewer is meant to register “okay, there’s a deeper argument here and he’s made it elsewhere” and move on.

I watched the video he flashed on the screen. It’s where he’s growing wheatgrass and chia seeds in his home. He plays a 4,000 Hz tone at 72 dB at his seedlings for twelve days, observes that the sound-exposed plants grew about 28% more biomass, and speculates about the mechanism. The video is threaded with reflections on Gödel’s incompleteness theorems, the replication crisis, and his own decision not to go to college.

In this video Jordan doesn’t just misunderstand the replication crisis, he inverts it. He says:

The replication crisis is more about researchers not being funded to reproduce their own or other studies and some of that is probably because of capitalism.

This is not what the replication crisis is. The replication crisis is the finding that when researchers do try to reproduce published studies, a huge fraction of them fail. The Open Science Collaboration’s 2015 project tried to replicate 100 psychology studies and got statistically significant results in only 36% of the replications. The Reproducibility Project: Cancer Biology tried to replicate 193 experiments and found that positive findings replicated only 40% of the time. Similar patterns have shown up across economics and other fields. The crisis is about the validity of the underlying literature, not the funding pipeline for replication.

Jordan has the direction reversed. He’s framing the crisis as “we haven’t run enough replications yet,” when the actual crisis is “we ran replications and a lot of published findings didn’t survive.” The first framing implies scientific consensus should be trusted pending more work. The second implies scientific consensus has already been partially invalidated, and we need to figure out which findings still hold.

This matters because the whole video is a manifesto arguing that institutional science is credentialist, capitalist-corrupted, and not to be trusted, while independent researchers with their own setups are the real source of knowledge. Jordan explicitly tells viewers “there’s a major war on truth and objectivity happening right now” (I definitely agree with him that in the exact moment I’m watching this video, a major war on truth and objectivity is occurring) and that the replication crisis has been weaponized “to discredit the effectiveness of research and scientific method.” He then dismisses the reputation of credentialed scientists as “the battlecry of a really poor or weak scientific argument” describes institutional gatekeeping as “probably a much larger threat to science than the replication crisis,” and presents his own n=1 unreplicated wheatgrass result as meaningful.

This is what Jordan tells viewers he’s “covered some of” in that link at 6:37. It’s not an argument that the fossil fuel industry distorted infrasound research. It’s a much broader argument that mainstream scientific consensus should be treated as corrupt and that viewers should trust independent YouTubers running personal experiments instead. Once you understand that, the infrasound video makes more sense. He’s not trying to engage with the decades of controlled studies, epidemiology, and public health reviews that all point the other way. He’s pre-emptively dismissing them. The Marshall studies, the Health Canada survey, the Finnish project, the NHMRC review, these aren’t findings to be grappled with. They’re products of the institutional capitalist machine, and the institutional machine has been compromised. What you should trust instead is a guy with a hydroponic chamber and a sine wave generator.

His own n=1 wheatgrass experiment is exactly the kind of unreplicated finding the replication crisis is about. If the crisis teaches us anything, it’s that results like this should be held lightly until independent labs try to reproduce them.

6:37 - So, yes, wind farms do cause infrasonic noise pollution, while fracking in Texas, Oklahoma, British Columbia is literally causing record-breaking earthquakes. And how does the infrasound from all this fracking affect people? We don’t really know ‘cause we haven’t studied it nearly as much as wind farms.

The fracking point is wrong. The infrasound research on ambient industrial sources is directly applicable to fracking infrasound. There doesn’t need to be a fracking-specific study for the same reason there doesn’t need to be a study on the effects of infrared light from each different very specific industry. We know the general effects of infrared light, and the general effects of infrasound. The framing that every new industrial source needs its own full-scale research paper before we can conclude anything is a way of creating permanent scientific uncertainty and dismissing all the actual research. He’s also trying to separate himself from the main way infrasound pseudoscience has been badly used: by attacking wind turbines. I don’t think he gets to have it both ways here. If infrasound is actually harmful, wind farms at close range to homes do also create meaningful amounts, and he’d have to at least acknowledge the trade-off that infrasound harms are real but worth it for building out wind power. Instead he wants to have it both ways.

Also, on the earthquake point, USGS says felt earthquakes directly caused by hydraulic fracturing are extremely rare, and the big rise in places like Oklahoma is primarily tied to wastewater disposal. Fracking-induced felt events have occurred in some regions.

6:52–8:10 - The Vic Tandy ghost story

6:52 - In the late 1990s, Vic Tandy, a respected lecturer and engineer, was working in a laboratory in Warwick, England that had a long reputation of being haunted. He felt a cold sweat and an intense feeling of depression. And out of the corner of his eye, he saw a blurry gray figure.

Yes, in the late 1990s, Vic Tandy was a lecturer working in a medical equipment lab. The lab had a local reputation for being haunted. People working there had reported feeling uneasy, seeing things out of the corners of their eyes, etc. Tandy himself had an experience like this. He suddenly felt cold and anxious and saw a blurry gray figure in his peripheral vision that vanished when he turned to look at it.

The next day, Tandy was clamping a fencing foil in a vice when he noticed the blade vibrating on its own. He realized a large fan in the lab was producing low-frequency infrasound around 18.9 Hz, and he hypothesized that this infrasound might be what was making people in the lab feel uneasy and see peripheral figures. He published this idea as a short article titled “The Ghost in the Machine” in the Journal of the Society for Psychical Research, a parapsychology journal, in 1998. This wasn’t “a study” the way Jordan describes it. It was a short article speculating about a possible mechanism. Just one guy’s anecdote, no experiment, no controls, nothing.

When someone eventually did try to test Tandy’s hypothesis rigorously (the Goldsmiths Haunt Project, which will be discussed more below when Jordan brings it up) the infrasound had no effect on reported experiences. Participants’ prior suggestibility was the sole thing that predicted whether they felt uneasy. This paper reinforced the idea that infrasound effects are a nocebo.

In both cases, Tandy visited places with strong local reputations for being haunted, where the people he interviewed had already been primed by each other to feel uneasy. It’s very likely this was a nocebo.

7:33 - A large fan was causing infrasonic vibrations around 18.9 hertz. You know what else resonates around 18 hertz? The human eye.

The eyeball resonance thing sounds scientific but isn’t at all. The original NASA work it traces back to was about mechanical vibration transmitted through the body at very high energy, not infrasound at normal environmental levels. It’s like seeing the frequency you need to push a kid on a swing to make them swing higher and higher, and concluding that very lightly poking their back at the same frequency would also work. Nobody standing next to a fan in a British laboratory was experiencing the kind of vibration those studies used that would actually harm your eye.

7:49 - This would later be published in a rather famous paper titled The Ghost in the Machine.

It’s famous in pop-science and paranormal-adjacent writing. It is not famous in the acoustics or medical literature in the sense of being a foundational finding. Serious infrasound researchers treat it as an interesting anecdote that has never been replicated under controlled conditions.

7:54 - Soon after this Vic visited another famously haunted place, a cellar in Coventry.

A few years later, Tandy wrote a follow-up article called “Something in the Cellar,” also in the same parapsychology journal. He visited a 14th-century cellar in Coventry that visitors had reported feeling uneasy in, put a microphone in the room for 20 seconds, found a 19 Hz peak in the ambient sound, and declared this confirmed his earlier hypothesis. The measured sound level was 38 dB at 19 Hz, which as Tandy himself acknowledges is 50 dB below the human perception level for that frequency. His explanation for why inaudible infrasound at 38 dB might still cause uneasiness was to speculate about “hypersensitive” individuals and to cite a New Scientist article about hypothetical military infrasound weapons. Again, this wasn’t a “study,” it was a guy going around with anecdotes and random hypotheses.

The methodology is not an experiment. Tandy did not expose anyone to the infrasound. He did not remove the infrasound and see if the apparitions stopped. He did not compare the cellar to other cellars. He did not blind anyone, recruit a control group, or do anything that would count as a test of the hypothesis. He collected stories from tour guides about visitors who had felt weird in a cellar that was already advertised as haunted, then measured the ambient sound in the room afterwards. This is not science.

The anecdotes themselves are also pretty bad evidence. Tandy notes in his own paper that the cellar has “a growing reputation” meaning visitors were arriving primed. Two of the “witnesses” Tandy relies on were self-identified witches who visited the cellar specifically to commune with the spirit world. A third witch was “frightened to death” and left. One key informant is the tour guide herself, who, on her own account, “found herself talking to” the presence during tours. These are self-selected paranormal enthusiasts reporting paranormal experiences, and their testimony is being used to support the causal claim that inaudible infrasound at 38 dB produces supernatural perception.

Another obvious problem is low-frequency resonances are a feature of almost every partially enclosed space. A cellar’s dimensions and an entry corridor will produce standing waves in the infrasound range by basic acoustics.

So the “Cellar in Coventry” paper is a ghost-hunter-reviewed anecdote-plus-ambient-sound-measurement, at sound levels the author admits are below any established threshold for perception, reported in a parapsychology journal, using informants who included self-described witches, in a cellar already known as haunted, and detected normal infrasound levels for cellars. And when it was tested rigorously later, it failed. This is the second piece of Tandy’s case, and it is, if anything, weaker than the first.

8:00 - It’s not the most conclusive study in the world, but pretty damn interesting

This is definitely the pattern yes.

8:10–10:09: Backyard recording

This section’s framing, that he can’t identify the infrasound source and therefore it’s mysterious and maybe significant, is silly. Any outdoor microphone capable of picking up infrasound is going to record a continuous soup of low-frequency energy from a ton of sources at once. Distant traffic on highways miles away. HVAC systems in nearby buildings. Wind pressure on the microphone itself, even with a windscreen. Ocean waves if you’re within a hundred miles of a coast. Thunderstorms hundreds of miles away. Trains. Aircraft. He does at least clarify at the end of this section that infrasound can travel far.

The fact that he can’t identify specific sources for specific rumbles is because infrasound propagates extremely well over long distances (which he acknowledges later) and that there are innumerable sources creating it constantly.

When he says “Wow, I am not jealous of the people who live nearby” he’s blurring the line between commenting on noise pollution vs. infrasound harms. The first one’s real, second one’s fake.

He throws in a wind turbine and makes what I think is a pretty obvious point that the wind turbine isn’t as loud as a train starting up at close range or the world’s busiest airport. He’s doing this in part to distance himself from the anti-wind activists on his side of the infrasound debate. I don’t think he should get to have it both ways here. If he thinks infrasound is a problem, he needs to acknowledge that wind farms create it, and that living near a wind farm exposes you to new levels of infrasound. He can say “that’s bad, but worth it to have the wind farm” or agree with me instead and say “that’s fake, build the wind farm, don’t listen to the cranks” but he shouldn’t get to imply that wind farm infrasound is fine because it’s not as bad as a train at close range. Throughout the rest of the video he’s implied over and over that similar levels of infrasound as wind turbines create are bad for you.

10:25–12:00 - The South Dakota detour

10:25 - But I really wanted to hear clear infrasound far away from any form of human infrastructure. And to do that, we had to drive over 1,500 miles to a particular spot south of Badlands National Park in South Dakota on the Oglala reservation.

10:55 - From where I’m recording, the nearest paved road and infrastructure of any type is about 11 miles away, and it looks like this. So not exactly a teeming metropolis. The sound up here is remarkably quiet.

He drives 1,500 miles across the country, sets up on a remote spot on the Oglala reservation, records almost nothing, and concludes that rural places don’t have much infrasound while cities and suburbs do. He’s implying infrastructure is polluting the world with invisible infrasound, and only in the most remote places can you escape it.

But he’s actually done nothing to legitimately establish that infrasound in cities is causing the symptoms he’s been implying throughout. The contrast he’s trying to draw, rural silence versus urban infrasound poisoning, is not supported either by the scientific consensus or even the consensus of the articles he himself has flashed on the screen so far in this video.

11:38 - But this does not mean that nature is infrasonically quiet. The earth can produce infrasound that humans can’t even begin to compete with. This distant thunderstorm, for example, made a symphony of various rumbles.

This is an important admission that he glides past. Large-scale natural phenomena like thunderstorms, earthquakes, volcanoes, ocean waves, and wind blowing over terrain, produce infrasound at levels that dwarf anything humans have built. The Earth is and has always been an infrasound-noisy place. Humans evolved and thrived in this environment. It doesn’t make sense from an evolutionary perspective that these sounds would be harmful to us. If ambient infrasound at ordinary levels caused the broad symptom list Jordan describes, every coastal population on Earth would be chronically ill.

12:00 - And there’s some early research and testing on the accuracy of infrasound to detect and predict tornadoes.

This part is actually real, phew! Infrasound can be used to detect and predict tornadoes.

12:08–14:32 - The Yellowstone segment

He then drives to Yellowstone and records infrasound from geysers, hot springs, and the volcanic caldera. This section is pretty, and there’s nothing wrong with it scientifically. These are real sources of real infrasound, and recording them is neat.

14:24–14:49 - The inescapability frame

14:24 - But aside from these extreme places or unusual events, the vast majority of infrasound that the average person is exposed to is caused by infrastructure. And unless you want to live here, there’s little chance that you’ll ever be able to escape it or even significantly reduce it.

This is the rhetorical goal of the road trip. He’s pushing that he went to the most remote place he could find, and only there was it quiet. He’s trying to imply that infrastructure is saturating the world with infrasound you can’t escape, and that’s harming you.

The claim that “the vast majority of infrasound the average person is exposed to is caused by infrastructure” is also just wrong if you include natural sources honestly. Wind, weather, ocean activity (detectable far inland as microbaroms), atmospheric turbulence, and thunderstorms produce enormous amounts of infrasound that propagates long distances. A person in a quiet rural area in the Midwest is constantly bathed in infrasound from distant weather systems, microbaroms from the oceans, and atmospheric gravity waves. The infrastructure-produced component is not clearly dominant in most settings, and it’s a strange claim to make unsupported.

Jordan has already walked viewers through fake papers, a nocebo symptom list, and the Tandy ghost story, so “infrasound is harmful” feels established, and now he can compare the infrasonic silence of nature to big evil loud society. He’s sticking the landing for the basic idea that viewers are being poisoned by something invisible and inescapable. But if we got all the authors of the studies Jordan has cited, they would all agree (except for Vic) that infrasound doesn’t seem to be a problem.

14:49–18:00 - The physics detour and SpaceX launch

Here Jordan gives a physics tutorial on why low-frequency sound propagates farther than high-frequency sound, followed by a demonstration where he uses open seismograph data to track a SpaceX satellite launch across southern California. He shows the launch detected at 3 miles, 6 miles, 27 miles, 60 miles, and eventually over 140 miles away.

The physics is correct, low-frequency sound does travel way farther than high-frequency sound, which is why you can hear distant thunder as a low rumble even when the higher-frequency crack is gone. It’s similar to electromagnetic radio waves, which have a way lower frequency than visible light, and can travel way farther. And a SpaceX rocket launch is, unsurprisingly, a loud infrasound event that can be detected far away.

But it’s doing the same rhetorical work as the tornado and Yellowstone segments. Here is a dramatic infrasound event propagating very long distances, done by Elon, an obvious bad guy (who to be clear I agree is bad). The implication is that this tells us something about how dangerous infrasound propagates through your neighborhood. But SpaceX infrasound doesn’t harm people for the same reason ordinary traffic, weather, and HVAC infrasound doesn’t harm people: even very loud sources, by the time they reach residential distances, are well below any level scientifically shown to cause health effects, and Jordan has produced literally zero evidence to contradict this.

17:18–18:00 - Helmholtz resonance

17:18 - it’s worth pointing out that in many cases, infrasound is actually louder indoors than outdoors, even if the source is coming from outdoors. And this is due to a phenomenon called Helmholtz resonance.

Helmholtz resonance is real, but jumping from “open car windows can throb” to “your house amplifies residential infrasound into something harmful” is a big leap. The phenomenon exists. The idea that Helmholtz resonance in residential buildings makes infrasound loud enough to be harmful is, once again, not supported by any evidence.

18:00–18:56 - The Liverpool Cathedral study

18:00 - In 2003, a group of scientists from the UK National Physical Laboratory held a live musical performance at the Liverpool Metropolitan Cathedral and then took a survey from the audience on how they felt during the music. In some of those music performances, they had a gigantic 23 ft infrasonic cannon tuned to 17.5 hertz, a frequency that didn’t resonate the building in a way where it would shake things or create audible overtones. and they played that tone at 90 dB while some of the music was played to people who were completely oblivious to what was actually going on. It turns out that people were quite a bit more likely to have felt uncomfortable during the performance when the infrasonic tone was playing. And 22% of the audience said that they felt some type of strange or unexplainable effect such as sadness, chills, or even feelings of fear and anxiety.

This is the “Soundless Music” or “Infrasonic” experiment, a 2003 concert at London’s Purcell Room (not Liverpool Metropolitan Cathedral, Jordan has the wrong venue). It drew about 700 attendees and collected 522 questionnaires. This study is often misunderstood as implying that the infrasound caused 22% of people to experience something strange.

The core problems with this result have been covered by infrasound researchers repeatedly. First, no peer-reviewed paper with the underlying data was ever published and there was no control group. Second, the 22% figure reflects the baseline rate of people reporting something strange during evocative music in a dramatic venue. People at concerts report feeling moved, weird, uncomfortable, or spiritually affected all the time, regardless of whether there’s infrasound present. The “22% reported unexplainable effects” statistic reads as impressive out of context but less so when you realize it’s approximately the rate at which people report unusual feelings during evocative music without any infrasound at all. Third, the Goldsmiths Haunt Project (which will be covered next) tested the same hypothesis properly under controlled conditions and found that infrasound exposure did not predict reported experiences, suggestibility did.

18:56–20:16 - The Goldsmiths Haunt Project

18:56 - In yet another study from Goldsmith’s College, four researchers attempted to make a haunted room by filling it with various infrasonic frequencies and EMF or electromagnetic frequencies.

The researchers built a “haunted room” at Goldsmiths College and ran 79 participants through it. Each participant sat alone in the room for 50 minutes. The room had two manipulations they could turn on and off independently: infrasound at 18.9 Hz (the Vic Tandy frequency), and complex electromagnetic fields of the kind that have been proposed as a cause of anomalous experiences. Participants were randomly assigned to one of four conditions: infrasound only, EMF only, both, or neither. Everyone was told the room might produce unusual sensations.

After the session, participants filled out a questionnaire about what they’d experienced: feelings of presence, unusual body sensations, dizziness, emotion, anything out of the ordinary. They also completed a standard psychological measure of suggestibility beforehand.

Jordan narrates the study setup accurately, but then he misrepresents what it found.

19:23 - Out of these groups, an average of 3.7 participants in the control group, those sitting in the empty room with nothing unusual happening, had some type of haunted experience. When infrasound was present, this number grew to an average of 5.2 people having a haunted experience.

What Jordan leaves out is the actual finding of the paper. The number of reported experiences was not statistically different across conditions. The authors explicitly concluded that the number of reports was unrelated to experimental condition, and that suggestibility was the most parsimonious explanation for the experiences participants reported. Jordan cites the 5.2 vs 3.7 numbers as if they supported an infrasound effect, but the paper they come from concluded that the effect was not present and that the reports reflected participants’ prior suggestibility, not the physical conditions they were exposed to.

19:48 - Now, in my opinion, the study is fun, but not very well organized and definitely not conclusive.

He hedges on the study’s conclusiveness only after having cited its numbers in a way that implies the opposite of what it found. The hedge is positioned so the viewer walks away with the 5.2 vs 3.7 number in their head and a vague sense that “the science isn’t settled but it’s suggestive” when the actual published conclusion was that infrasound did not explain the reports.

This is yet another massive misrepresentation of a scientific paper in the video. The paper’s authors specifically tested Jordan’s hypothesis, found it was not supported, concluded suggestibility was the explanation, and published this. Jordan presents the paper as if it found the opposite.

20:16–End - The gear section

The final third of the video is a discussion of his microphone setups, the Raspberry Shake seismograph, accelerometers, and a mention of his “hum” video. This section doesn’t make any big new scientific claims about infrasound harming humans. I’m mainly worried this is yet another way to look more scientific and legitimate.

Conclusion

In a single 25-minute video about how infrasound is bad, Benn Jordan cited as evidence:

• A survey of farm equipment that noted infrasound is safe at levels 100,000x what we experience even in noise polluted areas.

• A WHO document about noise pollution that doesn’t mention infrasound.

• A Finnish government project whose final report is titled Infrasound Does Not Explain Symptoms Related to Wind Turbines.

• A NIOSH workplace report that explicitly says the sound levels it measured aren’t known to cause harm and attributes the workers’ symptoms to annoyance and a psychological contract breach.

• A literature review exclusively about the bad effects of audible low-frequency noise.

• A survey checklist of symptoms as if it were the symptoms subjects actually reported having.

• A study of wind turbine maintenance workers at 60 to 83 dBA of audible occupational noise.

• A study of people living near oil and gas drilling getting mad at the 80 dBC audible noise.

• A heart-muscle paper demolished by two peer-reviewed responses linked from the top of the page he screenshotted, which found the results were caused by vibrating tweezers, that the acoustic dose was matched by a single breaststroke, and that the intensity used was hundreds of thousands of times above anything in a home.

• “Animal research” that turns out to be vibroacoustic disease, a hypothetical condition published almost entirely by two authors citing each other for forty years and rejected by every mainstream medical body.

• A parapsychology essay written as a personal anecdote by a guy primed by ghost stories, and its sequel, which mainly interviewed self-described witches in a cellar already advertised as haunted at sound levels 50 dB below the perception threshold, and found that they felt somewhat uneasy.

• An experiment that found no effect, but Jordan presents it as having an effect.

• A study that found 22% of people at a concert with evocative music had strange experiences.

• NASA work on eyeball resonance frequencies without mentioning that these harms only appear when you are literally inside a rocket.

The running theme across every study that actually addresses it is that the symptoms people report are real but not caused by infrasound. They’re driven by expectation, annoyance, context, and suggestibility, which is the signature of a nocebo.

He implicitly analogized infrasound to UV light instead of the obviously more comparable infrared light. He used Helmholtz resonance to imply buildings amplify residential infrasound until its harmful without giving any actual evidence of harm. He called infrasound “grossly understudied” when it’s one of the most thoroughly studied topics in noise research. He implied the fossil fuel industry suppressed infrasound research when the fossil fuel industry is in fact the main funder of the pseudoscience he’s promoting. He claimed fracking infrasound needs its own dedicated research before we can conclude anything, despite fracking causing the same levels of infrasound as a ton of other sources.

He told viewers he had a separate video explaining the politics of all this. In that video he inverts the replication crisis, dismisses relying on credentialed scientists as “a really poor or weak scientific argument,” calls institutional gatekeeping a bigger threat to science than replication failures, and presents his own n=1 wheatgrass result as meaningful.

He drove 1,500 miles, emitting the CO₂ equivalent of half a million ChatGPT queries, to record silence in rural South Dakota and make a point he could have established with a Wikipedia search. He acknowledged on camera that the Earth produces infrasound that dwarfs anything humans have built, and then concluded that human infrastructure is saturating the world with harmful infrasound we cannot escape.

And this is, from what I can tell, the lead-in to what I think is the most popular piece of media made about data centers in 2026 so far.

Which brings us to the second video.

Second video - Datacenters Behaving Like Acoustic Weapons

Well now we can finally talk about data centers I guess.

I’m going to work through this in roughly the order of the video, but it’ll be useful to separate three things that Jordan deliberately runs together:

1. Audible noise pollution from data centers, which is real and harmful.

2. Ambient infrasound from data centers, which his own data shows is unremarkable. This is a fake problem.

3. The symptoms residents are reporting, which are real, but whose cause is the audible noise plus stress, not the infrasound.

The video’s central move is to associate 3 with 2, whereas in fact it’s all caused by 1

0:00–1:29 - Intro

0:00 - If you’re in the demographic of people who generally get recommended my videos, then chances are that you are not a fan of data centers.

As I said at the start of this post, he’s using this nuance as a credibility shield. It works on viewers because it reads as a signal that he’s willing to acknowledge nuance, but it’s really a way of pre-buying permission to make crazy uncharitable claims about an industry the audience already dislikes.

1:04 - Spoiler alert, the results are terrifying.

Spoiler alert, they reduce to random statistical noise in a terribly designed study by Jordan

1:36 - regardless of where your ethical, environmental, political, or economic interests are, if a data center is being built nearby your home, you’re generally kind of f*****.

Obviously I disagree. Data centers (with a few exceptions, like xAI, which comes up later in this video) mostly don’t seem to have harmed people. I don’t think the citizens of Loudoun County, one of the top 5 wealthiest counties in the country and home to the most data center capacity in the world, are “generally kind of fucked.”

1:56 - Last year I started hearing about some mysterious and troubling symptoms from people who live nearby newly constructed data centers

Jordan seems somewhat addicted to flashing completely unrelated studies on the screen.

This time he flashes a Dutch survey of 190 people who self-identify as experiencing a perceptible low-frequency hum in their environment, compared to 371 people who don’t. The researchers didn’t measure any sound in anyone’s home. They didn’t establish that the complainants were even actually being exposed to what they thought they were perceiving. They didn’t test causation.

Most importantly, they didn’t mention data centers once.

What the paper does document is that the complainants are disproportionately older and introverted, which is the exact demographic signature of other “environmental sensitivity” syndromes where the attributed cause has repeatedly failed controlled testing. The paper’s own authors explicitly acknowledge that “non-acoustic factors including sociodemographic and individual characteristics or personality constitute substantial predictors for reactions to noise.” This is likely a nocebo, and Jordan once again takes a paper that documents self-reports and presents it as a paper that proves causation.

He then flashes the symptoms list on the screen as he says this. You would think that if someone is saying “symptoms from people who live nearby newly constructed data centers” and is flashing a list of symptoms on the screen, those people would live near data centers, or the study would focus on this in some way. But you’d be wrong!

2:07 - I started looking into seismograph readings from around the country, and there was enough correlation for me to dig pretty deep into an area of research I already had some experience in

Consider what that first video was like, this is the “experience” he’s referring to.

It’s true that data centers register on seismographs. So do rail lines, highways, hospitals, college campuses, shopping malls, factories, power stations, and large intersections. I suspect he’s going to draw some bad inferences.

2:24–3:22 - The symptom list, round two

2:24 - Being in the presence of elevated levels of infrasound has been shown to show spikes in cortisol levels... It can cause vestibular issues leading to loss of balance, vertigo and nausea. Then there’s something called vibroacoustic disease which suggests that infrasound can cause abnormal growth of extracellular matrices... It’s been shown to cause high frequency hearing loss, shortness of breath, anxiety, depression, and it even makes your heart have to work harder to pump blood.

He’s flashing the same study from last time that only measured audible low-frequency sound, not infrasound. Great! He’s doing the same trick of flashing a bunch of highlighted words without making it clear these are all about audible sounds. He seems more confident than last time. He also references the disproved heart study.

“Has been shown” is carrying most of the weight. Every one of these claims is either true only at audible levels orders of magnitude above ambient, drawn from the VAD literature that no medical body accepts, or flat-out contradicted by the controlled studies he himself flashed on screen in the first video. Literally everything he cited last time either contradicted these claims, didn’t address them at all, or was published in a journal about ghosts.

In the last video Jordan didn’t explicitly mention VAD pseudoscience, he just vaguely referenced it as animal studies, so he’s escalated in mentioning it directly here. Again, VAD is the pet theory of a small Portuguese research group whose work has not been replicated and which the peer-reviewed acoustics literature has specifically assessed as unproven.

2:57 - So, it’s not like this anxiety-based or obsessive compulsive thing where an unpleasant feeling or an unusual situation makes you manifest symptoms. We know that infrasound can tangibly harm your health in a very real and measurable way.

As we saw in the last section, we don’t know this at all. Basically all the literature Jordan cited in the last video said the opposite of what he claims here. He just flashed all the same sources again that all show completely unrelated results to what he’s implying or aren’t respected in the broader literature at all and rely on circular citations.

Here he’s more directly rejecting the nocebo explanation, whereas in the last video he at least passively mentioned it. He’s now implying this is established science. He’s telling the viewer, in advance, that anyone who raises nocebo is essentially accusing people who were harmed of making things up. This is what you’d expect from someone who knows the nocebo explanation is the most supported one: shield the claim behind real victims.

4:06–9:10 - xAI’s Colossus

xAI’s Colossus is a massive outlier among data centers, and there are a lot of real, widely-covered problems with it. The biggest is that it ran up to 35 unpermitted methane gas turbines for roughly a year as primary power generation, behaving like an unregulated mid-sized gas power plant in an already heavily polluted poor neighborhood. The methane turbines, the local air quality impacts on Boxtown, the nitrogen oxide emissions, and the permitting problems are all real, well-documented problems.

As a quick aside he also cites xAI using a lot of water, but doesn’t give a frame of reference for how much it is and just zooms up on a depleted water source. In reality, Memphis has some of the lowest water bills in the world, their water bills haven’t risen as a result of xAI, and Memphis isn’t considered a water stressed-area.

But none of the problems with xAI are about infrasound, and lies about infrasound cheapen the real harm people experienced.

7:46 – The lowest of rumbles is peeking out at like -28. And then if we go down to infrasound, the infrasound that you can’t hear is like 10 dB louder than that.

He’s reporting levels in what looks like digital audio reference (dBFS) rather than sound pressure level in the environment. Without a way to understand how he’s measuring sound, you cannot convert arbitrary recording levels into absolute sound levels, and he isn’t showing his method anywhere. When he says “infrasound is 10 dB louder than [audible rumble],” we don’t know what absolute level that represents in the environment. The entire infrasound literature is about absolute exposure levels. Relative ones aren’t useful here, and 10 dB above an uncalibrated reference tells you nothing about whether the infrasound is near the threshold of where we can hear it.

Of course infrasound is louder than the audible component at distance. Low frequencies propagate better than high frequencies. The lower the frequency, the less it diminishes over distance and the better it passes through walls. The ratio of low-to-high energy always increases as you move away from any industrial source. He made this clear himself in the brief wave physics section of his last vide.

8:02 – This in particular is a great example of how higher frequency sounds are dampened by distance much more than lower frequency or infrasonic ones.

Yes, correct. But notice what he does next: uses this correct physics observation to imply that infrasound is doing something special and harmful at distance, when all the physics actually shows is that low frequencies dominate at distance. What matters for our health is exclusively audible sound. He has yet to actually demonstrate any harms from infrasounds, but this entire video will now just assume infrasounds are harmful, resting on the terrible edifice he constructed in the last video.

8:26 – This rumbling is no good and would make your life utterly miserable. And then this rumble down here is happening at an amplitude that’s loud enough to shake the frame of your house.

If infrasound were actually shaking the frame of a house, that would be an extraordinary claim and easily falsifiable. It would also be perceived as a rumbling vibration by residents. He provides no measurements to support the “shake the frame” claim, and none of the residents interviewed mention their homes shaking. There has never been any accusation at all that Colossus shakes nearby houses.

8:44 – Okay. So, I’m going to delete everything above 20 hertz here. So, all audible frequencies. And now we’re only left with the infrasound. Now, I’m going to interpret the sample rate four times faster. And I’m going to make it even four times faster than that.

Same trick as the beginning of the first video. If you speed up infrasound and make it audible, it sound ominous. Just like if you look at the infrared light coming from your dog, your dog looks evil. It’s not evidence that the specific source he’s recording near is producing unusual infrasound. It’s evidence that infrasound exists, which no one disputes.

The xAI section is him listing a bunch of real harms, water, and then doing a bunch of fancy tricks with fancy but inaccessible sound terminology and equipment to support his unrelated baseless conspiracy theory.

9:10–15:18 - MARA / Granbury

Noise pollution from the Marathon Digital / MARA Bitcoin mine in Hood County has caused well-documented harm to nearby residents. Cheryl Shadden’s experience is real. The family whose daughter had seizures and whose symptoms resolved after they moved is real, and I’m not going to second-guess what they went through.

What I am going to push back on is Jordan’s causal story. He’s going to take these people’s real terrible experience with noise pollution and imply it’s about his crazy conspiracy theory instead.

10:15 – While the noise pollution is described as a 70 to 90 decibel jet engine-like sound, a large number of reports site symptoms ranging from vertigo to pulmonary embolisms to heart palpitations.

70–90 dB sustained is awful. Chronic exposure at those levels sits above the threshold at which WHO guidelines associate environmental noise with cardiovascular disease, sleep disruption, and a range of stress-related outcomes. These effects are well-documented in the science of audible noise pollution and do not require any infrasound to explain. A family living across the street from a 70–90 dB jet-engine-like sound will absolutely experience vertigo, heart palpitations, anxiety, sleep disruption, and cascading health problems. The MARA case is an audible noise pollution disaster.

10:39 – Sustained audible noise pollution can be a blight on any living thing nearby. But a lot of these symptoms don’t line up with the symptoms of just noise pollution exposure.

This is just completely wrong. The symptoms he lists, like vertigo, heart palpitations, anxiety, sleep disruption, and nausea, are exactly the canonical symptoms of chronic environmental noise exposure. There is a massive literature on this going back to the 1970s. Every major environmental health textbook describe this symptom cluster as the expected health consequence of chronic exposure to 70–90 dB environmental noise. He’s just lying here.

At 11:39 he starts playing more scary clips of sped up infrasound, with no evidence this is harming anyone, after leading in about how the residents told him about the real effects of noise pollution. What a dumb way to hide a baseless conspiracy theory behind people’s real suffering.

The residents of Granbury are suffering from real terrible noise pollution. Infrasound adds nothing to the explanation, and Jordan’s last video gave us not a single reason to believe it’s real in the 25 minutes it runs.

13:08 – Like, what physical symptoms have you felt just from your property being directly across the street from the data center? Motion sickness, dizziness, nausea, GI side effects. My ears ring and buzz. I’ve got permanent conduction hearing loss now. Um, nervous, anxious, worried.

Every symptom Cheryl describes is well-documented as a consequence of chronic high-level environmental noise exposure. The conduction hearing loss in particular is a direct audible-noise injury. There’s no good evidence that infrasound causes hearing loss. Jordan presents these symptoms as mysterious that infrasound can explain, but they’re not mysterious at all.

13:24 - Cheryl thought that I was just doing a story on the noise pollution as I was keeping my infrasound research close to my sleeve.

Okay, so this guy nods along to this person who was harmed by very real noise pollution, and lists classic symptoms of real noise pollution, and then adds “Oh by the way, I didn’t tell her that I’m actually just interviewing her to promote my bullshit unrelated conspiracy theory that I’ve never found a single good source to justify, and I’m just nodding along and am going to lie about her symptoms actually being caused by this other thing.” This is terrible! He lied to this real victim of noise pollution.

13:32 - Yet, every single one of these symptoms is commonly associated with excessive infrasound exposure.

Here the sleight of hand becomes explicit. He’s asserting that because the symptoms overlap with the infrasound horoscope-list, they must be caused by infrasound. But the symptoms also overlap, far more completely and with actual mechanistic support, with chronic audible noise exposure, which is unambiguously what’s happening at Cheryl’s property. The correct inference from “Cheryl has symptoms X, Y, Z while living across from a 70–90 dB jet-engine-sound source” is “chronic audible noise exposure produces those symptoms,” not “there must be a hidden inaudible cause.”

14:03 – One family, their daughter was 6 years old. She started having seizures in the summer of 24... they basically walked away from their home here, took out a second mortgage, moved to Somerville County... If you know, has her symptoms kind of gotten better since she’s left? They went away.

The family’s experience is real and terrible. But did infrasound cause the seizures? Chronic sleep disruption from audible noise in a young child can trigger a wide range of neurological symptoms, including lowered seizure thresholds in predisposed children. Severe family stress from a noise-blighted home can too. Moving to a quiet environment resolves the audible noise exposure, the sleep disruption, the household stress, and any potential infrasound exposure all simultaneously. The case is entirely consistent with audible noise and stress being the cause. It provides no evidence for infrasound specifically. Jordan’s never produced any actually good evidence that infrasound causes seizures.

Presenting this family’s story as evidence of an infrasound mechanism is both scientifically unjustified and strategically unhelpful to families in their situation, because it directs the legal and regulatory conversation toward a fake mechanism rather than the obvious provable one: MARA is producing 70–90 dB of environmental noise pollution at residential property lines, and that alone is enough to make people sick.

I really don’t like that he flashed a quick out of context clip of this woman at the very beginning of the video to imply that infrasound caused her problems.

15:18–17:58 - The Permian Basin

This section accidentally undercuts the video’s central claim.

17:32 - But while the infrasound levels were nowhere close to ideal, they still weren’t as bad as they were near the data centers in Memphis or Granbury.

A region with thousands of earthquakes per year, massive industrial fracking operations, and oil and gas infrastructure as far as the eye can see produces less infrasound than two specific data centers? Come on.

Fracking, large pipelines, oil rig drilling, and induced seismicity all produce vastly more low-frequency energy than cooling fans and generators. If Jordan’s measurements show otherwise, the most likely explanations are measurement artifacts from uncalibrated or differently-positioned microphones across sites, wind loading differences on the recording equipment, or distance differences between measurement locations he isn’t controlling for. Ridiculous.

17:39 - The infrasound levels at these two data centers were so much higher than the ambient values that it doesn’t even make practical sense to use a linear scale like 25,000% higher.

“25,000% higher” is a red flag. In physical acoustics, a 25,000% increase in pressure amplitude is roughly 48 dB, which would put the data center sites far above any residential or rural baseline anywhere in the developed world, including large airports, freeway interchanges, and active industrial sites. This is not plausible for ordinary data center cooling or power delivery, the main way they make sound,. A much more likely explanation: his baseline (Death Valley, random indoor hotel rooms, etc.) is at or below the noise floor of his measurement equipment, while his site recordings are well above it. Dividing “real signal” by “noise floor” gives you any number you want.

This is a recurring issue with amateur infrasound measurements. The noise floor of affordable infrasound-capable equipment is not zero, and the apparent ratio of “loud site” to “quiet baseline” is dominated by the noise floor of the gear. Published peer-reviewed data center acoustic measurements do not find anything remotely like 48 dB elevations over ambient at residential distances.

He’s an audio engineer. He knows what he’s doing here.

17:58–23:15 - The experiment

This is the centerpiece of the video, and is a disastrously bad experiment. I’ll walk through the problems one at a time because each one independently invalidates the claimed findings, and the combination is devastating.

Problem 1: The “haunted painting” priming

19:17 - I improvised a new story of why the owl painting was known to be haunted... I told them that the audio recording that they would be listening to was known to be haunted. I had them sit in this enclosed room while studying the painting and listening to the music for 3 minutes while paying attention to how they felt physically and emotionally.

He has pre-primed every single participant to expect unusual sensory or emotional experiences. This is exactly the Crichton-Petrie nocebo-priming paradigm from the first-video writeup, and it’s been shown repeatedly to produce symptoms on its own, regardless of whether any stimulus is present. His rationale (that varying the story randomizes a bias variable) is wrong. Randomizing the specifics of the priming across groups does not eliminate the priming. I’m confused about why he’s doing this.

A correct design would have one group primed negatively and one group given neutral or no priming, the way Crichton and Petrie actually did it. His design has no unprimed group. Every subject walks in expecting something weird to happen.

Problem 2: He isn’t blind

20:10 - And unless I saw water vibrating in a bottle or something, I often didn’t know if the infrasound was on or off myself.

“Often didn’t know” is not good enough. A study is double-blind when the experimenter does not know the condition, full stop, not when they usually don’t. If the experimenter can sometimes tell what’s up, it’s possible that their behavior, facial expressions, tone, and subtle cueing can leak information. And Jordan explicitly says he can tell via the vibrations. A real study would have a separate operator running the stimulus, ideally out of the room, with the experimenter truly unaware of condition until after data collection.

Problem 3: The stimulus itself

18:04 - I needed to figure out a way to reproduce the infrasound waves at 25 to 30% of the volume or amplitude near the data centers.

He’s playing back infrasound through “a gigantic specialized subwoofer in the room.” Commercial subwoofers, even specialized ones, do not reproduce sub-20 Hz content cleanly. They produce substantial harmonic distortion, port noise, cabinet resonance, and mechanical rattle, all of which show up as audible artifacts above 20 Hz that subjects can hear and that will absolutely affect their reported experience. Without a measurement showing the actual acoustic output in the room across the full spectrum, you cannot claim that subjects were exposed to “infrasound” specifically rather than to audible distortion products from an overdriven subwoofer.

His own video gives this away:

20:23 - Just put my DJI gimbal on the ground here. And Jeez, it’s kind of scary.

This implies the room is shaking. If the room is shaking, subjects can detect when the infrasound is on or off, and are probably going to feel weirder when it’s on.

Problem 4: The self-selection filter

21:33 - The experiment had over a 100 participants, but I disqualified anyone who missed a survey question or someone who seemed suspicious or said that they felt a vibration or where I was privy to whether the infrasound was on or off.

He dropped subjects who “seemed suspicious” or who reported feeling a vibration, based on his own after-the-fact judgment. I’m worried that “Seemed suspicious” is way too much of a subjective call by Jordan, who seems to have a bias in wanting the experiment to show results. And removing subjects who reported feeling a vibration specifically removes the subjects who noticed the confound identified in Problem 3, which is exactly the wrong direction of exclusion. Those are the subjects whose data is the most informative about what’s actually happening in the room.

We also don’t know how many of the original 100+ were dropped, which groups they were in, or what the results look like with them included. A valid pilot reports intention-to-treat results. This one doesn’t.

Problem 5: The effect sizes are tiny and the framing is misleading

To his credit, Jordan half-acknowledges this:

23:58 - the 33% more likely to experience nausea metric, the viewer may imagine a room with 100 people in it and then 33 of them experiencing nausea, which is very much not the case. The average nausea feeling score from 0 to 10 that each user put in for infrasound is 1.2. For the control group, it’s 0.9.

This is “33% increase” from 0.9 to 1.2 on a 0-10 scale. This is an 0.3 difference, in a sample of 74 with no correction for multiple comparisons across what appears to be a dozen or more possible symptoms, in a room where people can probably feel the floor vibrating when the infrasound is on. This is noise. These are the results that caused him to say at the beginning that “the results are terrifying.” He’s even misrepresenting the results of his own study.

24:55 - the average score in the control group for discomfort was 1.2. The average score for discomfort in the infrasound group was 4.8.

A 1.2-to-4.8 shift on a 10-point scale is a large effect, if it’s real. But consider what “discomfort” means in a room where a subwoofer is physically shaking the floor hard enough to move a DJI gimbal, where subjects have been told the space is haunted, where they’ve been warned the audio is spooky, and where the experimenter may sometimes know the condition and cue accordingly. The combo of all of these seems like the clear cause of the “discomfort,” not inaudible infrasound.

The discomfort result is exactly what you’d predict based on the nocebo studies. Prime people negatively, then give half of them a perceptible physical stimulus like a vibration they can feel through the floor or audible subwoofer distortion and they will report more discomfort.

With a dozen comparisons in an n=74 study, you expect chance alone to produce several “significant” differences. There’s no correction, no pre-registration of primary outcomes, no null-hypothesis framing. Some of the results point the “wrong” direction for his hypothesis (infrasound subjects were less creeped out and less spiritual) and he offers a post-hoc story to explain those away while treating the ones that point his way as meaningful.

Problem 6: Multiple comparisons

He reports roughly a dozen outcome measures (tingling, pain, tiredness, nausea, dizziness, creeped-out, spiritual, chills, irritability, eye irritation, lethargy, anxiety, sadness, discomfort), finds a spread of increases and decreases, and presents the increases as findings. With a dozen comparisons in an n=74 study, you expect several “significant” differences by chance alone. There is no correction, no pre-registration of primary outcomes, no null-hypothesis framing. Some of the results point the “wrong” direction for his hypothesis. Infrasound subjects were less creeped out and less spiritual, and he offers a post-hoc story to explain those away while treating the ones that point his way as meaningful.

What the experiment actually shows

If you prime a group of synthesizer-convention attendees by telling them a painting is haunted and some audio is spooky, then expose half of them to a subwoofer producing enough mechanical vibration to visibly shake objects in the room, the exposed group will report more discomfort and dizziness on a survey. This doesn’t tell us anything about data center infrasound. Jordan is just accidentally reproducing the nocebo literature.

23:15–25:43 - YouTube data vs research data

I’ll give him credit for this section, which is better than I expected. The acknowledgment that “33% more likely” on small absolute numbers isn’t meaningful is the right methodological instinct. The nausea shift from 0.9 to 1.2 on a 0–10 scale is, as he notes, not a room with 33 of 100 people vomiting. The recognition that the eye-irritation result is underpowered is correct. If the rest of the video had been pitched at this level of honesty, it would be a very different artifact.

But notice the structure: he does the methodological self-criticism after spending most of the video making the strong version of the claim, and he still ends up leaning on discomfort and dizziness as “profound” findings without acknowledging that those are precisely the measures most likely to be contaminated by the priming and the vibration confound. The honesty is real but selective. He applies it where it protects him from the “33% nausea” line being read literally, but not to the deeper question of whether his protocol measures what he claims.

25:13 - When reviewing that data, it’s hard for even the most skeptical person to come to any other conclusion than the presence of elevated levels of infrasound is likely to lower your quality of life.

Very very very very easy for the skeptical person, actually. The skeptical person concludes: priming produces reported discomfort, perceptible subwoofer vibration produces reported dizziness, and an unblinded single-operator study with post-hoc subject exclusion at a synthesizer convention is not a basis for policy claims about data center health effects.

25:43–end - The call to action

27:12 - Prior to this experiment, I feel like we have seen more than enough data to make infrasound monitoring and regulation just as important as air or water quality.

NOOOOOOOOOO!

Outdoor air pollution kills 4 million people every year. This sentence is elevating a fringe belief with zero evidence to the same level as one of the single largest killers on the planet. Crazy. After 40+ minutes of this guy misrepresenting studies and concepts almost by the second, this moment really made me lose it.

27:21 - Especially when you look at the studies pertaining to the cardiovascular effects, it’s a very real environmental hazard that has been poorly understood and grossly understudied.

Another repeat of his false claim that infrasound is “grossly understudied.” There are no good studies that imply cardiovascular effects. It’s fake.

28:21 - Now would be a great time to set up something like a Raspberry Shake and start storing a daily log of seismic and infrasonic activity. Because the vast majority of the responses to complaints and lawsuits is that there is no way of proving whether the infrasound wasn’t already present before construction.

If communities want to collect baseline acoustic data in advance of data center construction, fine, go ahead. I agree with critics of data centers that they often haven’t been careful about audible noise pollution, that this has caused real quality-of-life harm, and that communities should be able to monitor it.

But communities shouldn’t monitor infrasound for the same reason they shouldn’t monitor wifi signals over fears about wifi-intolerance or cancer. There’s no evidence for either. Monitoring would just waste everyday people’s time and leave them more confused and paranoid when they realized how much infrasound is around normally. It’s like hyping people up about the dangers of infrared light and sending them out to test it themselves. They’ll find a lot!

Conclusion

I found this second video to be way more objectionable than the first. Even though the first had nonstop lies about what every study flashed on the screen said, at least it didn’t have interviews with victims of real noise pollution Jordan was trying to use as evidence of his unrelated baseless conspiracy theory instead without even telling them. I found that pretty gross. I was also repulsed by his comparison of infrasound to outdoor air pollution, one of the great global killers right now.

My claim remains strong. Infrasound harms are fake, data centers or otherwise. People should focus on the real problems of noise pollution and not give an inch to baseless conspiracy theories promoted by chill-seeming guys using fancy-sounding technical terms and unrelated studies flashed on the screen to trick millions of viewers. Infrasounds have been used by all kinds of bad actors to oppose things we should all agree are good, like wind farms. Let’s not give the crazies more power here.

Some stray thoughts

If infrasound exposure near data centers caused this much harm, wouldn’t the absolute worst off people be people working in data centers themselves? They’re in the belly of the beast of data center infrasound, many have been working in data centers for way longer than the current buildout’s been happening, and it’s very easy to identify who works in data centers. They seem like the ideal people to look at for data center infrasound effects, but never get a mention.

Responses to this piece from the authors of the studies Jordan cites

Someone had suggested to me that since Jordan and I disagree about the content of these studies, why not just reach out to the authors themselves?

I reached out a few days ago. So far I’ve received one response from Swen Müller, who wrote the rebuttal to the heart study. Jordan cited the heart study but not the rebuttal, so this isn’t as good as having a direct quote from the authors themselves. Here’s Müller’s thoughts:

I’ll update this section as I get more replies.

Jordan’s response to this article was pretty goofy and reaffirmed that he can’t defend any of his misleading citations

Bluesky back-and-forth

Jordan responded to this post on Bluesky, and it went pretty badly. Opened with this

Then added this. Sure, I’d say people harmed by data center noise are vulnerable, but maybe that was too strong.

Here’s the next thing in the thread + my response, no reply from Jordan here.

He then accused me of hypocrisy here:

Which is a misreading of this post. I never said that it’s invalid to compare sound to light. Like he said, I do it over and over. I think it’s a useful comparison. What I specifically said was that it’s invalid to compare infrasound to the harm of ultraviolet light specifically, and I made that explicit here:

He’s setting up this framing about how things that are invisible to us can still harm us, and then very quickly moves from ‘ultraviolet light can obviously kill us’ to ‘and there’s new disturbing evidence that infrasound can harm us’ without acknowledging the obvious problem that infrared light does not really harm us at all unless it’s so powerful that it heats up our bodies in the way other normal heat sources do. This is the correct analogy to infrasound, not UV light. Just like infrared light can only harm us if it’s powerful enough that we physically feel it in the way we do other heat sources, infrasound can only harm us if it’s so powerful that it causes us to physically feel its presence in the way we do other very loud sources of sound. He’s not so subtly reaching way across this spectrum to say that because something that’s so high energy that we can’t see it can cause damage, maybe something that’s so low energy that we can’t hear it can cause damage too.

I replied making this point and didn’t get a reply. Jordan seems to have a pattern of misrepresenting articles.

I finally got a response here:

No response there so far…

There was one more thread where I got responses:

Instead of answer he just asked this:

We went back and forth for a little while and I asked him for a meta-analysis showing that infrasound was legitimate science. He sent the following links:

This is the same heart study I mentioned above. It’s an individual study, not a meta-analysis. “Cited 5 times” is a weird defense of the paper, it makes it look pretty bad. Two of those citations were the critical responses I posted above basically showing it was useless, and 3 other citations for a paper like this makes it seem like it hasn’t had any impact at all. I’ve been unable to find what those citations were. When I brought this all up he accused me of “moving the goalposts” but couldn’t actually explain why this study was useful beyond that it was a scientific study.

Importantly, he’s the one moving the goal-posts here, because I had specifically asked for a meta-analysis, not a random study.

The next link he sent was this one:

This one’s really weird. It’s not a meta-analysis or a study. It’s a single-author narrative review, basically a long op-ed.

The author is Michael Persinger. He’s best known for the “God Helmet”: a device he built that he claimed could induce mystical, religious, and paranormal experiences in subjects by applying weak magnetic fields to the temporal lobes. When an independent Swedish team ran a proper double-blind replication in 2005, they found nothing. Persinger’s response was essentially that the Swedes hadn’t done it right.

He also published, at length, on telepathy, precognition, “non-local” consciousness, and the idea that the Earth’s geomagnetic field modulates paranormal experiences. A substantial chunk of this guy’s career was in territory that everyone else in neuroscience treats as pseudoscience. He was a professor at Laurentian University until his death in 2018, but his research program was consistently well outside the mainstream and consistently failed when other labs tried to reproduce it.

Benn said “Their validity is supported by convergent quantitative biophysical solutions.” What that sentence actually means, once you trace the paper, is that Persinger’s claims about infrasound are supported by Persinger’s other work on biophysical mechanisms. There wasn’t any research done here.

Actual systematic reviews of infrasound (NHMRC 2015, Baliatsas 2016, van Kamp and van den Berg 2018, Marshall 2023) all converge on the conclusion that infrasound at the levels people are actually exposed to doesn’t produce the symptoms people attribute to it.

So when I asked for his best evidence, the second thing Jordan sent me was a 2014 narrative review, by the God Helmet guy, that vouches for its own conclusions by citing the author’s other papers.

Next was this one:

This paper found that infrasound, when delivered to chinchillas at 100 decibels, simultaneously with a chainsaw-level blast of audible 4 kHz noise, slightly increased cochlear hair-cell damage compared to the chainsaw blast by itself.

These are the paper’s results:

• At 108 dB of noise exposure, adding the 30 Hz tone slightly increased cochlear damage compared to 108 dB alone.

• At the 86 dB exposure, adding the 30 Hz tone did not increase damage at all.

• The authors’ proposed mechanism is that the 108 dB exposure perforates a membrane in the cochlea called the reticular lamina, and the low-frequency fluid movements makes the existing damage worse by mixing cochlear fluids through the hole. The low-frequency exposure only contributes to damage after a separate, extremely loud audible exposure has already wrecked the structure.

So the subjects are not human, the “infrasound” is audible to humans, the exposure levels are not remotely close to what someone living near a data center could encounter, and the finding is that a 30 Hz tone only contributed to damage when a separate exposure had already caused the damage. Alone, it did nothing.

And finally this one (after a mislink of a paper on PTSD that didn’t mention infrasound):

This is a link to a 2016 undergrad paper by Ryan Littlefield that was self-published on Medium and never peer-reviewed or published in a scientific journal. Semantic Scholar indexes it, but it indexes student papers indiscriminately. A paper being on Semantic Scholar tells you nothing about how legit it is.

I pushed back on each. Jordan’s final comment to me was anger that I was calling “published papers pseudoscience.”

He said these were from PubMed, but they’re actually both from PMC (ignoring the Medium article and single-author narrative by the God Helmet Guy he sent). PMC is an entirely separate paper archive with a much lower bar for inclusion, and mixing them up is a very common way to justify pseudoscience.

PubMed is built on MEDLINE, whose journals have to pass a scientific quality review by the National Library of Medicine (NLM). PMC accepts material under looser criteria, and, importantly, it also hosts NIH-funded author manuscripts that get deposited automatically regardless of whether the journal they appeared in has ever been reviewed for quality by anyone at NLM. According to NLM’s own staff, these manuscripts may come from journals that have never undergone scientific review, are outside the library’s traditional scope, or have failed to meet even PMC’s standards. Once a paper’s in PMC, it becomes searchable through the PubMed interface, which is how a paper from a journal no serious biomedical reviewer has ever vetted ends up being defended with the phrase “on PubMed.”

The Yale Medical Library flags this exact move is common enough that readers should watch out for it. A publisher can claim to be “in PubMed” when they’re really just searchable via PMC. It is a well-known laundering pathway for predatory and low-quality journals, and is popular among infrasound pseudoscience.

So yes, this all looks extremely pseudoscientific.

Jordan then blocked me and posted this:

After a lot of back and forth he wasn’t able to stand behind any of the studies he cited in his video. I rest my case that this video is full of fake information and pseudoscience, Jordan seems to know this, and is good at dodging this info and reframing any criticism as a “good guy vs bad guy” thing that will cause a lot of his audience to not look into it further. Jordan is knowingly promoting infrasound pseudoscience and is getting millions of views doing it.

Jordan’s blog post rebuttal of this post

Jordan then posted a blog post reply to this article, and it’s pretty bad. You can read the whole thing here. I’ll go down the list and respond to points from it.

But it seemed desperate and mean. I’m attacked for…creating a sundial and having “scientific looking microphones”.

Jordan is obviously knowledgeable enough about sound to be able to read and understand the incredibly simple language in the studies he’s flashing on the screen. I don’t know how it could be possible that he doesn’t know what they contain. I’m forced to conclude that he does know and is choosing to mislead the viewer with half an hour of wild intentional misinterpretations of very simple studies. So yeah, I think he adds scientific looking stuff to make himself seem more legit.

The author is constantly ping-ponging the people interviewed or featured in my videos as poor, stupid people, or people that are owed more empathy than I allotted. Something is just weird about this.

I don’t do this? I say he talks to poor vulnerable people who have been actually definitely harmed by real noise pollution. They’re not stupid at all. I do think people are owed not having their very real harms rebranded as the result of a pseudoscientific sham yes.

So after some sniffing around, I realize I’m being brigaded by someone deep in the Effective Altruism community.

Oh no how did he notice I was really trying to hide this one.

But please do browse his bibliography.
Andy lives in a parallel universe where datacenters don’t waste water, AI artwork is without victims, and using ChatGPT doesn’t harm the environment.

✅ Data centers don’t waste water.

✅ AI artwork is without victims.

✅ Using ChatGPT is not bad for the environment.

I love the idea that the only way you could possibly believe these is if you’re being paid to. Very intellectually walled-off from a lot of the world. I guess when you’re wrong, the correct people do live in a parallel universe.

This is one of the many takes that perfectly align with the board, contributors, and partners with Coefficient Giving

I only received a grant to write from Coefficient Giving a month ago. All the articles he cited above are from when I was just writing for fun.

Unlike most people with shadowy billionaire benefactors, mine and I are buddies on Bluesky. Here’s his input on my writing:

I spend a lot of time on Twitter interacting with people farther right than I am who think EA’s goal is to destroy all technological progress and stop AI, and people farther left than me on Bluesky who believe EA’s goals are to max out AI progress at the expense of everyone else. The truth is that most EAs think there’s a huge amount of both upside and downside to AI. If CG could, I think they’d slow the data center buildout significantly, but wouldn’t do it via spreading pseudoscience.

formally called Open Philanthropy, but changed after its close association with Sam Bankman-Fried was causing some well-earned skepticism.

This is not at all why Coefficient Giving changed its name. If they changed the name due to associations with SBF, why did it take them four years after FTX collapsed to do it? Open Philanthropy also never accepted any money from SBF.

The effective altruism movement, if we want to be gracious in calling it a “movement”, is a rabbit hole of reciprocal wealth.

Fair tbh

A lot of billionaires unload money into its related NPOs, and oddly enough, that money often goes into buying NVIDIA GPUs for an altruistic AI project, directly to Bill Gates to use philanthropically

Bill Gates isn’t involved really? I don’t know of any EA money that’s been used to buy GPUs or for altruistic AI projects?

or to grants for independent journalists and content creators to attack research that is being used to shift legislation away from things that the partners have interest in. Since it’s 2026, those things are data centers and AI.

I do wish I could get everyone telling me EAs are secretly paid to destroy AI and people who say EAs are secretly paid to max out AI progress in a room together to hash this out once and for all. The idea that EAs are paid to DEFEND AI is very funny. If you look up negative coverage of journalism funded in part by EAs, it usually focuses exclusively on how the journalism is more negative on AI than average.

But how did my infrasound content, of all things, end up as the target of a special interest group?
Well, as luck would have it, It turns out that I’m getting cited a whole lot in legislative meetings that are voting on datacenter zoning and moratoriums.
So while I’ll save my digging deep into the clusterfuck that is effective altruism for another time (I’m inspired now!), below is a quick graphic summarizing where this article came from and why it was written. Feel free to “do your own research”.

Jordan seems to have something of a fantasy of persecution. I didn’t receive any kind of command from above to write this. What happened was I was seeing more and more tweets about data center noise pollution, looked into it, thought “Wow this is terrible, this looks like a real problem” and then stumbled on Jordan’s videos. At first I only noticed a few problems, and was going to make infrasound a small end note of a post about the very real problems with noise pollution. But once I realized just how terrible Jordan was with his sources, and how popular this video was, I decided it needed its own post.

I’ve had a deep hatred of pseudoscience my whole life. In college I Turned Evil at a community meeting about the smart grid (essential for the green energy transition) where I got called a shill by crazies who had swarmed the meeting for not believing in WiFi intolerance. I decided these people absolutely could not be allowed to win anything ever. I respect real community concerns. I do not respect pseudoscience at all, anywhere. Now I get to debunk it it as my job.

The connection between EA and tech billionaires is not something I’m the first to document.

Yup, but the tech billionaires I’m associated with mostly oppose rapid AI progress and the rapid data center buildout.

Jordan shows a picture of local town meetings where he’s been brought up and says “That’s a lot of .gov action for a YouTuber who doesn’t understand science!” Well yeah, he successfully tricked people into believing his pseudoscience. Great.

I’m not going to go through Andy’s novel line-by-line because it’s a waste of everyone’s time. The entire framework of attacks like this are to “machine gun” information in a way that overwhelms the reader so they concede that the writer has an exhaustive understanding of the topic.

There was no way to document all the massive problems with Jordan’s two videos without making this post very long. A lot of people accuse me of either writing posts that are too long, or conveniently leaving things out. I suspect that if this post were shorter, Jordan would bring up every argument he made that I didn’t address. This is just a way of dodging all my very simple specific criticisms that he misrepresented every study he relied on.

1. On-screen graphics are not always citations, but sometimes may appear to be.

There’s a lot of time and tempo change between researching, writing, and editing. My infrasound videos have about 3TB of footage, sliced down into 2 videos running less than 30 minutes per piece. They may seem like elaborate deep dives, but you’re watching an entertaining overview of the actual deep dive. I spend an excruciating amount of time balancing the depth of research versus the probability that anyone will be interested in watching or learning about it. Some may disagree, but I think I’m doing a good job at this. More importantly, I’m always learning how to do it better with each video as I pay close attention to what makes viewers stick around.

My way of breezing through citations is using them as on-screen graphics. Unfortunately, not all graphics are consistent with what I’m talking about in a particular clip of the video. For example, an article making statement A may have a graphic or headline that fits better with statement B. It provides the viewer with continuity and visual confirmation, but if you comb through every on-screen article and compare it to the sentence coming out of my mouth, guess what? You’ll occasionally, if not seldomly, find that the 2 don’t perfectly line up.

But context is important. If a graphic isn’t confirming a point made in the video, that doesn’t mean the point in the video is a nefarious lie. It’s a truly wild stretch to assume that my videos are just factless voids of manipulation and Andy is the first person to catch on. Which brings me to the next segment.

This would be understandable if literally any of the articles he cited did what he claimed. It’s not that the timing was off, it’s that none of them backed up his claims at all and mostly contradicted them. This is an obvious simple dodge.

There is literally no way you can watch this section without believing that the paper he’s zooming in on that’s 100% perfectly timed with every single word he says is about infrasound, whereas in fact it’s entirely about audible sound, as explained here. This wasn’t sloppy editing, this was incredibly deliberate editing that showed a paper that didn’t at all say what he claimed in the moment. He’s just completely dodging these very obvious objections by claiming this was some cool editorial decision.

2. I fact-check. Always.

With the amount I complain about legal expenses, it’s probably not a secret that I often run my entire scripts past attorneys. And they catch a lot, sometimes transforming the video significantly, and in at least two cases, resulting in a video project being shelved. This isn’t always because of poor research, but because perfect is the enemy of good. And I need to be perfect sometimes for all the wrong reasons.

For example, I cannot drive up to a datacenter owned by the wealthiest person in history (who also happens to be a litigious asshole), pull out some equipment, measure infrasound, and make up readings or make claims about their health effects without a really good list of citations. If I make a single mistake in activities like this and when reporting on them, I’ll lose everything I have, including my channel. I do not receive any benefits of doubts when it comes to potential libel.

For months, I’ve had private investigators hired to park in front of my property. Are we to assume the company hiring them hasn’t found out about lawsuits yet?
So the idea that I’m intentionally misleading millions of people and legislators is not something I’m insulted by because in the context of my daily stressors, it’s comedy.

Well the fact-checkers clearly failed here, drastically, and Jordan hasn’t given us any counter evidence of that so far.

Throughout Jordan has what I think is a somewhat self-centered view that I’m just writing this to “stress him out” whereas in fact I would like him to stop promoting the same pseudoscience used to block wind farms. Throughout this exchange, he’s always speaking in “us vs them” language and is trying to identify me as being on a bad team instead of just looking at the facts directly. I think he knows what will happen if he does, so he’s somewhat desperate to just smear me as a tech billionaire stooge.

3. Categorical Dismissal vs. Scientific Uncertainty

Andy’s central thesis that sub-audible infrasound issues are “fake” is a definitive and hyperbolic claim that is impossible to sustain scientifically. In research, there is an infinite distance between “not yet proven without a doubt” and “fake”.
Anesthesia, consciousness, contagious yawning, and a whole library of things we experience are not completely understood. That doesn’t mean they don’t exist.

One thing that irks me is how heavily Andy relies on the “nocebo effect” (the idea that symptoms are caused by negative expectations).
Of course nocebo effect is a real confounding variable, but using it as a "catch all" to dismiss physical data ignores the actual physiological mechanisms routinely observed.

All of this happens within the confines of a world where Andy seemingly doesn’t have the ability to fact-check himself.
Recent research into mechanotransduction suggests that humans can "feel" sound through non-auditory pathways. Low-frequency noise can modulate pressure-sensitive ion channels (like PIEZO1), potentially causing neuroinflammatory responses or cellular stress even if the sound is below the threshold of "hearing" [Armand & Bikaran, 2025; MDPI, 2026].
This is just one of many examples of where infrasound is observed to be related to physiological responses that have absolutely nothing to do with one’s conscious perception or bias. “

Note: Due to IP restrictions, not all papers can be linked, only referenced.

But here is a more recent study pertaining to PIEZO1 and TRP4 channels.

It’s true that infrasound harms could technically exist, but because we haven’t found any after a half-century of serious studies looking for them all over, I’m forced to conclude that infrasound is as fake as other things in this category. Could be real, maybe? But no one (especially Jordan) has produced any evidence so far.

I’m not using nocebo as a catch-all, I’m adding it as an explanation after I show separately that all studies that hint at infrasound harms seem to be pretty quickly invalidated.

He links one MDPI paper by Dastan et al., and references a second “[Armand & Bikaran, 2025]” citation he says can’t be linked “due to IP restrictions.” Peer-reviewed science doesn’t have IP restrictions that prevent citing them, you just paste the DOI. I can’t find any paper by Armand & Bikaran 2025 on infrasound, PIEZO1, or anything like it. If anyone reading this can find it, let me know, but as it stands I’m going to treat it as made up until proven otherwise.

The MDPI paper he did link is a narrative review in Applied Sciences, not a study or meta-analysis, and not a finding of harm. It’s the same type of article as the Persinger paper Jordan sent me over Bluesky.

The paper says the following things:

1. Cellular mechanisms like PIEZO1 and TRPV4 modulation happen in experimental studies at the cellular level.

2. Cardiovascular effects in animals require “higher sound pressures.”

3. Neuroinflammation and memory effects in animals require “prolonged or intense exposure.”

4. Most importantly, this is a quote from the abstract: “Short-term studies in humans at moderate intensities have reported minimal physiological changes, with psychological and contextual factors influencing symptom perception.”

That last line is just perfectly describing my view of infrasound rather than Jordan’s. Moderate intensity infrasound (the kind people experience in their homes) doesn’t have an effect, and people’s experience of it is mostly determined by their psychology. It’s also saying that the cellular mechanisms people get excited about only produce damaging effects at high intensities over long durations in animals, which everyone has agreed on since the 1960s NASA work. Again, Jordan has linked a study that agrees with me, not him.

4. Comparing Light to Sound is Stupid (Except when Andy does it)

As much as I detest surveillance cameras indoors, I kind of wish I had one so I could see the look on my face as Andy repeatedly tried to cross-explain sound and light waves. His article does this 3 times before using the comparison of sound and light waves as a source to discredit my video.

In reality, there’s a lot of false equivalence that needs to be addressed when throwing shade on ultraviolet light comparisons. Light interacts with human tissue at a quantum/chemical level. Sound interacts with human tissue via mechanical resonance. UV light is blocked by skin while infrasound passes through solid structures with ease. This means that every human has their own resonant frequencies for their bones and organs that can be mechanically stimulated by external sources of mechanical pressure regardless of if the source is audible or not.

I already addressed this in the Bluesky section above (and had replied to Jordan on Bluesky saying exactly that, but he either didn’t read it or didn’t care). My complaint isn’t that he’s comparing light to sound, it’s that he’s comparing UV to infrasound when the obvious comparison is infrared to infrasound.

This ”every human has their own resonant frequencies for their bones and organs that can be mechanically stimulated by external sources of mechanical pressure regardless of if the source is audible or not” thing is just a restatement of VAD pseudoscience. The only way to get resonance frequencies to actually harm you is to apply enough power, and there’s just not enough power in infrasound in homes near data centers. Again, a kid on a swing has a resonance frequency for pushing them to get them higher, but if you don’t also push with enough force they won’t move back and forth at all. Same goes for your cells.

5. Wind-Turbines and Oversimplification

While not recently, infrasound research has been poisoned by special interest groups backed by energy companies trying to create arguments against wind turbines. I caught on to this very quickly years ago and had to be especially careful to avoid research papers primarily focused on wind energy.

But it’s also not difficult to do this for some very important reasons unrelated to politics or lobbying:

I’ve spent a gross amount of time playing Othello on my phone while sitting next to a device measuring wind farms and datacenters. Datacenters and cryptocurrency mines have a drastically different acoustic signature than wind turbines. Datacenters utilize high-velocity cooling fans and, in many cases, massive diesel generators in a 24/7, high-density configuration. They’re also typically much closer to residential zones. This is easily enough of a difference to segregate to 2 areas of study.

I think wind farms should be allowed to be built closer to residential zones, to the point that the infrasound from them would be louder than infrasound from data centers, because infrasound harms are fake. Would Jordan be okay with building them this close? He doesn’t give us a way to know. He’s trying to have it both ways here, but pseudoscience is like a biological weapon. You can’t use it to target hyper specific industries you don’t like and spare others. It messes with everything.

But it is tricking out, and we have very new research that directly studies datacenters. A 2026 report by the Environmental and Energy Study Institute [Communities Are Raising Noise Pollution Concerns About Data Centers] points to specific health clusters (vertigo, nausea, hypertension) in Virginia and Texas specifically linked to these industrial cooling arrays, noting that standard decibel meters (A-weighted) fail to capture the low-frequency energy that residents report that is literally “vibrating their homes”.

This article he mentions is good on real noise pollution. It’s a problem. I’ve recently been speaking to some city planner friends and have been kind of shocked at how difficult it is to regulate noise. I’ve upped noise pollution in my worries about the real issues with data centers quite a bit.

But again, this article is about audible low-frequency noise. It is not evidence of inaudible infrasound causing health effects. The Ontario environmental review board found this exact thing about wind turbines in 2011: dB(A) understates how annoying low-frequency audible noise is, which is why the EU moved to including C-weighted measurements. This article also does not support Jordan’s theory about inaudible infrasound. He seems kind of pathologically obsessed with linking articles that don’t show what he claims.

Now before we move on, it’s important to note that this “confusing concept of vibration and infrasound being related” was cited in “Health implications of the rapid rise of data centers in Virginia”, which was published the same week my datacenter video released. It also mentions datacenter water use, so I’m sure Andy has his work cut out for him in reaching to discredit it in the name of altruism.

This is one of the goated papers on data centers. It’s a reasonable mini-review that organizes a lot of the known real concerns about data centers, including air pollution from diesel generators, water stress, and audible noise pollution.

But Jordan is lying about what it says again. I’m coming away from this thinking he’s a pathological liar about research papers specifically.

The paper cites the JLARC report that Virginia data center noise is “between 40 and 59 decibels” and says this “approaches or exceeds” EPA’s 55 dBA day-night outdoor noise guideline. The health harms it then links to this noise exposure are the standard, boring, very real harms of chronic audible noise pollution: heart disease, tinnitus, hypertension, sleep disruption, cognitive and mental health effects. All of these are well-documented consequences of chronic exposure to environmental noise, and this is why I’ve said repeatedly throughout this post that data center audible noise pollution is a real problem.

The paper does not mention “infrasound” a single time, as far as I can tell. It does not mention “ground-borne vibration.” etc.

Jordan’s correct that I mostly disagree with the paper’s assessment of water issues. It basically just says that blowdown from data centers might be an issue and that they use a lot of water in the area. Both are technically true but don’t paint a good picture of the relative harm of either, which I think is pretty small, as explained at length elsewhere.

6. Infrasound and ground-borne vibration are often the same thing.

Andy seems to think that I’ve “confused” infrasound with ground-borne vibration, as if these 2 things aren’t a constant catalyst for one another. I’m actually perplexed that a former physics teacher would even argue this at the risk of discrediting themselves to anyone capable of using common sense.

I have no clue what he’s talking about here? Is this where I said that xAI probably isn’t shaking people’s homes? I have to throw my hands up and await more info.

For any 7 year olds reading, I’ll explain...

For the rest of us, as I explained in the video, it’s really hard to accurately measure the full spectrum of infrasound because [long explanation of microphone diaphragm shelves, vibration sensor weak spots, and Fourier harmonics].

Yeah it’s hard to measure, cosign.

7. The decibel problem.

I’ve already explained how challenging it is to get a quality infrasound measurement… Solving this problem is what I immediately turned my attention to when finishing my last infrasound-related video. I believe that there may be a way to accurately capture the full spectrum of pressure waves via infrared optical sensors and a collimated light source inside a self-contained portable unit. It’s something I’m actively working on. But to be clear…

This is a logistical problem, not evidence of pseudoscience.

This is a non-response to what I said.

Building a new portable infrasound sensor is completely fine. What I take issue with in his video is his experimental design was badly confounded by priming, un-blinded operation, a subwoofer he acknowledges made the room physically vibrate, and post-hoc removal of subjects who reported detecting the condition. None of those problems go away if he builds a better microphone. He could have a perfect measurement of the infrasound in his experimental room and his conclusions would still be unsupported for every reason I gave.

So when Andy read the words “low-quality data” in the World Health Organization’s analysis of infrasound exposure research, he didn’t seem to realize that the “low quality” was in the consistency of measuring the infrasound, not analysing physiological effects.

??? I didn’t quote any WHO “low-quality data” language in my post, and he didn’t link where he thinks I did. If he means the 2018 WHO Environmental Noise Guidelines, the WHO did say the evidence base for health effects at the frequency range they considered was low-quality in the GRADE sense, which is a technical term referring to overall confidence in effect estimates. The issues flagged were about study design, sample size, controls, risk of bias, etc.

His move is to frame every negative finding in the field as a measurement problem and every positive finding as real.

8. This “hit-piece” is good news.

The fact that someone is paid to spend an enormous amount of time pedantically combing through hours of video, reaching in every possible direction to discredit and insult me, my viewers, the cited researchers, and the people whose lives have been devastated by what appears to be infrasound means that our message has reached enough of the general public and their legislators to justify a low blow.

Classic conspiracy theorist move here. “They’re only criticizing me because the cabal doesn’t want me to win.” This is Alex Jones but higher status.

Coefficient Giving didn’t call me one day and say “take this guy out.” I’ve been posting about AI and the environment for a year and a half and now have a no strings attached grant from them to write more. I decide what I want to write about, and I despise pseudoscience and hate when people use stupid tricks like this to push it. I want it gone from the public debate.

Nobody is bringing a tablet to city council meetings and playing my videos. The presented information inspires them to roll up their sleeves, pick and choose what is reputable and relevant to their communities, and then present it formally to legislators.

Inshallah they actually read what the studies say.

Why doesn’t this organization just spend this money on researching and mitigating these things?

Like, wouldn’t actually lowering the water use of a datacenter help legislation pass in their favor?

Wouldn’t actually researching infrasound sources and reducing it where there’s growing evidence of harm make this all go away?

Lowering the water use of a data center often means increasing energy used in cooling, which means more air pollution and greenhouses gases. Don’t do it!

Again with the plea to take his dumb pseudoscience that he knows is fake seriously.

I’ve measured sound sources in hundreds of places, designed low-cost measurement devices, and interviewed hundreds of people suffering from something that I believe to be 100% real. And I’ve done it all on a shoestring budget provided by my videos and viewers under the umbrella of a nonprofit of which I do not claim a salary.

Popular YouTuber funds his own pseudoscience investigations. Hero!

What angers me the most about Andy’s post is that I could have spent this afternoon working on that optical microphone, but instead I’m feeling pressured to defend research that already holds its own ground.

One of the ways tactics like this are effective is by providing a distraction from the research.

I think Jordan should feel some pressure to not abjectly lie to his big audience yes. I’m the one defending the research from his crazy constant misrepresentations.

Response to “Andy Masley doesn’t understand how sound works”

Jordan also shared a blog from someone else claiming that this post misrepresents this science. The post was pretty bad and I wasn’t going to respond to it until Jordan shared it, but since he did I felt like I should put something together. Here’s my complete response (in another post):

There’s a lot of social permission to treat data centers as boogeymen right now

Misinformation is not always easy to identify based on the speaker. If someone looks and acts like Alex Jones, my alarm bells go off, but if someone presents themselves as a well-rounded everyday person who just wants to criticize giant powerful corporations and does it in a chill way, it’s much easier to nod along and not be as alert to ways they might be wrong. More importantly, a lot of people don’t seem to think it’s even worth worrying about the specifics of how they might be wrong if they’re targeting bad people. Who cares about the specific ways data center critics are getting things wrong? Data centers are evil. Jordan is clearly taking full advantage of this and knows that this will get him lots of views.

I think wind farm misinformation is obviously way way worse than data center misinformation. I’d much rather build a wind farm than a data center. But there are two key points here:

• Promoting pseudoscience is like releasing a biological weapon. It’s going to have broad unexpected effects on everything, including things you might like and support. I do not want more infrasound pseudoscience in the world. It will make a lot of people a little more dumb and neurotic and unable to consider real trade-offs well, and cause them to oppose a lot of things I want to see built out, especially wind turbines.

• Decisions about the environment are incredibly complicated, and involve a huge amount of trade-offs. If a pharmaceutical plant or a car factory is proposed near you, it’s important seriously weigh both the real upsides and downsides of both. There are places where building either could help a local community, and places where it can harm it, and getting the real benefits and avoiding the real harms involves understanding very nuanced details of where specifically the harms and benefits are. Throwing pseudoscience into the conversation just incapacitates people from making good decisions. I don’t want people to think of data centers like wind turbines, but I do want them to think of them like they do pharma plants or car factories: things with upsides and downsides that might in specific places be good or bad for local communities.

I want to ask Jordan and his audience “Why do you feel the need to be a goofy goober about this? There’s plenty of trade-offs to worry about with data centers and AI. xAI’s Colossus is a great example of a way a data center harmed people with air pollution, a real and terrible general problem. Why not just make a video about that? Why flatly make things up? Don’t you want to empower your side to know where the actual problems are? Don’t you want to treat your audience like friends and adults, who you wouldn’t lie to and want to empower to think seriously about the world?”

I’d like more people to think of pseudoscience and misrepresenting studies the way Batman thinks about guns.

As data centers become boogeymen for more people, who start to see them as encapsulations of everything wrong in society more broadly, there’s going to be more and more incentive to add on more fake issues, and I think this is mostly just going to incapacitate people from thinking more seriously about the downsides and upsides of the largest industrial buildout of my lifetime and maybe the most consequential new technology.

• “Training GPT-3 produced 552 tonnes of CO2, equivalent to driving 112 gasoline powered cars for a year” — Columbia Climate School

• “GPT-4’s footprint is roughly equal to the annual emissions of 1,550 US citizens” — Shop Without Plastic

• “GPT-3’s emissions are equivalent to the lifetime emissions of 8 cars — or 109 cars’ yearly emissions” — Truthout

I’m going to argue here that 1) The comparisons people make with training are often silly and misleading, and 2) When you make fair comparisons between training AI models and creating other products, the cost of training does not look unreasonably large at all. The way this is talked about often involves comparisons that would make literally any popular consumer product look ridiculous and wasteful. I think these comparisons are obviously goofy, and if I make more reasonable comparisons while keeping the numbers the same, you will see that training AI models is not some unique environmental catastrophe, and actually just blends into all the other ways society uses energy normally.

Contents

• What is training?

• Some bad comparisons

• Two useful comparisons

• A lot of popular coverage of this used terrible comparisons

• State of the art training runs for current models are larger, but don’t emit enough to change my point here, and they have way more users

• Should I also include all the other climate costs of training, like the physical infrastructure and failed training runs?

What is training?

Training an AI model basically means creating it. Untrained models are fed huge amounts of data to pick up more and more subtle background patterns. This involves connecting tens of thousands of specialized AI computers very close together to constantly communicate with each other over months, typically all housed in a single large data center. This uses a lot of energy.

Once a model is trained, it can be used over and over and does not need to be retrained. Newer more capable models are trained instead. Models are often used for months before being replaced by newer ones.

Importantly, spending this much energy is a necessary step to creating frontier chatbot models. We currently do not know of a way of creating models with similar capabilities using much much smaller training runs.

This uses a lot of energy by the standards of everyday buildings, but my argument here is that it doesn’t use much energy or emit that much by the standard of creating any other popular products.

I see training as being analogous to manufacturing a new product for users to buy. Just like physical objects people buy and use have to be manufactured first, AI models people interact with have to be trained first. Both are necessary steps in creating the final products the users interact with.

Some bad comparisons

Your personal emissions

Take a look at this graph, it and graphs like it have been very popular as a way of showing how bad AI is for the environment:

Importantly, this graph is about a very old 2019 model (way before commercial chatbots were available) that didn’t get much use. The stats from this old unused model made it into articles about AI’s environmental footprint 5 years later. Yet it still gets shared a lot to make points about current consumer chatbots. I’ll update the graph with the much much larger estimated training cost of GPT-4. I’m going to start with GPT-4 as an example, because this is the first model where everyone started freaking out about how much energy training uses, and how much it emits:

Much bigger!

I’m using a high-end estimate of GPT-4’s training, partly to include things like the cost of failed training runs. This cost is 15,000 tonnes of CO2.1

Is this a reasonable way to think about GPT-4’s emissions? Is it a useful comparison?

How reasonable does this other comparison look, where instead of training GPT-4 the bottom line shows my rough guess for the emissions of manufacturing every new copy of the iPhone 16?2

Does this tell you anything about whether you personally should or should not purchase an iPhone?

This is my big claim: It is not useful or reasonable to measure the emissions from the creation of a product used by hundreds of millions of people every day against your personal emissions, or the emissions of buying a single plane ticket. That tells you basically nothing about how damaging it is for you personally to use the product, unless you divide by the number of users.

This seems obvious, but almost all reporting on how much energy AI training runs use compare it to individual things people do. A better comparison is the cost of creating the AI model to the cost of creating other products used by hundreds of millions of people. When you do that, it becomes difficult to understand why the emissions of training AI models are receiving so much scrutiny:

Energy of cities

Another common comparison people make with AI training runs is the electricity costs of whole cities. For example, the MIT Technology Review notes that training GPT-4 “consumed 50 gigawatt-hours of energy, enough to power San Francisco for three days.” Is this useful? Well, the electricity homes, businesses, and industry in a single city use doesn’t give us too much useful context for how GPT-4 compares to other consumer products.

Again, take manufacturing iPhones. Manufacturing all iPhone 16s emitted approximately 12 million tonnes of CO2.3 Generating San Francisco’s electricity emits about 890,000 tonnes of CO2 per year. So all iPhone 16s emitted as much as 13 years of San Francisco electricity.4 Does this tell us much about whether you personally should buy an iPhone? Does it add much useful context at all?

Flights

One other common comparison is airplane flights, measured as the total plane’s emissions rather than emissions per capita. GPT-4’s training emissions were approximately as much as 38 full plane flights from San Francisco to Australia5.

NeurIPS is the largest annual ML/AI research conference. This year it’s being hosted in Sydney, Australia. Based on past numbers I’d estimate 15,000 people are going to travel to the conference this year6 from all around the world, enough to completely fill 50 planes. If the average attendee is flying from elsewhere, like San Francisco, this means that a single annual AI research conference will emit more than training GPT-4.

Imagine if an AI research conference was held, and it was so useful for attendees that it singlehandedly created GPT-4 from scratch, where only GPT-3.5 existed before. This seems like a pretty worthwhile get-together. Here, I don’t think people would even bring up the emissions. Right now few seem to be commenting on the emissions of the conference. I think that a headline “AI company emits as much as 38 planes to create GPT-4” is much more attention grabbing than “AI researchers from around the world take 50 planes to all hang out in Australia” even though the second is much worse for the climate and has likely been significantly less consequential for AI as a field, or everyday people’s lives.

Two useful comparisons

Total: Other products

Training is a necessary part of creating a product that hundreds of millions of people interact with every day. One way we can compare training AI models is to the emissions of creating similarly popular consumer products that hundreds of millions of people interact with. Here’s a graph of the total emissions of training GPT-4 vs my best estimates of emissions from the production of other consumer products:

Just like physical CDs and cases had to be manufactured for hundreds of millions of people to play Grand Theft Auto 5, GPT-4 had to be trained before hundreds of millions of people could use it. Both are necessary steps in making a popular product hundreds of millions of people use every day.

Per-user: CDs

I’m old enough to have memories of going out and buying CDs for computer games I wanted to play. One of the coolest objects I owned as a kid was a CD encyclopedia. The whole world’s knowledge at my fingertips!

There are no exact numbers for how many people were using ChatGPT over the time that GPT-4 was available, maybe somewhere around 100-200 million weekly users is a reasonable guess. Dividing GPT-4’s cost of training by this number of users gets to 75-150 grams of CO2 emitted by training per user. This happens to be a little lower than the emissions of manufacturing a single CD. It’s as if each weekly user of ChatGPT had purchased a CD to use it.

This is the cost of training everyone was freaking out about when GPT-4 was released. It’s like the company didn’t have to spend any energy on training at all, and to use ChatGPT you just had to buy it on a CD, which emits a little to manufacture. I don’t think that if GPT-4 didn’t involve training at all but came on a CD, anyone would be talking too much about the emissions of manufacturing those CDs, and yet those emissions are the same as the training cost that people speak so ominously about. This is very small!

A lot of popular coverage of this used terrible comparisons

Here’s a ton of quotes from different articles on the cost of training GPT-4 and previous models. Not one of them compared training to other consumer products used by hundreds of millions of people, or divided by the number of people to get a per-user number. I haven’t found any sources doing either. Read through these and think about whether the reader would come away with a better or worse understanding of the magnitudes involved. Would they know creating a model used by hundreds of millions of people had about the same climate impact as the CDs and cases for Grand Theft Auto 5?:

• “Training GPT-3 produced 552 tonnes of CO2, equivalent to driving 112 gasoline powered cars for a year” — Columbia Climate School

• “GPT-4’s footprint is roughly equal to the annual emissions of 1,550 US citizens” — Shop Without Plastic

• “GPT-3’s emissions are equivalent to the lifetime emissions of 8 cars — or 109 cars’ yearly emissions” — Truthout

• “That’s like a car driving 1.2 million miles” — Carbon Credits

• “500 tons of carbon dioxide emissions is the equivalent of around 600 flights between London and New York” — Cybernews

• “GPT-3 emitted carbon dioxide equivalent to 500 times the emissions of a New York-San Francisco round trip flight” — Analytics Vidhya

• “It’s the same amount of emissions as a single person taking 550 roundtrip flights between New York and San Francisco” — Carbon Credits

• “Training GPT-4 consumed 50 gigawatt-hours of energy, enough to power San Francisco for three days” — MIT Technology Review

• “15,000 tons is roughly the same as the annual emissions of 938 Americans” — Towards Data Science

• “Enough energy to power an average U.S. home for over 120 years” — Truthout

• “Training a single AI model can emit as much carbon as five cars in their lifetimes” — MIT Technology Review (headline)

• “Training GPT-3 consumed 1,287 megawatt hours of electricity (enough to power about 120 average U.S. homes for a year), generating about 552 tons of carbon dioxide” — MIT News

• “Training a single large model like GPT-3 can use over 1,200 MWh — enough electricity to power around 120 U.S. homes for a year” — Climate Impact Partners

• “Training GPT-3 emitted roughly 500 metric tons of carbon dioxide — the equivalent of driving a car from New York to San Francisco about 438 times” — Climate Impact Partners

• “Training GPT-3 produced 552 tonnes of CO2, which is equivalent to the emissions from 110 gas-powered cars over a year” — Smartly.AI

• “GPT-3 emitted carbon emissions equivalent to the lifetime impact of five cars” — Nature Scientific Reports

• “Training the bigger, more popular AI models like GPT-3 produced 626,000 pounds of carbon dioxide, equivalent to approximately 300 round-trip flights between New York and San Francisco — nearly five times the lifetime emissions of an average car” — The Sustainable Agency

• “This is similar to the yearly emissions produced by 120 passenger cars or 600 transatlantic flights per person” — Shop Without Plastic

• “Training AI models can emit more than 626,000 pounds of carbon dioxide equivalent — nearly five times the lifetime emissions of the average American car (including the manufacture of the car itself)” — Learning Tree

• “Training AI models can emit more than 626,000 pounds of carbon dioxide equivalent — nearly five times the lifetime emissions of the average American car” — Supermicro

• “Training a single AI model can emit 626,000+ pounds of CO2 equivalent… about 5x the lifetime carbon emissions of an average passenger car” and “training a single AI model can consume more electricity than one hundred American homes use in one year” — Carbon Credits

• “Training such a model requires… This is equivalent to the annual energy consumption of around 160 average American homes” — Frankly Kranky

• “GPT-3 consumed approximately 1,287 MWh of electricity—enough to power 120 US homes for a year… A single AI model training session can emit as much carbon dioxide as five cars over their entire lifetimes” — AI Energy Calculator

• “Training ChatGPT-3 is estimated to have required the equivalent energy consumed by an average American household for 120 years” — California State University Northridge Library Guide

• “GPT-3 consumed a staggering 1,287 megawatt-hours (MWh) during training. This resulted in about 502 metric tons of carbon emissions, equivalent to the emissions from hundreds of gasoline-powered cars in a year” — PatentPC

• “The process can emit more than 626,000 pounds of carbon dioxide equivalent — nearly five times the lifetime emissions of the average American car” — ACM Communications and separately SPE Journal of Petroleum Technology

None of these comparisons make sense, because basically all of them treat training as if it’s comparable to activities individual people do, instead of creating a product expected to get hundreds of millions of people using it. If instead we treat training GPT-4 like this, it fades into basically nothing. If you think it would have been worthwhile for the climate to prevent GPT-4 from ever being trained, you should also consider stopping the production of Lego for 2 weeks. If you’re not telling everyone to boycott Lego or Coke or Grand Theft Auto because of the emissions involved in creating them, you shouldn’t be telling people to boycott ChatGPT for the emissions involved in creating GPT-4. If you read “Manufacturing all the CDs for Grand Theft Auto 5 emitted as much as 160 American homes” I think you might be a little underwhelmed.

State of the art training runs for current models are larger, but don’t emit enough to change my point here, and they have way more users

Everything above uses GPT-4 as an example because that’s the model that kicked off all the panic about training emissions. But GPT-4 is old now. Current state of the art models are bigger and more expensive to train. Does this change anything?

Training costs are growing at about 2.4x per year for frontier models. The largest training run we have decent estimates for is Grok 4, which Epoch AI estimates used about 310 GWh of electricity and cost around $490 million, roughly 6–7x what GPT-4 cost. But Grok 4 is probably the worst-case scenario for training emissions, because xAI’s Colossus data center in Memphis was powered largely by mobile natural gas turbine generators, which emit about 0.49 kg CO2 per kWh, 1.3x higher than the US grid average of about 0.37 kg CO2 per kWh, which itself is higher than some of the grids where AI models are typically trained, like Oregon which relies more on hydropower. This is why Epoch estimates Grok 4 emitted about 154,000 tonnes of CO2, ten times GPT-4’s estimated emissions, despite using only about six times as much electricity.

Most other frontier labs don’t train on natural gas generators. They primarily train models in large data centers connected to the electrical grid, many of which are in places with more renewable energy. The same 310 GWh Grok 4 training run would emit about 115,000 tonnes on the average US grid, 75% of its actual emissions. On California’s grid it would only emit about 62,000 tonnes, 40% of its actual emissions. This means that a frontier training run of similar size to Grok 4, but run on a typical major cloud provider’s infrastructure, would likely emit somewhere in the range of 60,000–120,000 tonnes.

Not every frontier model uses as much compute as Grok 4, either. Grok 4 was trained on 200,000 GPUs and is probably the single largest training run to date. Other state of the art models may use significantly less compute, bringing their energy use and emissions lower. Hardware energy efficiency has been improving at about 40% per year, so the next generation of training runs won’t need proportionally more energy even if they use more compute.

So what’s a reasonable range for the emissions of training a current frontier model? Probably somewhere between 40,000 and 120,000 tonnes of CO2, with the wide range driven both by differences in the carbon intensity of the electricity and differences in the amount of energy used.

Is this a lot? Well, these models get used for months at least. I could just up the manufacturing numbers to a month each.

AI models still just don’t stand out much compared to other popular products.

For now I’ll use 80,000 tonnes as a round middle estimate for the current emissions of training an average frontier model.

Two things have also changed since GPT-4:

1. There are way more users. When GPT-4 was the main model, ChatGPT had roughly 100–200 million weekly users. Now ChatGPT has 900 million weekly active users. That’s roughly a 5–9x increase in users over the same period that training costs went up maybe 4–6x. The per-user training footprint has likely stayed roughly flat or even declined. 900 million users over 80,000 tonnes of CO2 for training is about 89 grams. This is still about half a CD. Even under Grok 4’s worst-case emissions of 154,000 tonnes, assuming similar user numbers would mean the per-user cost is about 171 grams, roughly one CD.

2. Even at the total level, the comparisons mostly don’t change. The most advanced AI models that huge numbers of users interact with for months at a time use about as much energy to create as it takes to manufacture 3–4 days of iPhones.

My basic point has so far not changed with larger models: creating an individual AI product used by hundreds of millions of people produces emissions that are small relative to manufacturing basically any other product used by hundreds of millions of people.

Training runs are going to continue to use more energy, but models are also getting more capable over time. Eventually this may be more like each user purchasing a piece of computer hardware instead of a CD, but for models so much more capable the investment will, I expect, be worth it.

Should I also include all the other climate costs of training, like the physical infrastructure and failed training runs?

Maybe my comparison to creating other products is unfair here, because there are lots of other analogous steps to manufacturing before the chatbot can be delivered to hundreds of millions of people. OpenAI also had to buy tens of thousands of GPUs, which had to be manufactured in fabs and assembled into servers. Those servers sit in data centers that had to be built out of steel and concrete. What’s the total carbon footprint of all that?

Let’s try to make this number as big as possible. We can include the embodied hardware emissions, the data center construction, the cooling infrastructure, and the experimental training runs that failed before the final model worked. Hugging Face found that when they accounted for all of these factors for the model BLOOM, total lifecycle emissions roughly doubled, from 25 tonnes to about 50 tonnes. The estimate I chose for GPT-4’s training cost is 3 times as high as many popular estimates already, so it likely includes these costs already, but let’s double it anyway to be safe, to get a total of about 30,000 tonnes of CO2 emissions from literally everything required to deliver a model to users.

Even at this extreme, training GPT-4 would still be smaller than building a single cruise ship. It would still be less than one day of manufacturing Nike shoes. It would still be less than a single day of manufacturing iPhones.

But I also didn’t include many of the additional costs of the other products too. I didn’t include building the factories or making the mining equipment for extracting raw materials or transporting iPhones to stores or the cost of the retail stores themselves… many accusations that “you’re leaving information out” are one-directional, and don’t consider that most things we do have lots of hidden costs depending on what we want to compare. Ultimately I think the direct comparison between training and creating other popular products is completely legit.

The per-user comparison doesn’t really change even under the most extreme interpretation. It raises the per-user cost of training GPT-4 from one CD to two. The reason the cost of creating a product used by so many people is so low compared to things like Coke and sneakers is that in comparison, computing is very energy-efficient, and thus it’s very rarely the most promising thing to optimize or cut for the climate.

It’s true that all training across all AI models is forecast to make up a significant fraction of US electricity use in the next few years, maybe rising to whole percentage points of US electricity usage. This is because a huge portion of total AI training will be concentrated in America, and because AI is entirely electrified, while most other industries aren’t. This post was mainly about the training of individual AI models. In the next post I’ll try to paint an overall picture of where training and AI more broadly fit into America and the world’s overall energy and climate situation.

I’m writing this as an emergency post because a new contender for drastically wrong common wisdom has just popped up.

There’s a new paper being shared and reported on called “The data heat island effect: quantifying the impact of AI data centers in a warming world.” It is the single worst writing and research on AI and the environment that I have read.

The central claim of the paper is that in the places where data centers are built, land surface temperatures increase by an average of 2°C (3.6°F) after the data center begins operations, creating what the authors call a “data heat island effect” that affects over 340 million people worldwide. In some extreme cases, data centers are found to increase land surface temperature by up to 9°C (16°F). The paper claims that a warming effect can be measured up to 10 km from the data center. This all sounds really bad! 340 million people affected by heat islands.

The paper also has alarming looking graphs that I suspect will do well in the media coverage:

There are three problems with this paper:

• The methodology is completely off-the-wall goofy and the study does not at all show what the authors claim. This will be the focus of this post. The authors very strongly imply this temperature change is due to data center waste heat, but I think it’s almost certain that it’s instead entirely due to the fact that the surface of any building, a data center or Wal Mart or house or Starbucks, is hotter than the grass that was there before. I think I can show conclusively using high school physics and math that the data centers are not heating the land around them. Data center heat exhaust could only heat land by at most 1% of what the authors imply. The paper itself doesn’t even consider this very obvious reading as a possibility.

• The paper’s framing is wildly misleading and will create a lot of additional confusion on top of the methodology itself being bad. The author’s suggestions from this paper is to “build more energy efficient AI algorithms” but the actual conclusion should be “Don’t build buildings if you don’t want those buildings themselves to get a little warm, even though they won’t affect anything around them.” There is a lot more incredibly shifty language the authors use over and over.

• I’m worried that the media takeaway from this is going to be that data centers are warming the air this much. If they were, that would be a complete disaster. To put this in context, if (another) data center were built near me in DC, and it heated the air by an average of 9°C, this would change the DC climate to the climate of Miami. A 2°C would be equivalent to moving DC halfway to Georgia, it would be like the region had fast forwarded through 50-70 years of climate change. In reality, the data centers are being measured to warm the land around them, which is very very different (explained below). It seems likely that a lot of people will misinterpret this. Look again at the graphs from the paper. The Y axes say “Normalised temperature increase,” not “Normalised land temperature increase” and if they’re shared without context, I suspect people will draw the wrong conclusions.

Contents

• The paper

  • A summary of my argument

  • What is land temperature?

  • The methodology

  • My core claim: this is literally just measuring hot surfaces of new buildings, and the soil and land around those new buildings probably hasn’t changed temperature at all

  • How much could a data center’s waste heat raise the land temperature around it?

  • The one remaining question: why the clean step change?

  • Other candidates for warming land

  • A few other bizarre methodological mistakes

  • The incredibly shifty framing the authors use

• The media is already running with this

• A parting thought: AI is useful for research

The paper

A summary of my argument

• The measurements in this paper are just showing the fact that buildings are hotter to the touch than grass. The data centers are most likely not heating the land around them. They themselves are just hot because they are buildings. If you include hot buildings in a very general measurement of the average total surface temperature of a large area, that average will go up slightly without telling you anything about the undeveloped land.

• Using simple high school physics, it’s easy to show that there is no way data center heat exhaust could account for more than at absolute most 3% of the observed average land heating, and any realistic assumptions drive that down to basically 0. The only other mechanism by which they could be heating the land fails. This leaves the other explanation: the data centers themselves are hot because they are buildings, the land around them remains at the same temperature, unless more buildings have been built on top of it to accommodate the data center.

• You would think the authors would have anticipated my criticism, but they don’t even consider it as an issue with their methodology at all.

What is land temperature?

NASA defines land surface temperature (LST) this way:

Land surface temperature is how hot the “surface” of the Earth would feel to the touch in a particular location. From a satellite’s point of view, the “surface” is whatever it sees when it looks through the atmosphere to the ground. It could be snow and ice, the grass on a lawn, the roof of a building, or the leaves in the canopy of a forest.

We need to understand land temperature as opposed to air temperature. They can be very different from each other. An obvious example is stepping onto black asphalt on a summer day. Even though the air temperature may be warm, if the air were as warm as the asphalt you’d be dead. Stepping onto a light-colored concrete sidewalk instead would be much more cool. Land temperatures can deviate a lot from the temperature of the air.

Changes in land temperature of 2-9°C don’t really impact human welfare directly. What matters for human welfare is air temperature. This temperature range invokes the low to very extreme catastrophic ends of the spectrum of what could happen with climate change, but that spectrum is about air temperature. Land temperatures change all the time without a corresponding significant effect on air temperature. If the most extreme land temperature changes happening in the paper were happening around you, the air around you would likely warm by a fraction of a degree at most. It would probably too small to measure or distinguish from normal day-to-day temperature variability.

How much does land temperature matter for the ecology of a region? Soil temperature directly affects plant health, microbial activity, moisture retention, and which organisms can survive in an area. If data centers really were warming the land surface by 1°C out to 4.5 km away, as the paper claims, that would be worth paying attention to ecologically even if it didn’t matter much for the people living nearby. A 1°C rise in soil temperature is pretty minor, but a 9°C would change which plants and microbes could survive. This would be a problem if it were real, but as I’ll try to show below, it’s not.

The methodology

The researchers used satellite data from NASA satellite sensors to measure land surface temperature at data center locations worldwide from 2004 to 2024. They looked at how land temperature changed at each site before and after the data center began running, and found a clear jump of about 2°C on average right at the moment the data center came online, as shown on this graph (the orange region shows the range from the very minimum to maximum measured temperature changes, the red line is the average, and the error bars represent the 95th percentile bounds, meaning the vast majority of measurements fell within those much smaller bounds):

The paper also claims that this effect extends up to 10 km from each data center, with a measurable 1°C land temperature increase as far as 4.5 km away using the same MODIS sensor:

The authors present this as evidence that data centers create broad climate zones comparable to urban heat islands.

The authors are somewhat ambiguous about what’s causing this temperature increase. They say that their study “relies on the assumption that AI hyperscalers might have an impact on the LST of their locations because of the heat that they would release as a result of the high power demanding applications.” So this makes it sound like they mean to imply that it's the heat exhaust of the data centers (evaporated water or from air cooling systems) that's causing this increase. However, they never explicitly state this, and they never consider the alternative explanation that the temperature increase is simply caused by the construction of the building itself. The possibility that replacing grass with a large building and parking lots might explain their results is completely absent.

Their mitigations section confirms what they think is going on, though. Their proposed solutions are all reducing operational energy use. They recommend more efficient AI algorithms, adiabatic circuits, and passive radiative cooling coatings for data center buildings. These are solutions to the problem of “servers produce too much heat” not “a building was built.” If the authors believed these numbers were primarily driven by land cover change, the solution would be “paint the roof white” or “plant vegetation around the perimeter.” Instead, they propose redesigning AI models themselves. This tells you they believe operational waste heat is the cause, even though their methodology can’t actually support that conclusion. But their methodology is goofy goober.

This is in some ways worse than if they’d made an explicit causal claim, because there’s nothing specific to argue against. The most obvious alternative explanation for their findings, that buildings are hotter than fields, isn’t mentioned in this paper. They don’t control for it as a possibility.

My core claim: this is literally just measuring hot surfaces of new buildings, and the soil and land around those new buildings probably hasn’t changed temperature at all

My argument here is going to be that this paper is just measuring the fact that buildings are hotter than grass. The land around the data center that isn’t either buildings or road has probably not warmed. It doesn’t matter at all that this study is about data centers. I’d expect similar measurements for Wal Marts.

If you build a building, whether it’s a data center or Wal Mart or house or Starbucks, and you point a NASA satellite at the building, you are going to measure a warmer surface temperature exclusively caused by the material the building made of, not any heat exhaust from what’s happening inside the building.

However, the difference between the surface temperature of buildings and grass is much larger than study’s finding of 2°C. It’s on the order of 20-35°C on a sunny day. Why is the study’s number so much smaller? The satellite sensors this study used (MODIS) can’t actually look at an area as small as a data center. MODIS pixels are roughly 1 km² (or 0.5 x 0.5 km in the enhanced product the paper claims to use). A data center and its parking lots might occupy a fraction of that pixel, with the rest still being undeveloped land, roads, or vegetation. So what the satellite sees is a blended average, maybe 10-20% data center and parking lot surface, and 80-90% unchanged land.

Let’s assume buildings and roads are on average 20°C warmer to the touch than grass. To get an average land temperature change of 9°C (the most extreme outlier in their data) on a 0.5 x 0.5 km unit of grass, you would need to build 112,500 m² of new building and road surface, about 28 acres. Many hyperscale data centers are larger than this. For example, QTS Atlanta's campus covers over 95 acres, DC BLOX's Atlanta West campus is 55 acres, and the Stargate campus in Michigan will occupy 250 acres. Even a typical hyperscale data center occupies around 10 acres of building footprint alone, before accounting for parking lots, substations, access roads, and security perimeters, which can easily double or triple the total paved surface. 28 acres is a mid-sized hyperscale campus. This on its own would explain the most extreme outliers.

What about the 2°C average they find? To get a 2°C increase, you would only need about 25,000 m², or roughly 6 acres of building and road surface within the satellite's field of view. That’s a pretty normal size for a non-hyperscale data center.

What about the increase in temperature farther from the data center, up to 10 km away? Well, two things:

• Data centers are mostly not built in very isolated places. They need power infrastructure, fiber connectivity, and road access. Places with this stuff are also attractive for warehouses, logistics hubs, office parks, and residential development. When you look at the area within 10 km of a data center, you’re typically looking at a region that has a lot of new building in general, and that all registers as warmer to a satellite’s thermal sensor, regardless of what the buildings are used for. The paper’s smooth spatial decay curve is exactly what you’d expect from a development footprint that’s densest near the infrastructure hub and thins out with distance.

  

  Importantly, this is not what you would expect if the effect were being driven mainly by waste heat from a single facility. Waste heat in air does not spread outward in a neat, slowly fading circle. It behaves more like a plume: it rises, mixes, and disperses through an ever larger volume of atmosphere, causing its temperature effect to fall off in a strongly nonlinear way with distance. The exact decay depends on wind, turbulence, exhaust height, and mixing conditions, so there’s no single universal graph here. But we know for sure that if this were really a direct waste-heat signal, you would expect a pronounced curved drop-off, not a broad, almost linear decline over many kilometers like what we actually see. The paper’s red line looks much more like a development gradient than a plume. A smooth, nearly linear decrease is exactly the sort of pattern you would expect if the land closest to the data center were also the land most likely to pick up new roads, roofs, parking lots, warehouses, and other warm impervious surfaces, with that development footprint thinning gradually as you move outward.

• Data centers also frequently trigger or coincide with significant new infrastructure construction in the surrounding area. A hyperscale data center consuming 100 MW or more needs dedicated power infrastructure. These are large industrial installations with their own footprints built within a few kilometers of the data center they serve. Access roads also need to get built or widened. Fiber corridors are trenched. Supporting commercial services like equipment suppliers, security contractors, food service for construction crews and staff set up nearby. The data center brings an ecosystem of development with it. All of that replaces vegetation with built surfaces, and all of it would register as increased land surface temperature to a satellite, having nothing whatsoever to do with the thermal output of the servers.

The paper doesn’t control for any of this. They don’t compare their data center sites against other types of commercial construction. They don’t attempt to separate the surface change from operational heat output. They don’t even discuss the possibility that what they’re measuring is a land cover change rather than a thermal pollution effect. The entire paper treats the LST increase as if it were caused by the servers running inside the building, when the far more parsimonious explanation is that it was caused by the building itself. And if it is caused by the building itself, there’s no reason to expect the land around it to be heating at all, for 2 reasons:

• Land mostly doesn’t absorb heat from nearby buildings. If you step from a hot asphalt parking lot into cool grass, that grass is likely the same temperature as grass a mile away from the asphalt.

• If the land were heating up, this would imply that the buildings themselves would need to be much cooler to keep the average as low as it is. Data centers and the roads and infrastructure around them would need to be significantly cooler than all other normal buildings, or just be very small, and we know they’re not small.

There’s some very minor effects big buildings have with the flow of air and wind around the land around them, but it’s pretty marginal, and again this would change the study to be about any large buildings at all.

I think I already have a strong case here, but I have some high school physics and math in the next section that in my opinion completely confirms it.

How much could a data center’s waste heat raise the land temperature around it?

MODIS satellites take their measurements around 10:30 AM and 1:30 PM. During the day, the sun is shining directly on the ground. The ground absorbs solar radiation and gets hot. The air above it is cooler than the ground surface. This means heat is basically always flowing upward, from the ground into the air.

Data centers emit heat into the air. The way they would heat the land is by raising the temperature of the air enough that less heat from the ground flows into the air, and stays in the ground instead. The colder the air, the more heat is lost from the ground, and the warmer the air, the more heat stays in the ground. It’s like if you touch an ice cube to your hand, and then touch a lukewarm object to your hand. In both cases, your hand is losing heat energy, but it loses it much faster when in contact with the ice cube.

How much would the data center need to heat the air to cause significant amounts of heat energy to stay in the ground?

The math (skip if you’d just like the conclusion)

MODIS is measuring land surface temperature. The graph in the study shows that over an area of 1.6 km, land temperatures were elevated by 1.5°C. Could a data center’s waste heat plausibly deliver enough extra heat to keep the surface around it about 1.5°C warmer over a radius of roughly 1.6 km?

Consider a large hyperscale data center drawing 100 MW of power. Nearly all of that electricity eventually becomes waste heat, that’s 100 million joules every second leaving the facility. A circle with a 1.6 km radius around the data center would have an area of about 8 million square meters.

Imagine that every single watt of heat the data center produces somehow makes it to the ground surface within that circle, spread perfectly evenly. None of it rises into the upper atmosphere. None of it blows beyond 1.6 km. Every joule goes straight into the ground. How much heating would each square meter of ground receive?

100,000,000 watts ÷ 8,000,000 m² ≈ 12.5 watts per square meter

Now, does 12.5 W/m² produce 1.5°C of surface warming? No, because the ground doesn’t just absorb heat and get hotter forever. As the surface warms, it loses heat faster, it radiates more infrared energy and sheds more heat into the air through convection. The surface will only warm until the extra heat loss matches the extra heat input, and then it stabilizes at a new temperature.

How quickly does the surface shed extra heat? Engineers quantify this with a “heat transfer coefficient” that measures how many watts per square meter the surface loses for each degree it warms. For outdoor surfaces with even a light breeze, this coefficient is at least 10-15 W/(m²°C), and often higher. This means that for every degree the surface warms, it loses an additional 10-15 watts from every square meter.

So to hold the surface 1.5°C warmer than it would otherwise be, you need to deliver:

1.5°C × 10 W/(m²·°C) = 15 W/m² (using the most favorable coefficient)

or

1.5°C × 15 W/(m²·°C) = 22.5 W/m² (using a more typical value)

Our absolute ceiling of 12.5 W/m², which already assumed 100% of the data center’s heat reaches the ground, falls short of even the most favorable estimate of what’s needed. And in reality, the vast majority of the heat rises, disperses, and never reaches the ground surface within 1.6 km. A generous estimate for daytime conditions would be that 2-5% of the data center’s total heat output reaches the ground surface within 1.6 km. That gives us:

12.5 W/m² × 0.02 to 0.05 = 0.25 to 0.625 W/m²

Which would produce a surface temperature increase of:

0.25 to 0.625 ÷ 12.5 = 0.02 to 0.05°C

That’s roughly 1-3% of the observed 1.5°C signal. And that’s for a data center way larger than average and almost no wind.

Conclusion of the math

Even if we assume that the average data center in this study was a 100 MW hyperscaler (which it definitely wasn’t, since the vast majority of the 11,000+ facilities in the study’s database are much smaller enterprise and colocation centers drawing closer to 10-20 MW), and the wind is always gentle, the math still doesn’t work for waste heat. With a 100 MW data center, only about 1-3% of the observed temperature increase could plausibly be attributed to waste heat warming the surrounding ground. The remaining 97-99% is explained simply by the building and its pavement being warmer than the grass they replaced, which raises the satellite’s blended average temperature reading without actually warming the surrounding land at all.

And 100 MW is an extremely generous assumption. The study’s database spans 2004-2024 and includes facilities of all sizes. A typical facility in this dataset likely draws 10-20 MW or less. At 20 MW, waste heat could account for less than 1% of the observed signal. In other words, for the actual average data center in this study, the temperature increase is almost entirely, over 99%, explained by the fact that a building exists where a field used to be. The ground around these data centers almost certainly hasn’t warmed in any meaningful way. The remaining temperature change to the land itself caused by heat exhaust from the data center is so tiny that it cannot have any meaningful effect on the ecology or human wellbeing in the region.

The one remaining question: why the clean step change?

There’s one aspect of this that’s still somewhat confusing and doesn’t square with my story. Look again at this graph:

This jump is somewhat significant. An average increase of 2°C is about 1/10-1/20th of the difference between grass and a building, but also we’re assuming the building is taking up a small part of the satellite image. It is weird that this jump doesn’t occur until after the data center becomes operational. If this were construction, you’d maybe expect more of a slow gradual curve up?

The strength of the other evidence leads me to suspect that something’s off here. There’s no smoking gun, but I have a few explanations for what might be happening:

• The stuff that creates the large majority of the building’s new hot surface area is usually added at the very end of construction: roofs and parking lots. It would make sense to me that during construction a roofless building where the ground hadn’t been paved wouldn’t become as hot as a building with a full roof and large paved parking lot. If these are added in the last month and add the major part of the heat, it would appear like a step change. But this could be wrong. Some people who know more about architecture than me reached out to say the roofs are likely not what’s causing this, they’re often cooler than the construction site. I’m kind of at a loss here, could this be the parking lot? It’s hard to find reliable data on which parts of construction absorb the most heat. I need good sources on this but can’t find many. If you have tips, send them!

• The authors say they are measuring temperature changes relative to when each data center became “operational,” but that may be much less precise than it sounds. They don’t seem to know if operational here means “first server energized,” “first meaningful IT load,” “certificate of occupancy,” “roof completion,” or “substantial completion.” They’re just using a database date from the IEA-linked source they cite, and then centering all of their analysis on that single date. It would make sense to me that a lot of what’s listed in the data base might just be when the data centers wrapped up construction, which would align with when the roof and parking lot were added. The database they say was provided to them by IEA isn’t publicly available, so I have no way to check this. Data Center Map, the largest collection of public info on data centers online, says they track data center lifecycle stages including Land Banked, Planned, Under Construction, and Operational, but then immediately warns that these lifecycle stages “should only be considered indicative, as it is very hard to track when stages change.” If data center map considers its own data on this to be very hard to actually pin down, I’d suspect that IEA doesn’t have a significantly better private collection of all 11,000 data centers.

So my best guess is that the actual operational opening time varies a lot compared to the numbers the researchers are using, there’s going to be a clear jump in heat around the time the roof and parking lot are built, (←I’m now convinced that crossed out part’s wrong and this graph is a deep mystery to me) and therefore it’s likely that if you’re just looking at month by month snapshots, the roof + parking lot jump might appear to occur after the data center “became operational.”

One other important point is that this step change actually doesn’t look like what you’d expect to result from a data center’s energy draw, because most data centers are opened and then slowly filled with servers and powered on more and more over a period of months or even years. If this were the result of emitting more heat, you would expect the line to continue to go up and up as the data centers drew more power, but the actual average line appears to be flat after only a month of operations. The entire industry's ramp-up pattern is gradual, not instantaneous, at every scale, even for smaller colocation (not hyperscaler) data centers.

Data centers also vary enormously in their power draw, but the error bars where almost all observed heating effects occur are incredibly tight along a narrow band. I would expect any heating caused by data center power consumption to vary as much as the power consumption does itself, but this graph doesn’t show that at all. Instead, regardless of the size, data centers seem to have mostly shockingly similar effects on warming. Again, this is a sign that it’s the full physical infrastructure buildout of data centers and the stuff around them that might be causing this instead, though obviously data center physical size also varies a lot. This one’s very confusing in general.

Looking into this more, I learned that there are very few studies like this on other large buildings that emit a lot of heat. The clearest comparison I can think of is nuclear power plants, which can emit 10 to 20 times as much waste heat as a 100 MW data center. I’ve never heard of a nuclear power plant’s heat emissions raising the temperature of the ground around it, even though by the same logic it should be having 10-20 times the effect of the data center. However, there seems to be no info on this in either direction as no one’s seriously studied it.

I’m unsure here so will circle back if I get better evidence.

Other candidates for warming land

I would be pretty shocked to find out that the heating exhaust is able to warm the land this much after all this converging evidence, but my building explanation isn’t super satisfying as an explanation of the step change mentioned above. Other candidates for ways the data center could actually be warming land include:

• Groundwater and irrigation changes. Data center construction often involves significant grading, drainage modification, and sometimes groundwater pumping. If the local water table drops or irrigation patterns shift in surrounding agricultural land, vegetation dries out and LST rises. This could plausibly create a gradient extending several km, especially in arid or semi-arid regions where many data centers are built. Would be interesting if this were about water use, though I think this is pretty unlikely as there are a few aspects of water that don’t fit the patterns we’re seeing here.

• Vegetation stress from construction dust and runoff. During and after construction, dust deposition and altered drainage can stress vegetation in a radius around the site. Stressed or sparse vegetation has higher LST. This is temporary but could persist for years if the site continues expanding.

• A statistical artifact from MODIS point spread function. MODIS doesn’t have a perfectly sharp pixel boundary, its effective spatial response bleeds signal from bright/hot pixels into adjacent ones. A very hot data center pixel could artificially elevate apparent LST in neighboring pixels purely as an instrument artifact. This would produce a smooth radial decay that has nothing to do with actual ground temperature. MODIS was designed for global-scale monitoring rather than facility-level analysis, so this could maybe be a nontrivial contributor.

A few other bizarre methodological mistakes

Thanks to Rohit for pointing these out:

• They measure population affected by counting residents within km of each data center measured, but they don’t seem to check for overlaps at all. There are a lot of places like Loudoun county where data centers are all within 10 km of each other, so you’d get a lot of overlapping circles and people double counted in each one. The population affected number is likely inflated by double counting.

• The paper says it used "a reconstructed MODIS LST dataset (produced by NASA) acquired worldwide from 2004 to 2024 over an enhanced 500m resolution grid" and cites Metz et al. But the Metz et al. paper describes a European product with 250m spatial enhancement, not a global 500m product. So the citation doesn't match what they claim to be using.

The incredibly shifty framing the authors use

Over 99% of the observed land temperature increase seems to be coming from an omission in the author’s methodology. They’re assuming the land has warmed, but it hasn’t, and they’re only observing that there are additional warm buildings in the areas they are analyzing. They did not control for this at all and don’t even reference it as a possibility.

The authors strongly imply multiple times that the causal mechanism for this temperature rise is the high energy usage of the data centers. Here’s every time they imply that the effect here is due to heat emissions caused by high energy usage:

• Title: “The data heat island effect: quantifying the impact of AI data centers in a warming world”

  • This frames the phenomenon as caused by AI data centers specifically, not by construction or land development. A better title would be “quantifying the impact of large buildings.”

• Abstract: “we focus our attention on the heat dissipation of AI hyperscalers”

  • Explicitly names heat dissipation from operations as the focus of the paper.

• Abstract: “the land surface temperature increases by 2°C on average after the start of operations of an AI data centre”

  • “after the start of operations” implies the operations are causing it, not the construction.

• Introduction: “the type of human activities established and operating dramatically influence the impact of UHI on environment and communities”

  • Frames the effect in terms of activities operating, not buildings existing.

• Introduction: “data centres are expected to be one of the most power-hungry activity in the next decade”

  • Connects the phenomenon to power consumption.

• Introduction: “AI data centres are in vast majority relying on fossil fuel use”

  • Connects to energy source and emissions, implying operational energy is the driver.

• Introduction: “the steep growth of AI training and use for various applications would directly translate into high net impact on emissions”

  • Links AI workload growth to environmental impact.

• Introduction: “the inefficiencies and nonidealities of AI hyperscalers operations would cause their emissions to rise even more”

  • Again, operations.

• Section 2: “This study relies on the assumption that AI hyperscalers might have an impact on the LST of their locations because of the heat that they would release as a result of the high power demanding applications that they would be used for”

  • This is the most explicit statement. The assumed mechanism is heat released by computational workloads.

• Section 3: “These results are dramatically impressive, especially considering that the typical LST increase caused by the quintessential example of compound of anthropogenic activities – the urban heat island effect – has been estimated in the 4 and 6 °C interval”

  • Comparing to UHI, which is understood as an operational/activity-driven phenomenon.

• Section 3, Bajío case study: “The stable climate, low seismic activity, and proximity to North American markets made the Bajío region a great hub for AI data centres. Nevertheless, it has been recorded a serious LST increase trend”

  • Implies the data center operations caused the regional warming. In fact, this region fits the Andy Masley theory of More Buildings = More Hot Buildings much much better. In addition to building some data centers over the last 20 years, the Bajío underwent one of the most dramatic industrial buildouts in Latin America. The region received $37.6 billion in accumulated foreign investment from 2004 to 2019, with nearly two-thirds going to manufacturing. While national industrial activity grew at 1.7% per year from 2003-2016, Bajío industry expanded at 4.6%. Toyota, Honda, General Motors, Bombardier Aerospace, and hundreds of other manufacturers built large-scale operations across the region. Industrial parks, automotive plants, logistics centers, highways, rail networks, and supporting commercial infrastructure were constructed at an enormous scale during the exact period the paper attributes its warming trend to data centers. The LST increase in the Bajío is far more plausibly explained by the massive conversion of agricultural and rural land to factory rooftops, warehouses, parking lots, and paved roads than by the waste heat of data centers. The paper sees “data centers + warming” and draws a causal arrow between them, while completely ignoring the fact that this region was transformed from an agricultural center into one of North America’s most important manufacturing corridors during the same period.

• Section 3, Aragón case study: “The region is becoming a critical node for AI, cloud computing, and, increasingly, specialized server manufacturing. At the same time, the region has recorded an anomalous increase of approximately 2°C”.

  • Same story as Bajío, different country. Aragón has undergone a massive energy and industrial buildout during the study period. The region has become a multi-sectoral hub (from data centers to logistics and automotive) with companies like AWS, Microsoft, Blackstone, and others developing strategic projects there. In the last year alone, 651 new megawatts of wind and solar photovoltaic capacity were installed, a 7.1% increase over 2023. Aragón has been described as potentially the “Saudi Arabia of solar power in Europe” and bp alone has around 1.1 GW of solar projects at various stages of development or construction in Aragón. A single solar cluster near Zaragoza installed 615,000 bifacial solar panels across 650 hectares of land, with two overhead transmission lines of 18 km and 20 km constructed to connect it to the grid. That’s hundreds of hectares of dark solar panels, wind turbines, substations, transmission infrastructure, access roads, and associated industrial development blanketing a previously rural, sparsely populated region, exactly the kind of massive land cover change that would produce a clear LST increase when viewed from a satellite. The paper attributes the warming to data centers and ignores all of this. The Andy Masley theory of Buildings are Hotter than Grass is more parsimonious with the data.

• Section 4: The entire mitigation section proposes solutions to reduce operational energy consumption: more efficient algorithms, adiabatic circuits, model pruning, dynamic power response. These are solutions to “servers produce too much heat,” not solutions to “a building was constructed and is hotter than grass.”

• Section 5: “We focus our attention on the heat dissipation of data centres, which is directly connected to the energy consumption required for the operations of the AI hyperscalers”

  • Explicitly links the effect to operational energy consumption.

• Section 5: “Since the trends of data centre energy consumption are expected to show a steep growth in the foreseeable future, the data heat island effect could solidly become an additional factor for environmental and industrial sustainability”

  • Implies the effect will worsen as energy consumption grows, which only makes sense if operations are the cause. If it were land cover change, the effect would be determined by building footprint, not energy consumption.

The paper also repeats the claim over and over that 343 million people are “affected” by the data heat island effect. They find this by just counting everyone who lives within 10 km of a data center in the study’s database. They don’t bother to see whether those people experienced anything bad at all, like a change in air temperature, health outcomes, energy costs, or anything else. By the same methodology, you could count everyone living within 10 km of a Walmart and announce that billions of people are “affected” by the Walmart being a hot building regardless of whether it’s affecting anything else. In reality, I think 343 million people are being affected by the “If you build a building the building gets hot in the sun and doesn’t really affect anything else” effect.

Another incredibly shifty move the authors make is using "AI hyperscalers" and "AI data centres" throughout the paper, but the database they use (from the IEA) contains data centers of all types. Most data centers built between 2004 and 2024 weren't doing AI workloads at all, especially in the earlier years. The framing makes it sound like this is about AI specifically, but almost all of what they’re measuring is traditional non-AI focused data centers. It’s incredibly jarring to read “across all AI hyperscalers analysed during the 2004-2024 period” as someone who knows anything about the timing of the data center buildout. AI-focused hyperscaler data centers didn’t appear until after 2020.

I don’t understand the sociology of how something like this gets published. These are serious people! Going down the list of authors:

• The lead author Andrea Marinoni is an associate professor of applied remote sensing at the Arctic University of Norway with a visiting fellowship at Cambridge.

• Benjamin Horton is the Director of the Earth Observatory of Singapore at Nanyang Technological University, a leading sea-level rise researcher, IPCC author, Fellow of both the American Geophysical Union and Geological Society of America, and winner of the EGU Plinius Medal. His research was cited by President Obama in the 2015 State of the Union Address. He has 260+ peer-reviewed publications including in Science, Nature, and PNAS. He’s a very serious climate scientist!

• Erik Cambria is a Professor of AI at NTU, IEEE Fellow, Visiting Professor at MIT Media Lab, Provost Chair in Computer Science, 80,000+ citations on Google Scholar, listed in Clarivate’s top 1% of scientists, and was featured in Forbes as one of “5 People Building Our AI Future.”

• Pietro Lio’ is a well-known computational biology and AI researcher at Cambridge.

• Jocelyn Chanussot is a prominent remote sensing researcher at Université Grenoble Alpes, IEEE Fellow, with expertise in signal/image processing for remote sensing. He’s arguably the author whose expertise is most relevant to the methodology.

What happened here? I’m somewhat cynical and think that academics want attention on their work and personas, media attention is good for them, and making hyperbolic claims about data centers is a really great way to get media coverage right now, even if they have to use goofy methodology to get there.

The media is already running with this

Despite this paper not being peer reviewed and only being up on arXiv for 10 days, the media is already running with it. Notice that most of these headlines leave the reader with the impression that it’s the air that’s warming:

• CNN: Scientists have found an alarming environmental impact of vast data centers

  • CNN adds that data centers are “making life hotter for more than 340 million people.” They also add “The findings are particularly alarming, the scientists say, because AI data centers are set to boom over the next few years, and these temperature rises come as planet-warming pollution is already making heat waves more extreme around the world.” This is so incredibly deceptive. This study told us nothing about heating air. “Already making heat waves more extreme” paints a picture of people’s lives getting warmer as a result of this effect.

  • There were also a lot of local CNN affiliates that ran copies of the story, like here and here and here.

• Benton Institute for Broadband & Society: "AI data centers can warm surrounding areas by up to 9.1 degrees Celsius"

  • WHAT?! They ran with the absolute craziest outlier in their headline?!

• Bode Living: “AI Data Centers Are Generating Significant Local Heat, Study Reveals”

• New Scientist shares this story on their social media feed

• More Perfect Union wins the most deceptive presentation award, simply announcing “Major data centers are creating “heat islands” that warm the surrounding area by up to 16 degrees Fahrenheit, according to a new study.” They did exactly what I thought they’d do: leave out that this is a land effect and let the reader infer that this must be talking about the air temperature. They do nothing but lie on this topic.

• Gadget Review: "Scientists Find That Vast Data Centers Are Creating Dangerous Heat Islands" Opens with "Those ChatGPT queries and smartphone photo enhancements come with an unexpected side effect: they're literally heating up the planet in ways you've never considered"

• Slashdot: "AI Data Centers Can Warm Surrounding Areas By Up To 9.1C"

• AOL UK: "Data centers are creating 'heat islands' on land around them – warming them by up to 16 degrees, researchers warn"

• Gizmodo: “Data Centers Create Heat Islands Stretching 6 Miles, Study Finds.”

  The subtitle is: Heat islands can worsen air pollution and increase heat-related deaths.

• Fortune: “Data centers are so hot their ‘heat island’ effect is raising temperatures up to 6 miles away and impacting 343 million people worldwide, study finds” This one does link to this post as criticism at least!

• Futurism: "Data Centers Causing Huge Temperature Spikes for Miles Around Them, Study Suggests" This one also quotes me, though the title is clearly wrong.

What is going on with the media ecosystem around data centers? It’s crazy. The average person’s understanding of AI has already been wildly warped by bizarre obviously wrong coverage of data centers over and over. I’d like to make sure more new crazy wrong ideas don’t seep into the public’s beliefs. I want to stop this one from escaping containment if I can.

A parting thought: AI is useful for research

This paper didn’t pass the sniff test for me when I first read it. The first thing I did after I read it and had drawn some conclusions myself was upload it to Claude and ask it for anything that seemed off about the methodology. Claude responded like this:

This paper has some pretty significant methodological problems. Let me walk through the main issues.

The core identification problem is fatal. The paper’s entire causal claim rests on showing that LST rises ~2°C after a data center begins operations. But they’re measuring land surface temperature at 500m resolution from MODIS satellite data — and what happens when you build a data center? You replace vegetation or bare soil with a large dark-roofed building, parking lots, and paved access roads. That land use change alone would produce exactly the LST signature they observe, completely independent of waste heat. They’ve essentially rediscovered that buildings are warmer than fields when viewed from space. LST measures the radiative temperature of the surface itself, not ambient air temperature, so a rooftop baking in the sun will read dramatically hotter than the grass it replaced. They never attempt to distinguish this from actual thermal pollution.

No control group. A basic difference-in-differences design would compare data center sites to similar sites where, say, a warehouse or distribution center was built instead. They don’t do this. Any large commercial construction project — an Amazon fulfillment center, a Walmart, a factory — would likely show a similar LST step-change. Without controls, they can’t attribute anything specifically to data centers’ operational heat output versus generic land development.

The spatial decay curve is suspicious in the wrong direction. They find the effect “reaches up to 10 km” and drops to ~30% at 7 km. But think about what actual waste heat dissipation looks like: data centers exhaust warm air that mixes rapidly with ambient air. The thermal plume from even a large facility would be negligible beyond a few hundred meters in terms of measurable air temperature change. A 10 km radius of influence would require an implausible amount of heat. What would produce a gradually decaying LST signal over kilometers? The ancillary development (roads, commercial buildings, housing) that tends to cluster around data center campuses. They’re likely picking up the general suburbanization/industrialization footprint.

The 2°C figure is misleadingly framed. They compare their 2°C LST increase to the urban heat island effect’s 4-6°C range as if these are comparable phenomena, but UHI studies typically measure ambient air temperature differences (or at least carefully distinguish LST from air temperature). A 2°C LST increase in a single MODIS pixel that went from grassland to a commercial roof is trivially explained and tells you almost nothing about whether nearby residents are experiencing warmer temperatures.

The 343 million affected people claim is essentially meaningless. They count everyone living within 10 km of any data center globally and associate them with whatever LST change was observed. But they haven’t shown that these people actually experience any temperature change. LST at a data center rooftop doesn’t translate to air temperature in a neighborhood 5 km away. The population figure is just “people who live somewhat near data centers,” repackaged as an impact metric.

The case studies are cherry-picked and don’t support the argument. They mention the Bajío region in Mexico, Aragón in Spain, and Ceará/Piauí in Brazil as places with “anomalous” warming trends that coincide with data center presence. But these are regions with well-documented climate trends driven by broader factors — land use change from agriculture, deforestation, urbanization, and regional climate variability. Pointing at regions that have both data centers and warming trends, without controlling for anything else, is textbook confounding.

They conflate “AI hyperscalers” with generic data centers. The paper uses “AI hyperscalers” and “AI data centres” throughout, but the database they use (from the IEA) contains data centers of all types — colocation facilities, enterprise data centers, telecom facilities. Most data centers built between 2004 and 2024 weren’t doing AI workloads at all, especially in the earlier years. The framing makes it sound like this is about AI specifically when their data can’t support that distinction.

The seasonality removal and filtering methodology is underspecified. They mention removing seasonality and outliers but don’t describe how. With 500m MODIS LST data, cloud contamination, emissivity variations by land cover type, and viewing angle effects are all serious issues. The paper doesn’t discuss any of these, and “aggregating to monthly scale” doesn’t solve them — it just smooths over them.

The “step function” in Figure 2 could easily be an artifact of construction timing. If the “start of operations” date is close to when the building was physically completed, then the step change in LST simply marks when a large impervious surface appeared in the satellite’s field of view. You’d expect exactly this kind of abrupt transition regardless of whether the servers inside were even turned on.

The mitigation section is a grab bag that undermines the paper’s seriousness. They pivot from their empirical findings to discussing adiabatic circuits, model pruning, carbon-aware inference, passive radiative cooling coatings, and a “Matryoshka model of intelligence” — a philosophical framework by one of the coauthors. None of this connects rigorously to their findings. It reads more like a literature review of the authors’ other work than a coherent response to the specific problem they’ve identified.

Fundamentally, this paper presents a land-use-change signal as a climate impact finding. A rigorous version of this study would need to measure actual ambient air temperature effects (not just LST), include matched control sites, and disentangle construction from operations — and it would likely find a much smaller and more localized effect.

I didn’t rely on any of this being true. I did a deep dive on each to see if it held up and came up with additional arguments (like the math for how much energy the ground could take in), but ultimately I was given most of the issues with this paper within about a minute of uploading it to an AI chatbot, and major news organizations like CNN apparently missed these. My skills at finding these errors have in large part been automated away and you can download the same skills by using Claude or ChatGPT or Gemini. I claim that this is strong evidence that AI is, in fact, sometimes useful for research.

• A popular book summary of everything important in the IPCC report, written for a mass audience. Best I can find are the summaries published by the IPCC itself, which don’t seem to be reaching many people.

• A clear customizable visualization of how any one activity compares to your total daily and annual emissions, to help you figure out what to cut.

  • Edit: I made it!

  • The visualizations that do exist are pretty clunky.

  • A lot of people I know worry about the emissions of flying in planes, but almost none can tell me how those emissions compare to their daily lives (a flight from NYC to San Francisco emits in expectation about as much as 2 weeks of your regular lifestyle, or about 1/10th of a car’s annual emissions).

  • More importantly, few people seem to know how advocating for systematic change to the energy grid can pay off in ways that have hundreds of thousands of times as much impact on the climate as any lifestyle change you could make. I’ve been vegan for 10 years, but that basically rounds to zero compared to my friends working on smart grid tech. More people should know about how big this gulf is.

  • It is kind of an overwhelming thought that in the fourth decade of serious climate communication, the average well-educated person is still basically spinning a roulette wheel in their head to make random lifestyle cuts for the climate with no real sense of how the options compare to each other. Eat less meat or fly less? Cut chatbots or recycle? So few people actually know anything about how these compare.

• A map of new smart grid and green energy infrastructure in your area you could go and advocate for.

• An updated version of Sustainable Energy – without the hot air. My favorite climate book, partly because when you read it you become confused about why so little climate communication is able to achieve anything like the same level of simple clarity.

• An 80,000 Hours-like comprehensive website trying to rigorously compare the actual impacts of different climate careers specifically (I know the Effective Environmentalism network aims to do something like this).

• A big public estimate of how much emissions have been prevented, in total, by the climate movement, by government action, by new technology vs. policy vs. lifestyle changes etc. I asked Claude and it estimates based on some papers that the climate movement specifically has prevented maybe 8 Gt of CO2, a little over a year and a half of the US’s emissions. I think this is way too low because it’s not including government policy, which Claude estimates to have prevented 34 Gt, and surely the movement is having lots of indirect impact there. Is any of this correct at all? Zero clue, I’d like a better source on this, but I can’t find one!

Most of these don’t seem that hard to make and maintain. I might try to make the second one if no one beats me to it. In general, a lot of people thinking at high levels about this stuff seem to be asleep at the wheel when it comes to how much they can change the conversation by just putting out clear, accessible, but also comprehensive and highly numerate info about climate change.

In a more sane world, I wouldn’t have been able to build the big audience I have by saying over and over again “The emissions of a computer program aren’t really in the same ballpark as driving a car,” but too many people who know about this stuff weren’t saying anything. I got lucky and had the time and energy to type it down, and was also motivated a lot by the effective altruist observation that almost no one is bothering to publicly compare very simple numbers involved in big global problems, and that those numbers do matter quite a bit and should be more action-guiding than they often are. The basic EA intuition is that a lot of people who talk about big high-status issues are coasting on saying the correct social passwords in order to be seen as good people, that we easily deceive ourselves with these passwords into thinking our own beliefs are way more thought out than they actually are, but also that the issues themselves are often very real, getting them right matters, and taking the numbers seriously is a way of breaking through the status games we play with ourselves and others.

You should strongly defer to actual expert consensus unless you have an extremely good reason not to, but you need to figure out what the actual expert consensus is and not trust that social passwords align with it. I was lucky in college to learn a lot about climate change and what the actual expert consensus was, so I could take the problem seriously while also ignoring the people confidently telling me that the US would fall into a civil war over water by the early 2020s. That was just a password, and the actual expert consensus was available to me if I was willing to seek it out instead of coasting on this new fake way of signaling that I was a serious clear-eyed person. So much of my success in blogging in the last year came from the simple decision to not go along with the social password “You should definitely be mindful with your chatbot use.”

I’m not bothered at all if people don’t like EA, but I do think this simple story of the world it tells is basically true and a way of springboarding yourself way above the fray in big debates about global problems like this. Don’t call it EA if you don’t want to, don’t feel like you need to give EA itself any status, but if you’re thinking about diving into climate communication, keep this basic story in mind. You can break out of the mental stupor of coasting on high-status passwords with a blank Google doc, a calculator, and a goal to seriously communicate how simple quantities compare to each other (with humility and strong deference to expert consensus where it exists). As I’ve learned over and over, there’s a huge appetite for this way of thinking, and you should assume that so many people are coasting along without bothering to look for the gaps in the discourse that these gaps are actually very easy to find and fill as soon as you start seriously looking for them. If you step out of trying to say the right social passwords about climate (or another big issue you care about) and just start free writing about what everyday public resources could exist for it but don’t, I think you’ll be surprised at how easy it is. Even just a simple rule like “Make the resource you wish your past self had” on its own goes a long way.

As an aside, Our World in Data is a light in the dark on good climate stats, and if you go back through my posts you’ll notice how reliant I am on a lot of the climate data they’ve compiled. Draw inspiration there! The Effective Environmentalism network also think about this stuff a lot. Project Drawdown’s another good example. So is this Founders Pledge report on climate philanthropy. There are lots of others, but not enough.

Update 11/17/2025: the author took the time to respond to me below. While I’m very grateful, the materials she sent actually seems to confirm my main criticism and I’m now very confident a key number in the book is 1000x too large and needs to be revised. I summarize everything in my reply here.

Intro

I was taking a break from posting about AI and the environment, but after reading parts of Karen Hao’s book Empire of AI, I’ve stumbled on such wildly misleading claims that have so far gone unaddressed that I’ve felt the need to counter them here. Within 20 pages, Hao manages to:

• Claim that a data center is using 1000x as much water as a city of 88,000 people, where it’s actually using about 0.22x as much water as the city, and only 3% of the municipal water system the city relies on. She’s off by a factor of 4500. This is the single largest error in any popular book that I’ve found on my own, and to my knowledge I’m the first person to notice it.

• Imply that AI data centers will consume 1.7 trillion gallons of drinkable water by 2027, while the study she’s pulling from says that only 3% of that will be drinkable water, and 90% will not be consumed, and instead returned to the source unaffected.

• Paint a picture of AI data centers harming water access in America, where they don’t seem to have caused any harm at all.

• Frame Uruguay as using an unacceptable amount of water on industry and farming, where it actually seems to use the same ratio as any other country.

• Frame a proposed data center in Uruguay as using a huge portion of the region’s water where it would actually use ~0.3% of the municipal water system, without providing any clear numbers.

These are all the significant mentions of data centers using water in the book. Read in this light, the chapter becomes somewhat ridiculous, because the rest includes descriptions of brutal acts of torture and plunder under colonialism, and then frames data center water use as a continuation of that same colonialism. If instead you see data centers using water in other countries as part of a simple trade the countries are making to get more taxable industry in the area, and that doesn’t seem to harm water access, the central narrative thrust of the chapter becomes false.

This book has been very popular, and influenced the AI/environment conversation. One of the most common replies I’ve received to my water arguments is that I should read it. So I did, and I came away kind of shocked at how badly it covered the issue. Hao’s points about water are all in the chapter “Plundered Earth.”

My only ask for people writing about AI and the environment is that at the end, readers are left with a more accurate picture of how energy and water is being used overall in the regions covered, and not left with a contextless impression of AI as a huge environmental offender in places where it’s not. Every single time water’s mentioned in this book, the reader is left with a worse understanding than they came in with.

I make my case for each with easily accessible numbers. All quotes and page numbers are taken from the Kindle edition.

Contents

• A misleading projection of how much water AI is expected to use

• A massive factual error: Hao claims a data center would use 4500x as much of a city’s water as the actual value

• A weird framing of Uruguay’s water use

• Misleading presentations of data center water issues in America

• Conclusion

A misleading projection of how much water AI is expected to use

The first time the book goes into detail on water is here:

The land and energy required to support these megacampuses are but two inputs in the global supply chain of data center expansion. So, too, is the extraordinary volume of minerals including copper and lithium needed to build the hardware—computers, cables, power lines, batteries, backup generators—and the extraordinary volume of potable—yes, potable—water often needed to cool the servers. (The water must be clean enough to avoid clogging pipes and bacterial growth; potable water meets that standard.) According to an estimate from researchers at the University of California, Riverside, surging AI demand could consume 1.1 trillion to 1.7 trillion gallons of fresh water globally a year by 2027, or half the water annually consumed in the UK.

(pp. 277-278)

The study mentioned is “Making AI Less Thirsty.” The study does not say that AI demand will consume 1.1-1.7 trillion gallons of water annually. Hao seems to be getting this number from this part of the study:

Even considering the lower estimate, the combined scope-1 and scope-2 water withdrawal of global AI is projected to reach 4.2 – 6.6 billion cubic meters in 2027

4.2-6.6 billion cubic meters = 1.1-1.7 trillion gallons.

Withdrawal is very different from consumption. Withdrawal means the amount of water taken from a local source. Consumption is the amount of water taken and not returned to the local source. Many of the ways we use water (especially power plants) withdraw water temporarily from local sources, use it briefly, and then return it unaffected. This is called non-consumptive withdrawal. It’s like diverting part of a river to run near a mill where you have some water wheels, and the diverted water then returning to the main flow after. Consumptive withdrawal is like sucking up and evaporating the water. The main water issue for most regions is water consumption, not just water withdrawal.

How different is consumption and withdrawal for AI? Very! The next sentence in the study Hao cites says:

Simultaneously, a total of 0.38 – 0.60 billion cubic meters of water will be evaporated and considered “consumption” due to the global AI demand in 2027.

This is equal to 100–158 billion gallons, only 10% of the number Hao reports. If someone’s writing a book that covers AI water use in depth, I would expect them to know the difference between consumption and withdrawal, especially when the consumption number is given in the very next sentence of the study. It’s very weird that Hao misidentifies withdrawals as consumption here.

Why are withdrawals so much bigger than consumption for AI? It’s mainly because the authors of the study are measuring how much water offsite power plants use to generate electricity for AI as part of its total withdrawals number. The vast majority of water that power plants use is withdrawn and then returned to the local source (non-consumptive use on this graph).

So the withdrawal number is going to be way bigger than the consumption number. Withdrawal is just not nearly as much of an issue for local water access as water consumption.

But the way this is being measured gets weirder. Even just measuring consumptive water use, the amount of water actually removed permanently from a local source, the study authors still find that the vast majority of water consumed by AI is offsite in nearby power plants that data centers draw from. For the US average, only ~15% of the water AI is consuming is actually happening in data centers themselves (the on-site water column).

Basically none of the water power plants use is potable. The only potable water used for AI is in data centers themselves. So the study Hao is citing is not only saying AI will consume just 10% as much water as she says, only 15% of that 10% will be drinking water.

Read Hao’s paragraph again. It strongly implies that AI will be using 1.1-1.7 trillion gallons of drinking water:

The land and energy required to support these megacampuses are but two inputs in the global supply chain of data center expansion. So, too, is the extraordinary volume of minerals including copper and lithium needed to build the hardware—computers, cables, power lines, batteries, backup generators—and the extraordinary volume of potable—yes, potable—water often needed to cool the servers. (The water must be clean enough to avoid clogging pipes and bacterial growth; potable water meets that standard.) According to an estimate from researchers at the University of California, Riverside, surging AI demand could consume 1.1 trillion to 1.7 trillion gallons of fresh water globally a year by 2027, or half the water annually consumed in the UK.

(pp. 277-278)

In reality, the study she’s citing specifically says that the water used by AI data centers themselves (the only place potable water is used) could consume 150 – 280 billion liters. This is 40-74 billion gallons. This is just 3.6% of Hao’s number. Further, the study notes that only 80% of the water used by Google’s data centers is potable, brining the number down 32-59 billion gallons, just 3% of Hao’s number. Here’s how much of Hao’s number is non-drinkable water returned to the source unaffected (blue), non-drinkable water consumed by power plants and data centers (red), and drinkable water used in data centers themselves (yellow).

The paragraph compares water to half the UK, and leaves the reader to infer that this is all potable drinking water. London uses 2.6 billion liters of water every day, 690,000,00 gallons, which is 252 billion gallons per year. So the actual amount of potable drinking water this study Hao is citing is projecting that all global AI will use is at most 20% of just the London water system, and 1.5% of the UK’s water use.

Hao repeats this misconception in her Atlantic article AI is taking water from the desert:

Public data hint at the potential toll of this approach. Researchers at UC Riverside estimated last year, for example, that global AI demand could cause data centers to suck up 1.1 trillion to 1.7 trillion gallons of fresh water by 2027.

This leaves the reader with the very strong impression that the physical data center buildings themselves will “suck up” and not return up to 1.7 trillion gallons. In reality, 90% of this “sucked up” water will be returned to the source it was drawn from, and only 3% of it will be in physical data centers themselves.

I have some disagreements with the study itself. Among other things, a lot of the water it measures AI “using” offsite is actually water evaporated in lakes dammed by hydroelectric plants to generate electricity, but it ignores the water these lakes recapture from rain. In similar studies this evaporated water ends up being 2/3rds of the total water use estimate of AI in general, so I think the actual offsite use should be significantly lower. The study is also two years old and I think we just have better estimates based on way more data now, specifically this report. We also have better numbers for chatbots measured by companies themselves, which show them using significantly less water than the estimates we had available in 2023. I also argue here that whether water is potable doesn’t really matter, because what actually harms the availability of drinkable water is access to total freshwater, not potable water. It’s relatively easy to turn freshwater potable.

A massive factual error: Hao claims a data center would use 4500x as much of a city’s water as the actual value

The chapter later turns to Chile, focusing a lot on the municipality of Quilicura on the outskirts of Santiago. The chapter covers ways that colonialism has seriously harmed the people and local environment and water there in the past, and then turns to the Google data center built there. The Google data center is framed as a continuation of the colonialism. It then goes on:

This unique combination—a history of neglect and a precious water source—created fertile ground for the blossoming of several environmental activist groups who were used to being watchdogs and were fiercely protective against the extraction of their resources. That summer, as Google filed a report with Chile’s environmental agency for approval of its data center—a largely rubber stamp process—MOSACAT, a water activist group, began combing through all 347 pages of the filing. Buried in its depths, Google said that its data center planned to use an estimated 169 liters of fresh drinking water per second to cool its servers. In other words, the data center could use more than one thousand times the amount of water consumed by the entire population of Cerrillos, roughly eighty-eight thousand residents, over the course of a year. MOSACAT found this unacceptable. Not only would the facility be taking that water directly from Cerrillos’s public water source, it would do so at a time when the nation’s entire drinking water supply was under threat. In 2019, as with Iowa and Arizona, Chile was already nine years and counting into a devastating and historically unprecedented megadrought.

(pp. 288-289).

Look at this line again:

In other words, the data center could use more than one thousand times the amount of water consumed by the entire population of Cerrillos, roughly eighty-eight thousand residents, over the course of a year.

Hao justifies this number in the notes section at the end of the book:

In other words, the data: The Google environmental impact report to SEA stated that the data center could use 169 liters of potable water a second, or 5,329,584,000 liters a year. According to the water service authority in Cerillos, the municipality consumed 5,097,946 liters in all of 2019, the year Google sought to come in; 5,329,584,000 liters a year divided by 5,097,946 liters a year equals 1,045.

(p. 454)

Something isn’t adding up here. 5,097,946 liters a year is 13,966 liters per day. Dividing by 88,000 residents says that each resident is using 0.2 liters of water per day. That’s about this much:

This implies that a city of 88,000 people is using as much water per day in total as a single shower head left running. Adults need to drink ten times as much as this, 2-4 liters of water per day, to stay alive. The average citizen of Chile buys 180 liters per day from their municipal water systems. I think the actual amount of municipal water supplied to people in Cerrillos is 900 times as much as Hao is claiming here.

I can’t find the study Hao is referencing, but I did find this document from the local government saying that the potable water consumption for the region of Maipú, Cerrillos, part of Estación Central (which the water system seems to serve) in 2019 (the year the study was conducted) was 54,148,639 m³. That’s 54,148,639,000 liters in a year, 148,352,000 liters per day. The total population the system is serving seems to be ~650,000 people.

So each person using this system is using ~230 liters of water per day. That’s way more in line with Chile’s average water use. The reason this is larger is probably that it also includes commercial buildings.

Hao got this number wrong by 3 orders of magnitude and reported the data center as using 1000 times as much water as a city of 88,000 people. I can’t get over how crazy this is. She implied that a single building is using as much water as a city of 88 million people. That’s over 4x the entire population of Chile! Readers are leaving this book assuming that Google built a single building in Chile that’s using 20% of all the residential water on the whole continent. No wonder everyone’s freaking out so much about AI and water use.

I think I know what happened: this municipality seems to report their water use in cubic meters. Each cubic meter is 1000 liters. Hao’s number reported in liters is noticeably very close to the cubic meter value I see in my sources. I suspect she got her wires crossed and somehow recorded m^3 as liters, making the city appear to use 1000x less water than it does. But she and others should have caught this in editing. I’m not aware of any building anywhere that uses 1000x as much water as a nearby city. That would be crazy! Note: Since writing this Hao replied here, and the material she sent confirmed to me that this is actually what happened, I explain here. I expect the book to be corrected.

But it gets worse! This is not the only misleading thing about how this number is presented.

This 169 L/second number was reported by a local paper as the maximum permitted amount. Nowhere in Hao’s writing does it make it clear that this number is the very maximum upper bound of how much water the data center will use, not the average normal water use. Reporting the maximum permit for water draw as the normal amount it will actually draw is very misleading. Here’s my explanation for why from my main post on AI and water:

Many articles about current or future AI data centers report the number in the water permit they apply for as the amount of water they actually use day to day. But this almost never happens.

When a data center is being built, the company needs to obtain water use permits from local authorities before construction. At this stage, they have to estimate their maximum possible water consumption under worst-case scenarios:

• All cooling systems running at full capacity

• Peak summer temperatures

• Maximum IT load (every server rack filled and running)

• Minimal efficiency from cooling systems

The permit needs to cover this theoretical maximum because regulators want to ensure the local water infrastructure can handle the demand and that there’s enough water supply for everyone. It’s easier to get a higher permit upfront than to come back later and request more, so data centers are incentivized to aim high.

Actual water usage is always significantly lower than what the permits allow, because they’re designed with the absolute worst conditions in mind. But many popular articles about how much water data centers use give the number on the water permit, not how much the data center actually uses.

We don’t have information on how much water the data center was actually expected to use, but we can at least compare it to other Google data centers permitted for the same amount of water to get a rough guess. Google’s data centers in The Dalles Oregon were permitted to draw the same maximum amount of water per day. The actual amount they used in a year was 275 million gallons, which is 0.75 million gallons per day. So another data center with this same permitted amount ended up using just 20% of its actual permit.

Maybe the data center would use more water because of different climates? The only times when data centers seem to use significantly more water is during heat spikes at the hottest times of the year, and The Dalles’s hottest month is on average warmer (88 F average in the Dalles, 71 F average in Cerrillos).

This all points to the actual expected water usage of this data center being around 1 million, not 4 million, gallons per day. 1 million gallons per day is 3,800,000 liters per day, or 1,767,000,000 liters per year.

So in a place where the local water system sells 54,148,639,000 liters per year, this data center’s normal operations would have raised the total water system’s demand by 3%.

Compare Hao’s sentence on this:

the data center could use more than one thousand times the amount of water consumed by the entire population of Cerrillos, roughly eighty-eight thousand residents, over the course of a year.

To what I think is the correct description:

A municipal water system serves 650,000 people. If a data center is built in the region, it would raise the system’s annual demand by 3%, equivalent to the water used by 19,500 residents.

Still significant, but drastically different from the number Hao gives. This is the main story of a data center harming water access in the book, and the main number given for its water usage is 4500 times as high as the correct value. Ridiculous!

This book has been reviewed almost 1000 times on Amazon:

This also shows up in an interview did with Tech Policy Press:

That’s where Google wanted to add its data center and then tap into that freshwater resource to cool its data centers. The amount of water that they were proposing was to consume more than a thousand times the amount of water that residents in that community would typically consume. They went through, they fought tooth and nail just to get Google’s attention because they not only had to make enough noise to pressure Google Chile, they had to make enough noise to then get that all the way back to headquarters in Mountain View.

And from what I can tell not a single person has noted “Oh hey the central story of a data center harming water access is assuming its water use is 4000x as big as it actually is.” Way, way, way too many people are reading about AI water issues uncritically. Why is it so easy for me to so quickly stumble on these gigantic errors in popular writing, that it seems like no one else has found? This book has been really popular with environmentalist critics of AI. Surely someone would have noticed this mistake?

Hao says here that she spent time with activists in Chile focused on data centers. Surely she would have noticed somewhere that these numbers were off. She specifically says they were “fighting tooth and nail to stop these data centers from taking all their drinking water” so conversations must have focused on the amount of water the data centers would actually use. How did this go so wrong? Further, she has a degree in mechanical engineering from MIT with a minor in energy studies and was a senior editor on AI at the MIT Tech Review. Surely she knows that there’s just no way any single physical building we’ve ever built would use as much water as 1000 times a city of 88 thousand people. The data center with the most water used in the world as of writing seems to be Google’s Iowa data center, with a billion gallons used every year. Even this one is just using half as much water as Cerrillos. This new data center would have to be 2000 times as large as any previously existing data center to make her claim true. How did someone covering AI for a living for a decade, with an energy studies degree from MIT, who spent time in the region itself with water activists, not immediately clock that this is way too big?

A weird framing of Uruguay’s water use

The next section covers Uruguay. It opens with a discussion of drought in the country, and includes a strange observation:

The water crisis emerged from the compounding effects of climate change and a failure of the state’s allocation of freshwater resources: In Uruguay, more than 80 percent of the country’s fresh water goes to industry instead of human consumption—most notably, cash crop agriculture. These include industrial farms for soybeans and rice, and for trees that feed into paper production. Most such farms are run not by local companies but by multinationals that export what they grow and show little accountability for Uruguay’s natural environment. Their activities deplete the nutrients in the soil, making it more difficult to grow actual food, and pollute the country’s water streams with a volume of fertilizers that makes Uruguay one of the world’s largest per capita fertilizer consumers and causes unusually high rates of cancer.

(p. 293)

The reason I think this is strange is that in basically every country, including the US, 80% of water is used on industry and commercial buildings rather than household consumption. I think a lot of readers don’t know that. This framing implies Uruguay is holding back water from locals who need it more. In reality they have the same split of water as every other country. Water’s very important to agriculture, industry, and commercial buildings. Hao presents this as a sign that Uruguay’s water is being used by big evil industries and being kept from the people. In reality it seems to have the same split of water as any other country.

The chapter then goes to a sociology researcher named Daniel Pena.

So when Google arrived, Pena was vigilant. During his regular scans of the Uruguayan environmental ministry’s website, which lists major industrial projects, he came across the company’s proposal for the data center. Pena had read about hyperscalers using potable water, even during major droughts, and the activism of communities like MOSACAT that had resisted the projects. But when he downloaded the details of the project, the water numbers were marked as confidential. After submitting a public information request, which he had successfully done around twenty times, the ministry continued to withhold the numbers, saying they were proprietary information. Pena wondered what they were hiding and worried about the precedent it would set for other cloud companies that would inevitably begin to eye Uruguay, following Google’s lead, for their own expansion. So he evoked the water clause in the constitution. With the help of a lawyer friend who was willing to work pro bono, he sued the ministry.

In March 2023, four months later, Pena won the case in a surprising victory. The environmental ministry revealed that Google’s data center planned to use two million gallons of water a day directly from the drinking water supply, equivalent to the daily water consumption of fifty-five thousand people. With much of Montevideo receiving salt water in their taps not long after, the revelations were explosive. Thousands of Uruguayans took to the streets to protest Google and all of the other industries that had led the government to squander the country’s precious freshwater resources.

….

Near the end of 2023, Google silently updated its proposed data center in Uruguay to use a waterless cooling system and said it would reduce the facility to a third of its size.

(pp. 294-295)

I’m not going to lecture locals in a far away country what they should and should not build in their country, but I do want to know how nefarious Google was being here. Again, I suspect this 2 million gallons per day was an upper limit, and if it’s anything like other Google data centers the number’s probably closer to 400,000 gallons per day. The municipal water system serving the city of 1.7 million people there seems to use 500,000 m³ per day. This is 132 million gallons. So if the data center had been built, it would be using about 0.3% of the municipal water system. Again, not nothing, but the reader is left with zero sense of scale here. It should be pretty easy to provide that.

Misleading presentations of data center water issues in America

Hao repeatedly mentions data centers in America built in water stressed areas. Each mention I think is misleading. Take this example from Iowa:

Altman and other executives never brought up the data centers’ environmental toll in company-wide meetings. As OpenAI trained GPT-4 in Iowa, the state was two years into a drought. The Associated Press later reported that during a single month of the model’s training, Microsoft’s data centers had consumed around 11.5 million gallons, or 6 percent, of the district’s water. GPT-4 had trained there for three months. (A Microsoft spokesperson said the company is working to increase its water efficiency by 40 percent above its 2022 baseline and to replenish more water than it consumes across its global operations by 2030, with a focus on the water-stressed regions where it works.)

A month of using 11.5 million gallons means each day OpenAI used 380,000 gallons of water. Corn in Iowa uses between 0.1-0.2 inches of water to grow per day. 0.1 inches of water over 1 acre is 27,154 gallons. So OpenAI was using as much water as 14 acres of an Iowa corn farm, or 0.02 square miles. The average Iowa corn farm is 346 acres. This amount of water is equivalent to Sam Altman purchasing 4% of a single Iowa farm to grow corn for his employees. Here’s that area on a map (the yellow box).

Does Hao’s paragraph get this magnitude across? If you heard a company had bought 4% of a corn farm, how big of a problem would you assume this is for regional water access? What if the tech company were using this 4% to grow something that half a billion people would use every single week for a year?

The book turns next to Arizona.

Arizona, too, faces a severe water crisis. In 2022, as Microsoft laid the groundwork for Phase 3, a study in Nature Climate Change found that the Southwestern US had been facing the worst drought it had seen in over a thousand years. That drought, combined with severe mismanagement, has drained the Colorado River, which Arizona and six other states rely on for fresh water, to dangerously low levels. Without drastic action, the river could cease to flow. The shortage compounds a power crisis, as climate change has slammed the region with relentless record-breaking temperatures and families have cranked up their air-conditioning. The region relies in part on hydropower from the Hoover Dam and water-cooled nuclear power plants. In other words, it needs water to produce more energy. In 2023, the Phoenix metro area hit multiple new heat records as well as the worst year for heat-related fatalities, which surged at least 30 percent from 2022 to over six hundred dead. “All things,” says Tom Buschatzke, the director of the Arizona Department of Water Resources, “are converging in a challenging direction.”

(pp. 280-281)

Notice that Hao does not give a number for how much water Arizona data centers are using. Here’s a section of my main water post where I look at how much water data centers in Maricopa County (the main place they’re being built there) are using compared to the total county water and golf courses, and how much tax revenue they’re bringing in per unit water:

Take Maricopa County in Arizona. The county is home to Phoenix, and is in a desert where water is pumped in from elsewhere. It’s also one of the places in the country where the most new data centers are being built.

Circle of Blue, a nonprofit research organization that seems generally trusted, estimates that data centers in Maricopa County will use 905 million gallons of water in 2025. For context, Maricopa County golf courses use 29 billion gallons of water each year. In total, the county uses 2.13 billion gallons of water every day, or 777 billion gallons every year. Data centers make up 0.12% of the county’s water use. Golf courses make up 3.8%.

Data centers are so much more efficient with their water that they generate 50x as much tax revenue per unit of water used than golf courses in the county:

So even though data centers are using 30x less water than golf courses, they bring in more total tax revenue:³

Some people see this, and react with something like “Well I don’t think golf courses OR data centers should be built in the desert.” At some point this becomes an argument against anyone living in deserts in the first place. If you want to have a gigantic city in the desert, like Phoenix, that city needs some way of supporting itself with taxes, and giving jobs to the people who live there. Most industries use significant amounts of water. If Phoenix is going to exist, it’s going to need private industries built around it that are using some water. We have two options here:

• Build industries that generate huge amounts of tax revenue relative to the water they use. Data centers fall into this category (though they don’t provide many jobs).

• Do not build cities in the desert in the first place.

Arguments against data centers existing in the desert because they harm water systems there also often apply to building cities in the desert in the first place. It’s fine and consistent to say that Phoenix shouldn’t exist because it’s unnaturally pumping water from hundreds of miles away, but it’s inconsistent to say that Phoenix should exist, that its water bills should be kept as low as possible, but also that no industries that use any water should be built there.

Simply reporting data centers as “being built in water scarce areas like Arizona” I think leaves out way too much important context to leave readers more informed.

Hao opens the water section with this observation:

Another study found that in the US, one-fifth of data centers were already drawing that water before the generative AI boom from moderately or highly stressed watersheds due to drought or other factors.

(p. 278)

This is true, but data centers are subject to the laws of supply and demand like any other business. In places where water’s more scarce and expensive, data centers are more likely to use other methods of cooling. In Arizona, where lots of data centers are being built, they’re using very little water relative to many other industries, and the tax revenue they’re bringing in. Overall, I haven’t found a single place where the normal operation of a data center has caused any issues for water access anywhere in America. We have plenty of other industry in medium and high water stress areas. Those communities also benefit from the tax revenue industry and commercial buildings bring in. But you wouldn’t know that from the book’s coverage.

Conclusion

Hao is smart. She has a degree in mechanical engineering from MIT with a minor in energy studies and was a senior editor on AI at the MIT Tech Review. If she wanted to give readers a full complete picture of where AI data centers are fitting into the broader environmental picture, this would be extremely easy for her. But every individual mention of water in Empire of AI leaves the reader less informed.

More broadly, this is a terrible sign for just how bad the public understanding of AI water use is. This book has been out for 6 months, received 1000 reviews on Amazon, was recommended by Time, The New York Times, Vulture, The New Yorker, The Economist, Financial Times, and Kirkus Review. Hao also thanks her fact-checking team in the acknowledgements:

To my incredible fact-checking team: Lindsay Muscato, Matt Mahoney, Rima Parikh, and Muriel Alarcón. All four of them fastidiously combed through the draft, cross-checking the labyrinth of details against documents and sources, and stress-testing my word choices. Matt also supported early research in my book, and Lindsay fielded many calls from me to serve as the most patient sounding board, while Rima somehow turned her fact-checking notes into standup comedy. They are all lifesavers.

(p. 424)

And yet after all these people read the book, I’m somehow the first person to notice and comment on the fact that no building, anywhere on Earth, is using anywhere close to 1000 times as much water as a community of 88,000 people use. How did this happen? Why did no reader, anywhere, let the author know about this massive glaring issue?

I think this happened because almost no one covering this issue is actually just looking at the numbers on AI and water use. The people who do are mostly experts less engaged in the public conversation. Almost everyone covering this seems to just be going with vibes. The fact that so, so many journalists, self-identified environmentalists, and members of the public could just nod along to the claims here and let them slip by provides a useful intuition for how it could possibly be that after all the ink spilled on it by so many professional people, the AI water issue is still somehow fake.

More by me

• All my writing on AI and the environment

• The AI water issue is fake

• Data centers don’t harm water access at all anywhere in America

• More Perfect Union videos are wildly deceptive on data center water use

• Contra the UK government, please don’t delete your old photos and emails to save water

In the past few months I’ve spoken to a lot of people facing objections to using chatbots, including a surprising number of people who want to buy chatbot access for their large organizations and have been shot down because of worry over the environmental impact of individual prompts. I think it’s crazy that this is still happening and want a much shorter post readers can send to people who are still misinformed about this. Here it is!

Using ChatGPT isn’t bad for the environment

Using chatbots emits the same tiny amounts of CO2 as other normal things we do online, and way less than most offline things we do. Even when you include “hidden costs” like training, the emissions from making hardware, energy used in cooling, and AI chips idling between prompts, the carbon cost of an average chatbot prompt adds up to less than 1/150,000th of the average American’s daily emissions. Water is similar. Everything we do uses a lot of water. Most electricity is generated using water, and most of the way AI “uses” water is actually just in generating its electricity. The average American’s daily water footprint is ~800,000 times as much as the full cost of an AI prompt. The actual amount of water used per prompt in data centers themselves is vanishingly small.

Because chatbot prompts use so little energy and water, if you’re sitting and reading the full responses they generate, it’s very likely that you’re using way less energy and water than you otherwise would in your daily life. It takes ~1000 prompts to raise your emissions by 1%. If we assume each response is ~100 words, and you read at the speed an average American does, and writing the prompts and waiting for the response took you no time, it would take you 6 hours and 30 minutes to read all the responses. So you would use half your waking day on an app that in total caused 1% of your emissions. If you sat at your computer all day, sending and reading 1000 prompts in a row, you wouldn’t be doing more energy intensive things like driving, or using physical objects you own that wear out, need to be replaced, and cost emissions and water to make. Every second you spend walking outside wears out your sneakers just a little bit, to the point that they eventually need to be replaced. Sneakers cost water to make. My best guess is that every second of walking uses as much water in expectation as ~7 chatbot prompts. So sitting inside at your computer saves that water too. It seems like it’s near impossible to raise your personal emissions and water footprint at all using chatbots, because using all day on something that ends up causing 1% of your normal emissions is exactly like spending all day on an activity that costs only 1% of the money you normally spend.

There are no other situations, anywhere, where we worry about amounts of energy and water this small. I can’t find any other places where people have gotten worried about things they do that use such tiny amounts of energy. Chatbot energy and water use being a problem is a really bizarre meme that has taken hold, I think mostly because people are surprised that chatbots are being used by so many people that on net their total energy and water use is noticeable. Being “mindful” with your chatbot usage is kind of like filling a large pot of water to boil to make food, and before boiling it, taking a pipet and removing tiny drops of the water from the pot at a time to “only use the water you need” or stopping your shower a tenth of a second early for the sake of the climate. You do not need to be “mindful” with your chatbot usage for the same reason you don’t need to be “mindful” about those additional droplets of water you boil.

Some people think tiny parts of our emissions “add up” when a lot of people do them. They add up in an absolute sense, but they don’t add up to be a larger relative part of our overall emissions. If AI chatbots are just a 100,000th of your personal emissions, they are likely to be around a 100,000th of global emissions as well. We should mostly focus on systematic change over personal lifestyle changes, but if we do want to do personal lifestyle changes, we should prioritize cutting things that are actually significant parts of our personal emissions. That’s the only way we could reduce significant amounts of global emissions too.

The reason AI is rapidly using more energy is that AI is suddenly being used by more people, not that AI stands out as using a lot of energy per person using it. Personal chatbot usage is a tiny fraction of AI’s total energy energy and water footprint, it’s being used for way more. It’s like if the internet had been invented a second time and people were rapidly coming online.

The reason AI data centers use a lot of energy is that they are built to collect huge amounts of individually tiny computer tasks in a single physical place. This makes them more energy-efficient than other ways of doing the same things with computers. If we’re going to do things with computers, we should prefer that data centers manage a lot of it. Every time you interact with the internet, you’re using a data center in the same way you use any other computer. Globally, the average person uses the internet for 7 hours a day, but data centers only use 0.23% of the world’s energy. It’s a miracle of optimization that something we spend half our waking lives on can use less than a 200th of our energy. Computers in general have been ridiculously optimized to use as little energy as possible, so we should assume that the things we do on them will not be significant parts of our carbon footprints. It does not matter for the climate where emissions happen. If I’m right that individuals using chatbots are emitting way less than they would doing other things, then all the emissions caused by chatbots in data centers would have actually still happened, and there would have been a lot more of them, if people boycotted chatbots instead. So chatbots in data centers are often reducing emissions, they just concentrate the reduced emissions so we can see them all in one place. This makes them look bad, but they’re often preventing way more emissions that would just be more dispersed.

Data centers do put more strain on local grids than some other types of buildings, for the same reason a stadium puts more strain on a grid than a coffee shop: the stadium is serving way more people at once. Data centers are building-sized computers that tens of thousands of people are using at any one time. The reason they stand out is that they gather a large amount of aggregate energy demand into a tiny place, not that they’re using a lot of energy per user. In the equation (Total Energy) = (Energy per Prompt) x (Number of Prompts), energy per prompt is low, but the number of prompts in a data center is extremely high, so the total energy they use is high. This means that your personal use of AI is adding extremely tiny amounts of energy demand, and of all the things you can cut to reduce your emissions, it’s one of the very least promising. The fact that chatbots as a whole are using a lot of energy tells you nothing about whether you using it personally is wasteful, for the same reason that tens of billions of dollars are spent on candy bars globally, but you purchasing a candy bar isn’t financially wasteful. Deciding that you’re going to stop using AI for the sake of the climate is like going around your home and randomly unscrewing a single LED bulb, or pausing your microwave a few seconds early to save the planet. It’s so small that it’s a meaningless distraction.

The vast majority of AI’s effects on the environment will come from how it’s used, not from what happens in data centers. Amazon and Google Maps both have big impacts on the climate. Amazon might help or hurt a lot, and Google Maps optimizes a lot of car trips, but also might encourage more driving. But no one in debates about Amazon or Google’s climate impact says “The most important issue is the energy costs of running this website in data centers.” That would be crazy, because the websites are tools that cause people’s behavior to change, which leads to much larger changes in the physical world. If you’re concerned about AI’s impacts on the climate, the main question should be how using AI can help or hurt the climate, not the (tiny) costs of running AI in the first place.

• I was going completely crazy with the number of everyday people I was meeting who suddenly all had wildly inaccurate confident beliefs about AI, energy, and water. All of these beliefs were easily disprovable with simple easy-to-find statistics and comparisons, but I wasn’t finding anyone online doing it. Every piece of news coverage I was reading had ridiculous comparisons (the first one I read literally said that ChatGPT is now using more than twice as much energy as a whole person) and no one doing the simple David MacKay move of actually putting the numbers in context.

• I had a decent background in the general facts involved after following climate change for 15 years and teaching physics for 7 (if you’ve enjoyed my explanations here you can find my full explanation of all high school physics on my YouTube channel) and it was satisfying to use my stored-up knowledge. It was nice to get out some really fundamental ways I think about climate to a big audience.

• I was getting a lot of great feedback. My original two posts (here and here) have been collectively read 230,000 times, and posting the first was the reason my blog took off:

  

  It’s put me in touch with a lot of really cool people in tech and journalism and environmentalism, and has been extremely fun. I’ve developed a reputation as the AI water guy.

  

• This was one of the first topics I was able to do very rapid, complex research using chatbots as aids. A lot of people who don’t use them much don’t know how easy it’s become to just have them make huge catalogs of relevant sources and to double check all the sources they give. Here’s one of many examples:

  

  I feel like I’m writing with a whole team of researchers by my side now. I owe a lot of this blog’s success to AI chatbots. At some point I’ll write an update to how I use them.

Now I’ve developed a pretty huge catalog of takes on AI and environmentalism more broadly. I always wanted a topic where I could imitate Piero Scaruffi and leave a big collection of takes important to me people could explore for themselves, and now I have that.

I’ve run up against a big wall of reasons to take a break from the subject:

• Everyday people are converging on the correct objections. Much more than before, I’ve noticed a lot more people saying “AI water use is mostly fake, personal prompt costs are mostly fake, but AI as a whole is going to put a huge strain on our electrical grid going forward that’s going to matter a lot for the green energy transition, local electricity prices, and air pollution.” To be clear I don’t think I caused that, it’s just that the conversation has found its footing more than it had a year ago.

• I’m working on AI and the environment in other places. I have a few pieces forthcoming in more established places than this blog and want to give them my full “AI and the environment” energy right now.

• I’ve said most of what I want to say about it. Some of the new objections to AI data centers are getting weirder and weirder. Some people are now saying the land use is an issue, which is so wild that I don’t even think it deserves a post on its own (they’re so incredibly compact relative to what they do, and such tiny tiny parts of America’s full land footprint. We have a lot of land!). I’ve addressed what I think I can address, and other stuff (like household electricity prices) are complex enough that I feel like I can’t do them justice with the limited time I have right now. I’ve really really really been thorough in addressing every last point. Every now and then some new big article will come out with some massive incorrect claim about AI and the environment, and every single time I feel like my past blog posts have already addressed all the goofy moves the author is making.

  

  I can have a hit tweet on the topic every month or so when a new article comes out, that takes me ~10 minutes to research and type, and reach a pretty large audience. This leads into the next point.

• I’ve gotten most of the “explore” out of the way, what’s left to do is “exploit.” In the explore-exploit trade-off, I feel like I’ve covered a lot of the ground and learned where most of the contours of the debate now. What’s most impactful to do now is mash a button on the most important points.

  

  This is me saying “It doesn’t use that much water tho” over and over again. Because I feel like I’ve found the most important points, I’ve mostly just been saying the same few simple things (listed here) over and over in response to different specific arguments. This can get old, and is starting to not feel valuable.

• My impact on this topic could be bigger elsewhere. I think if I wanted to actually maximally change the conversation on this, I should make some kind of YouTube overview. I don’t have the capacity for that right now, but if I’m being honest and wanted to build the biggest audience and change the most minds, that’s where I’d do it. If you’d like to run with a lot of my points on YouTube, I think there’s a ton of low hanging fruit in the debate there.

• I’m burned out on this topic. Back when I was teaching physics, the word “physics” had become painful for me to hear, because I’d hear it so often every single day. Now typing the word “data center” is equally painful. I kind of just need a break.

• I’d like more people to be doing this. I really think what I’m doing here is straightforward and a lot of people could be doing the same. I’m flattered to get people messaging me about this.

  

  But I’d also like more people to run with these basic moves I keep making over and over and apply them in new places! I’ve gotten some nice messages like this:

  

  and I’d like more people to do the same. Almost no one anywhere is actually writing about this thoroughly for a broad audience. It’s either dense academic papers of mostly correct takes, or wildly off coverage from people who seem to mostly be following whatever others have already written.

• I should prioritize more important stuff. I am into and a representative of effective altruism, a key tenant of which is that most of your impact in life comes from what you choose to prioritize, and most people very quickly get obsessed with a very local maxima of impact they happen to discover and lose the opportunity to do a lot more good elsewhere. Posting about this has been a fun hobby for me, but I think I need to orient my hobbies in more impactful directions. I’d like to focus more on:

  • Misuse risks from advanced AI systems.

  • How seriously to take the basic x-risk case from advanced AI (my “p(doom)” has always been low but nonzero, and I need to see how justified that is).

  • Animal welfare and ethics.

  • The general DC AI policy space.

  • AI and China, and China in general.

  • General lifestyle posting (I think this short post is one of the best things I’ve written and I’d like to write more like it).

  • Ideas from philosophy that have been important to me.

• I need to give more time to big work projects over the next few months. In the words of one of my closest friends:

  

I’m willing to bet that you haven’t read every last post I’ve written on this topic. If you’d like more posts from me on this, first circle back and read the other ones. Maybe upload them all into an AI model and have them give an AI take from my perspective on a new issue you’re trying to understand. Or best of all, just start doing these deep dives yourself! I’m mostly using high school-level math and physics here.

Thank you to my legions of new followers. I’m so excited to be participating in the debates with all of you about this amazing, wild, sometimes terrifying new technology and its implications for society and the future. It’s crazy to have such a big audience from all over and I’m excited to keep this adventure going. If you ever want to chat, your bar for reaching out to me at AndyMasley@gmail.com should be low.

A bounty for mistakes

In writing this blog, I have also developed many critics. While some have been really useful, others make what I think are wildly false claims about what I’m doing here. Take this one:

All my claims are based on external trustworthy sources I link, and if I link myself it’s only to show an argument I’ve made somewhere else using external sources.

I stand behind everything I’ve posted on this blog. To demonstrate, I offer a simple bounty: I will send you $50 if you can find:

• A legitimate source for a statistic I cite where I lie about the number it’s giving or I get it wrong, which invalidates a core argument I’m making.

• A source for a number I cite that’s illegitimate, and that most legitimate sources disagree with, and this updated information invalidates a core argument I’m making.

If I did a calculation wrong and 0.0000102% should actually be 0.0000103%, I’m not going to consider that to be invalidating. But if I say the median chatbot prompt uses 0.3 Wh, and it actually uses 300 Wh, that does invalidate a core claim I make. A few more rules:

• All claims need to be analyzed by when I said them. If I make a claim about ChatGPT in June, and Sam Altman comes out with a new model in December that’s 100x as energy intensive, my claim in June isn’t wrong, just outdated.

• I’m vetoing very specific disagreements of what counts as “using” water. Whether to count rainfall on crops or into lakes dammed by hydroelectric power stations turns out to be very relevant for the water debates. The way both are reported leave huge error bars in both directions and I’ve tried my best to take a middle ground on them. I’d be interested if you disagree with my takes, but I worry that both extremes have such reasonable people on them that I could go back and forth between sending both $50 as I oscillate between two extremes.

• I have a few pieces I typed fast (like this one) where I haven’t circled around to cite some of the stats in them, because I had gone over them in previous posts. I ask for a reprieve if I don’t cite a source in a newer post but I had cited the source for it previously, and the source is legit.

• I used some rough back of the envelope math for the graphs in my very first post that I want to revisit, so I’m walling those off specifically. Sorry fam.

If you find something, email me at AndyMasley@gmail.com. I’ll be happy to have a back and forth with you if I disagree with your take and why. You can feel free to publicly shame me if I’m evasive. If you’re right, I’ll send you $50 and post the update on a wall of shame. I might be a little slow to get to you, ping me if I don’t respond the first time, publicly shame me if you don’t hear back in a few weeks.

My main worry hasn’t been that AI’s being unfairly attacked (I’m to put it mildly ambivalent about the future of AI). It’s more that readers are coming away with wildly inaccurate beliefs about where AI and data centers fit into the broader environmental picture. Getting this right matters a lot, because it’s very hard to keep people’s attention on climate and the environment for long. If they get distracted by a relatively small issue, they miss the opportunity to do a lot more good elsewhere. Environmental problems are best understood as a series of complex trade-offs. We need to have a correct picture to manage those trade-offs well.

My ask isn’t that you as a journalist present AI positively or negatively, it’s that you aim to give readers an accurate picture of its place in the environmental issue, and not leave them less informed about where the most pressing environmental problems are in their communities and the world more broadly.

These are all my requests. I go into more detail on each below:

• The big ultimate goal: readers should leave your story with a better understanding of how data centers fit into the broader environmental problems a region or the world is facing.

• Don’t ever share contextless large numbers.

• Readers should leave understanding that the energy and water used by individual prompts doesn’t meaningfully add to their personal carbon or water footprints.

• Compare data centers to other industries and commercial buildings, not household use of energy and water.

• Criticize AI specifically, but don’t imply that it’s inherently weird or bad to spend physical resources on digital products.

• Don’t imply that data centers make the computer processes inside less efficient.

• Don’t imply that data centers are new or uncommon.

• Don’t leave readers to infer for themselves that data centers have caused specific catastrophes that haven’t actually happened.

• Water-specific asks

  • Don’t report a data center’s water permit as the amount of water it will actually regularly use.

  • Don’t frame the offsite water use as “hidden” that the companies are dishonestly keeping secret.

  • Don’t use “straining local water systems” or “exacerbating drought” as synonyms for “using any water at all in a high water stress area” without clarifying what the actual harms are.

For context, this recent critical story on data center buildouts checks my boxes for great coverage. I think more writing should aim for this level of quality.

The big ultimate goal: readers should leave your story with a better understanding of how data centers fit into the broader environmental problems a region or the world is facing.

If a reader leaves your story convinced that data centers are the main unique threat to a region’s water, when golf courses in the region are using 30 times more, readers have been misinformed. This often happens in coverage of data centers in Maricopa County Arizona, where data centers are framed as a unique water catastrophe for the region, but are actually using 0.12% of the county’s water, whereas golf courses use 3.8%. If you write a story about Maricopa County or somewhere else, a reader shouldn’t later be shocked to see this graph:

This is a good general rule of thumb. If after reading your article, a reader can feel shocked and confused to see some simple data completely clash with the vibes of what they just read, the article has miscommunicated the problem.

Don’t ever share contextless large numbers.

This is the singular sin of bad writing on the environment. There are 8 billion people. Anything used by even a small fraction of people globally will use huge amounts of resources by the standards of one individual person, to the point that literally anything in society can overwhelm us with its scale. If you don’t give readers a sense of proportion, you’re leaving them in the dark and unable to make good decisions about the environment. Don’t merely write “10 million gallons of water per year” or “200,000 bottles of water per day” in a vacuum. This is a huge amount of water by the standards of any individual, but it’s also about 10% of the water a large car factory might use. If you drop a lot of huge contextless numbers on your readers, they will basically always be alarmed, whether there’s a “thousand” or “million” or “billion” involved. A reader leaving your article should not be surprised to later find out that the data center you described uses as a car factory.

My favorite book on good environmental communication (that I’d assign to everyone covering this if I could) is Sustainable Energy - Without the Hot Air by David MacKay. The book’s available for free online. This quote from the introduction captures how useless and harmful it is to share contextless large numbers:

This heated debate is fundamentally about numbers. How much energy could each source deliver, at what economic and social cost, and with what risks? But actual numbers are rarely mentioned. In public debates, people just say “Nuclear is a money pit” or “We have a huge amount of wave and wind.” The trouble with this sort of language is that it’s not sufficient to know that something is huge: we need to know how the one “huge” compares with another “huge,” namely our huge energy consumption. To make this comparison, we need numbers, not adjectives.

Where numbers are used, their meaning is often obfuscated by enormousness. Numbers are chosen to impress, to score points in arguments, rather than to inform. “Los Angeles residents drive 142 million miles – the distance from Earth to Mars – every single day.” “Each year, 27 million acres of tropical rainforest are destroyed.” “14 billion pounds of trash are dumped into the sea every year.” “British people throw away 2.6 billion slices of bread per year.” “The waste paper buried each year in the UK could fill 103448 double-decker buses.”

If all the ineffective ideas for solving the energy crisis were laid end to end, they would reach to the moon and back.... I digress.

The result of this lack of meaningful numbers and facts? We are inundated with a flood of crazy innumerate codswallop. The BBC doles out advice on how we can do our bit to save the planet – for example “switch off your mobile phone charger when it’s not in use;” if anyone objects that mobile phone chargers are not actually our number one form of energy consumption, the mantra “every little helps” is wheeled out. Every little helps? A more realistic mantra is:

if everyone does a little, we’ll achieve only a little.

Readers should leave understanding that the energy and water used by individual prompts doesn’t meaningfully add to their personal carbon or water footprints.

Simply reporting a ChatGPT prompt uses “ten times as much energy as a Google search” doesn’t give a reader much useful information. It’s fine to report on its own, but it shouldn’t be used to imply that ChatGPT is going to become a noticeable part of a person’s daily carbon footprint. Google searches use tiny amounts of energy. Wouldn’t it be weird to hear a climate scientist say you should limit your Google searches? It’s like saying that a digital clock uses 1 million times as much power as an analog watch. That’s true, but it’s such a minuscule amount of power that they both round to zero. Readers should be left understanding that their individual prompts are not really affecting their carbon or water footprints at all. The average American would have to prompt ChatGPT 1000 times per day to raise their carbon footprint by 1%. They would have to prompt it 8,000 times per day to raise their water footprint by 1%. But if they were doing this, they would likely be skipping other activities that were way more harmful to the environment. I’ve read too many articles that give readers the sense that using chatbots can somehow meaningfully contribute to their individual environmental footprint. Even if you try to include every last way chatbots affect the environment, the numbers just don’t add up. It is wildly misleading to imply that using AI raises your personal carbon or water footprint.

Compare data centers to other industries and commercial buildings, not household use of energy and water.

Because AI and the digital economy are large general industries in America, their proper context is other industries, not large multiples of personal lifestyle things individual people do. It would be weird for me to compare the American auto industry’s energy use to the energy I use in my home. Better comparisons would be to the steel industry, or agricultural industry, or other places where I can understand how car manufacturers fit into the big picture of America’s industrial use of energy and water. Simply reporting “Wow, the auto industry uses millions of times as much energy as Andy! Shocking.” would give readers no useful information.

But most media I’ve seen about AI data centers seems to only ever compare AI energy and water to massive multiples of personal household use, not to other normal industries. Take this recent example from the BBC:

The video announces that all data centers in Scotland are using 27 million bottles of tap water every year. That’s a weird way to talk about a large industry and doesn’t give the viewer any context for how data centers compare to other industries. If we make the comparison, we find that all data centers in Scotland combined are using just 4% of the water used by a single large car factory. That’s just 0.004% of Scotland’s total water use. This is a much more helpful comparison for readers. Many viewers would likely be confused if after watching this video, they discovered that everyone involved had been talking about a single car factory in Scotland. It gives viewers a wildly misleading impression of where water’s actually being used. Because most water use is in agriculture and industry, not in households, you should be sure to give readers context for how much resources data centers are using by comparing them to factories, farms, and other things that use larger amounts of water.

This is a very easy move to make. It doesn’t add many words to your piece, and gives readers a much more complete picture of how data centers compare to other industries.

These are some specific ways data centers are often talked about that other industries aren’t:

• Anything referencing “community water” as in “data centers want to tap into local community water.” Most water in America is used for large scale agriculture and industry, not households. Any industry could be said to be “tapping into community water.” It would be strange to read an article about the Detroit auto industry “tapping into Detroit’s community water” even though the auto industry there uses millions of gallons of water per day. It’s understood that most parts of the country have both industry, commercial buildings, and homes, and in fact places are often benefited by having more industry and commerce.

• Ominous stories of data centers “using more than 20% of a town’s water.” This only sounds shocking if you don’t think of data centers as any other industry. If a factory or large college were built in a small town, and you found out either were using 20% of the town’s water, this would make sense. The main business in the town would use a big chunk of its resources. This can often be beneficial, as it gives local utilities more money to invest back into improving service.

Criticize AI specifically, but don’t imply that it’s inherently weird or bad to spend physical resources on digital products.

The digital economy more broadly has been a gigantic civilizational achievement for transmitting valuable information. It’s part of what’s contributed to dematerialization where over time we’ve had more economic growth using less physical resources and energy. Digital goods are valuable because information is valuable. You can hate AI and think it’s all bad or useless, but don’t imply that it’s wasteful to spend physical resources on digital goods in general. Computers are the most efficient way to deliver information. Information is valuable. The value of a book is mostly in the information it contains, not the physical paper or ink that make it up. It’s okay and good to spend physical resources to acquire information, especially when doing it digitally is saving physical costs elsewhere.

Don’t imply that data centers make the computer processes inside less efficient.

Data centers are the most efficient ways to do large-scale computing. This point seems obvious, but I’ve met a lot of people who follow AI and the environment in their spare time who seem to think that the environment would be helped if we didn’t use data centers and ran more computer processes in our homes. This would actually be pretty terrible, everything we did would use way more energy. I’m not sure where they’re getting this idea, but the ways data centers are framed sometimes seems to contribute.

Don’t imply that data centers are new or uncommon.

Massive AI data centers are new, but readers should also understand that everything they do online besides AI has always used energy and water in data centers. I meet a lot of people who talk as if AI is the only digital product we use that uses water. There are ~5,400 data centers in the country. 99% of Americans live within 50 miles of a data center. They are ubiquitous and normal. Make sure to distinguish massive new AI data centers from all the others.

Don’t leave readers to infer for themselves that data centers have caused specific catastrophes that haven’t actually happened.

A very common move in a lot of reporting on data centers and water is to say that data centers “can create problems for community access to water” or “data centers can drain local aquifers” without once mentioning that of all 5,400 data centers in the US, none of them have done either. It’s true that they technically “can” create problems for access to water, because they use water! This is true of any industry. Anything that’s clearly meant to let a reader infer something has happened that hasn’t confuses them and leaves them less informed.

Water-specific asks

Don’t report a data center’s water permit as the amount of water it will actually regularly use.

Many (I’d say the majority) of articles I’ve read on data center water use frame the data center’s water permit as the amount of water it will actually use. This is wildly misleading. Water permits are difficult to amend, so data centers request the absolute most water they would need in extreme situations to give themselves a safe upper bound.

Don’t frame the offsite water use as “hidden” that the companies are dishonestly keeping secret.

The companies that run data centers often don’t have access to exactly how much water the power plants they draw from on the grid use. If a data center reports “We will use x amount of water” to the public, it seems reasonable for the public to assume that they are only talking about the water used in the building. If data centers become more efficient, that means that they use less water relative to the power plants around them. But this also means that the “hidden” cost in the power plants is a much larger multiple of the data center’s water. Talking this way effectively punishes the most water-efficient data centers, and implies they’re actually worse because they have a much larger proportion of “hidden” water offsite. This is backwards. I explain this more here.

Don’t use “straining local water systems” or “exacerbating drought” as synonyms for “using any water at all in a high water stress area” without clarifying what the actual harms are.

Literally everything that uses water in a drought-prone area is “exacerbating the drought” and “straining the local water system.” But using those terms can give readers an inaccurate picture if they’re using an extremely small amount of the water. There are many articles referencing data centers “exacerbating drought” in Phoenix that don’t mention they’re only using 0.1% of the water there. Find more specific ways to get across how much they’re adding or not adding to the problem.

Everything I’m saying is actually just a few points repeated over and over

There are actually just a few simple points I’m trying to drive home in all these posts, expressed over and over in different situations:

• For your personal footprint, compare AI to your total carbon and water footprint and see if it has any effect at all. Don’t compare it to incredibly efficient tiny things like Google searches.

• For local and global impact, compare AI to other industries and applications, not to your personal lifestyle. Manufacturing iPhones uses huge amounts of power in total. It would be ridiculous to compare manufacturing all iPhones to the number of flights an individual person takes. Creating a new product for a billion people is a more relevant comparison to training an AI model. Many data centers use the same amounts of water as other industries etc. Always contextualize large numbers.

• The collective environmental footprint of society is made up of the footprints of individuals. We should behave in the ways we’d want the group to behave. We should prioritize the things that actually help climate the most, because some things have many orders of magnitude more effect than others.

• All environmental and climate impacts of all industries are “new” in the sense that every day, all industries add new carbon emissions and air pollution to the environment. Treating AI data centers as special just because the buildings themselves are new normalizes much larger sources of emissions and harm that coincidentally were already happening.

• AI and the environment is confusing to think about because of the environmental paradox of data centers: they both put uniquely large concentrated demand on local grids, but are also tiny efficient parts of the global energy grid. They can look big and evil on the ground, but are actually remarkably efficient.

• In general, computing is one of the most resource-efficient parts of society.

• Most of AI’s effects on the environment will probably be caused by how it’s used, not by the physical operation of the data centers where it’s run.

All of these points seem really simple individually, but I wasn’t finding any popular commentary on AI and the environment online that checked all these boxes in the way I wanted, which was a big motivation to post about it more. There are a lot of great industry reports written for more technical audiences, but a lot of the popular coverage I was reading went against basically all of these points.

Why these posts are so long

I’m capable of writing short posts. The reason the AI and environment posts tend to be so long is that for each one:

• I find very few other places online collecting what I want to say about the topic in a single place. I figure it might be useful to other people to have these points collected in individual places rather than spread across 20 posts.

• I’m faced with a choice. If critics read my post, they’re going to come away with one of two takes based on how much I write:

  

  Between these two, I’d rather just aim beyond that red line, clearly label the contents of the post, give readers the option to skip around, and make it really long rather than come across as dodging some important point. Still, I find that a lot of people still announce that I’m dodging some point that I very clearly label. One of the most common reactions to my first post on this was “How convenient, he left out the cost of training where most of the energy is used” even though I had a clearly labeled training section. Can’t win ‘em all.

• There is just a ton of nuance to get across in getting these basic points across, and I want to make it clear where I’m coming from. Since college I’ve worried that the environmentalist movement is too quick to label things “whataboutism” and was finding that was a very common objection, so getting down what I think about that and why I worry environmentalists overshoot here seems useful. I have a backlog of 15 years of following and reading about climate and environmentalist ethics, so I’d rather get it out there if it’s relevant.

• I could just leave people with that list of basic ideas above, but I keep having a lot of positive reactions and feedback to each new post, so it seems like actually doing the deep dives is providing value.

This isn’t actually the single most important problem

The reason I’ve been writing about this a lot more recently is mainly that I’m finding I’m having some nonzero influence on the conversation and building up a big audience who seem to value it. Both my original post and cheat sheet post each have 100,000 views now, definitely the most anything long I’ve written has gotten. 3 of the 5 top images for “ChatGPT and environment” are from my blog. Neat!

This has been a really fun surprise. I’m mashing this button as long as I think I can actually add new useful stuff to the conversation and get positive reception. Older readers know that I used to blog about a much larger range of stuff. I want to get back to that. I’ve been pretty busy recently and in choosing what to write about, it does still seem like there are some big gaps in the AI & environment debate I’d like to focus on addressing. But there are a limited number, and once I have more time I want to broaden what I post about again. This is temporary.

More people should be doing this!

I’d really like more popular writing on AI and the environment from people who can go beyond “Oh AI’s weird and new, and is using a lot of energy and water. Must be bad, end of story.” There’s a lot of great technical writing on AI and the environment, but things aimed at regular people are still shockingly scarce. If you do a YouTube or TikTok search, most of what you see is pretty misinformed. If you have skills in either, that might be a niche you can fill.

I presume most journalists are honest, but also might be stretched thin or in a topic they don’t know much about

I don’t want to come off as implying that there’s a vast journalistic conspiracy against data centers. I’ve been roasting a lot of recent coverage, but I think the problems I’m identifying mostly come from journalists being somewhat new to the issue, and running with previously established common wisdom about data centers, like the idea that they use a lot of water. Less charitably, it might be that big weird new buildings using a lot of power and water just gets clicks, but I presume almost everyone covering this is authentically interested in the truth and is just slipping up because the public understanding is bad.

There are exceptions, but I really don’t want to come off as implying that journalists as a whole are actively choosing to be deceptive about this. I think they’re sometimes just taking shortcuts based on pre-existing coverage.

What would cause me to think that AI is actually bad for the environment

I would change my mind about AI and the environment if I saw good reasons to believe 1 of these 4 points:

1. Using AI raises individual people’s carbon or water footprint in expectation.

   The numbers are so tiny that if people switch from using AI to most other activities their emissions probably go up. Right now the numbers imply that if AI is using a large amount of your energy and water, you’re using it so much that you’re not doing other activities that are much worse for the environment.

2. AI data centers are using an inordinate amount of power and water relative to the actual number of people using them (hundreds of millions), the tax revenue they generate for the communities they’re built in, or other major American industries.

3. Problems specific AI data centers have created for local communities (xAI’s data center in Memphis might have added significant pollution) couldn’t be easily fixed with strict regulation and pricing.

4. AI isn’t actually valuable at all, or at least isn’t valuable per unit of energy and water it costs. This is the place I’m least likely to change my mind. I use chatbots hourly at least. They’re a big reason I’ve been able to learn so much about this so fast in the first place.

Something I’m not willing to entertain is the point that AI is bad for the environment because it emits at all, or uses any water. Under that definition, literally everything we do is bad for the environment, and saying “AI is bad for the environment” is identical to “AI is something people do.” I have a more detailed breakdown of the definition of “bad for the environment” I’m using here.

The idea that AI’s water usage is a serious national emergency caught on for three reasons:

• People get upset at the idea of a physical resource like water being spent on a digital product, especially one they don’t see value in, and don’t factor in just how often this happens everywhere.

• People haven’t internalized how many other people are using AI. AI’s water use looks ridiculous if you think of it as a small marginal new thing. It looks tiny when you divide it by the hundreds of millions of people using AI every day.

• People are easily alarmed by contextless large numbers, like the number of gallons of water a data center is using. They compare these large numbers to other regular things they do, not to other normal industries and processes in society. They aren’t aware of how much water society uses on other normal industries.

Together, these create the impression that AI water use is a problem. It is not. Regardless of whether you love or hate AI, it is not possible to actually look at the numbers involved without coming to the conclusion that this is a fake problem. This problem’s hyped up for clicks by a lot of scary articles that completely fall apart when you look at the simple easy-to-access facts on the ground. These articles have contributed to establishing fake “common wisdom” among everyday people that AI uses a lot of water.

This post is not at all about other issues related to AI, especially the very real problems with electricity use. I want to give you a complete picture of the issue. I think AI and the national water system are both so wildly interesting that they can be really fun to read about even if you’re not invested in the problem.

Importantly, I am not saying that it’s impossible for data centers to ever cause any problems with water. They require careful planning in the same ways other large industrial buildings do. What I am saying is that right now no reasonable forecasts imply that data centers will rise to become a significant problem for water access in the US. Almost all complaints about their national water use are basically just saying “We should not have a new large industry in America using water.” The tax revenue per gallon from data centers is just so high that in many places they are among the best new buildings possible to benefit a community experiencing water scarcity, because any other industries using the same amount of water would generate way less tax revenue. Critics of data centers need to carefully weigh their water costs against the massive amounts of revenue they can bring in for everyday people, not just look at the water costs alone. The debate about building data centers should involve reasonable conversations between ecologists, economists, and city officials, not everyday voters shouting down local meetings with misleading statistics.

Contents

• A few important definitions

• AI water use isn’t an issue on the national, local, or personal level

  • National

  • Local

  • Personal

• How big of a deal is it that data centers use potable water?

• Do data centers poison water supplies?

• Using AI can save way more water than is used in data centers

• It’s okay to use water on a digital product

• The social value or harm of a tool isn’t the final word on how harmful it is to the environment

• There’s a trade-off between water and energy for data center cooling systems. For the climate, water’s often preferable

• What about all those news stories about AI harming local water access?

  • Every popular article about how AI’s water use is bad for the environment in the last year has had a wildly misleading framing

  • 5 common misleading ways of reporting AI water usage statistics

  • Some examples of great news coverage

• Further reading

A few important definitions

Suppose I take a cup of water from a lake, and then immediately dump it back in. That doesn’t seem bad. Now I take a cup from the lake, and this time I evaporate it. That seems worse. Now I take a cup and spend some resources on making it drinkable. These all have very different costs and effects on the water system. We need words to describe it.

Water is complicated, but not too complicated. There are a few key definitions to understand. First:

• Consumptive use removes water from a local system. Taking the cup of water and evaporating it is consumptive use. Evaporated water mostly does not return to its original source.

• Non-consumptive use temporarily takes water from a local system, and returns it later unaffected. Taking the cup of water from the stream and pouring it back in is non-consumptive use.

Growing food is an example of consumptive use. Some of the water becomes part of the food itself. When the food is shipped away, the water leaves the local system.

Many data centers rely on evaporative cooling. This is the way water is consumptively used in data centers. They do not immediately evaporate all the water they are using. Most of it circulates through the cooling system repeatedly before evaporation.

Many reports on AI’s water use do not only include water in data centers, they also include the water consumed by the power plants data centers draw from. This leads to a second important distinction:

• Direct use: The water used inside AI data centers themselves to cool servers.

• Indirect use: The water used in nearby power plants to generate electricity the data center uses.

The U.S. electricity sector uses approximately 50 trillion gallons of water each year — enough to cover all of Pennsylvania in five feet of water. However, most of that use is non-consumptive.

The average kilowatt-hour takes 4.35 L (1.15 gallons) of water to generate. Unless electricity is sourced entirely from wind and solar, most activities have an indirect water cost. A digital clock has a direct water cost of zero, but an indirect water cost of 0.2 L of water per day. Every three days, the clock accounts for a bottle of water’s worth of consumption at a nearby power plant.

It is standard practice in environmental reporting on AI to include both onsite and offsite water usage. On average, the water data centers use to cool servers is only about one-fifth of the water required to generate the electricity the servers use. Put another way, almost all (80%) the reported water used by AI occurs during the generation of electricity. Data centers primarily consume water indirectly, similar to any other industry.

Consumptive use can harm total access to freshwater, but freshwater sources are also regularly being replenished. The full balance of and access to freshwater is beyond the scope of this report. The key takeaway is that evaporating freshwater removes it from the local supply, although new freshwater is added elsewhere. There is ongoing debate regarding the severity of this issue for total U.S. freshwater supplies. The U.S. has more abundant and cheap freshwater than most other countries.

Another key definition is potable water: freshwater treated enough that it’s safe for human consumption. It’s not actually very costly to turn freshwater potable. It’s much more costly to turn saltwater potable. Most water used in physical data centers themselves is potable, because it needs to be very clean to flow through the cooling systems without harming the pipes.

So of the ways AI uses water:

• The vast majority (maybe 90%) is withdrawn, freshwater (not potable) that is indirectly (offsite) used non-consumptively in power plants (it’s returned to the source unaffected)

• Less (maybe 7%) is withdrawn freshwater (not potable) that is consumed (evaporated) indirectly (offsite) in the power plants to generate the electricity AI uses.

• And less (maybe 3%) is withdrawn freshwater that’s then treated to become potable, used directly (onsite) in physical data centers themselves, and consumed after (not returned to the source, evaporated).

This post is mainly going to focus on the ways AI causes water consumption, not non consumptive withdrawals, because removing water from a source is a significantly bigger problem than temporarily using it and then returning it unaffected.

There are a lot more terms in water management, but these are most of what you need to understand the issue. Another key point is that data centers don’t meaningfully pollute the water they use, so terminology around pollution doesn’t come up.

Finally, it is important to note that most activities consume water. Most U.S. electricity is generated by heating water to spin turbines. Water is used in the production of most physical objects. The majority of the average person’s consumptive water footprint stems from food consumption. Domestic water use accounts for less than 8% of an American’s daily footprint. An American’s total consumptive water footprint is approximately 422 gallons of fresh water per day.

AI water use isn’t an issue on the national, local, or personal level

National

All U.S. data centers (which mostly support the internet, not AI) used 200–250 million gallons of freshwater daily in 2023. The U.S. consumes approximately 132 billion gallons of freshwater daily.1 The U.S. circulates a lot more water day to day, but to be extra conservative I’ll stick to this measure of its consumptive use, see here for a breakdown of how the U.S. uses water. So data centers in the U.S. consumed approximately 0.2% of the nation’s freshwater in 2023. I repeat this point a lot, but Americans spend half their waking lives online. A data center is just a big computer that hosts the things you do online. Everything we do online interacts with and uses energy and water in data centers. When you’re online, you’re using a data center as you would a personal computer. It’s a miracle that something we spend 50% of our time using only consumes 0.2% of our water.

However, the water that was actually used onsite in data centers was only 50 million gallons per day, the rest was used to generate electricity offsite. Most electricity is generated by heating water to spin turbines, so when data centers use electricity, they also use water. Only 0.04% of America’s freshwater in 2023 was consumed inside data centers themselves. This is 3% of the water consumed by the American golf industry.

How much of this is AI? AI uses approximately 20% of the electricity in data centers, though this is obviously forecast to increase significantly. Water use roughly correlates with electricity, because offsite water is used exclusively for generating the electricity data centers use, and onsite water is used to cool chips after electricity flows through them and heats them up. For now I’ll assume AI also uses 20% of the total water data centers use, with large error bars.

So AI consumes approximately 0.04% of America’s freshwater if you include onsite and offsite use, and only 0.008% if you include just the water in data centers.

So AI, which is is now built into every facet of the internet that we all use for 7 hours every single day, that includes the most downloaded app for the 7 months straight, that also includes many normal computer algorithms beyond chatbots, and that so many people around the world are using that Americans only make up 16% of the user base, is using 0.008% of America’s total freshwater. This 0.008% is approximately 10,600,000 gallons of water per day.

I’m from a town of 16,000 people. It looks like this:

All AI in all American data centers is collectively using 8 times as much water as the local water utility in my town provides to consumers.2 You should be exactly as worried about AI’s current national water usage as you would be if you found out that 8 additional towns of 16,000 people each were going to be built around the country.

Here’s data center water use compared to a lot of other American industries:

Note: the steel number on this graph is I think incorrect, my source misread withdrawals as consumption. The mining number is also withdrawals. Everything else is correct. I’ll circle back on this later to update the graph.

And here’s a comparison to how much water different American agricultural products use, the main way water is used in America. All of these are the amount of water irrigated crops use and aren’t measuring rainwater:

Forecasts imply that American data center electricity usage could triple by 2030. Because water use is approximately proportionate to electricity usage, this implies data centers themselves may consume 150 million gallons of water per day onsite, 0.12% of America’s current freshwater consumption.

So the water all American data centers will consume onsite in 2030 is equivalent to:

• 8% of the water currently consumed by the U.S. golf industry.

• 8% of the water consumed by U.S. steel production.

  • Since writing this I think the source significantly overestimated US steel production’s water consumption and this is actually steel’s water withdrawal, which is very different.

• The water usage of 260 square miles of irrigated corn farms, equivalent to 1% of America’s total irrigated corn.

If you found out that U.S. steel production was expected to increase by 8% in 2030, the amount that would cause you to worry about water is how worried you should be about data center water usage by 2030.

How much of this will be AI? Almost all this growth will be driven by AI, but because AI is only 20% of data center power use, its growth will have to be huge to triple total power usage. One forecast says AI energy use in America will be multiplied by 10 by 2030. Because water use is proportionate to energy use, we can multiply AI’s water use by 10 as well.

So in 2030, AI in data centers specifically will be using 0.08% of America’s freshwater. This means it will rise to the level of 5% of America’s current water used on golf courses, or 5% of U.S. steel production, or be about 173 square miles of irrigated corn farms.

The average American’s consumptive lifestyle freshwater footprint is 422 gallons per day. This means that in 2023, AI data centers used as much water as the lifestyles of 25,000 Americans, 0.007% of the population. By 2030, they might use as much as the lifestyles of 250,000 Americans, 0.07% of the population. Not nothing, but 250,000 people over 5 years is just 4% of America’s current rate of population growth. If you found out that immigration plus new births in America would increase by 4% of its current rate, would you first thought be “We can’t afford that, it’s way too much water”?

This is more contentious, but all this is in the context of AI potentially boosting U.S. and global GDP by whole percentage points. Most forecasts imply that AI will boost total U.S. GDP by at least 1%. If we judge industries by how much they’re contributing, data center direct onsite usage will collectively be 0.08% of consumption by 2030, but contributing at least 1% to GDP. Maybe this won’t happen, but in worlds where this doesn’t happen, AI companies won’t be able to afford a huge buildout either. If AI is a bubble, the bubble will have to pop sometime before AI data center water usage hits 10x what it currently is in America. Your predictions for how much water AI will use and your predictions for how much real economic value it’s going to provide have to be related in some way.

Local

Data center operational use of water has had minimal impact on water access where they're built

Because data centers are using the same normal amounts of water as many other industries, there is only one place I’ve found where data centers seem to have partially contributed to a rise in water costs: Newton County, Georgia, where the NCWSA’s own documents cite the second Meta campus as one factor driving accelerated capital improvements. Even there, Meta is one of several large industrial users alongside a pharmaceutical plant and a Rivian factory, and other factors like inflation and housing market uncertainty are also driving the increase. Everywhere else I’ve looked, data centers have not raised household water costs or harmed local water access. I have a much longer deep dive on that here. I won’t repeat all the arguments. If you’re skeptical, I’d suggest reading that first.

The only exceptions to this rule are the construction of data centers, which has in a few place caused issues for local groundwater. This is bad, but it’s purely an issue of constructing a large building. It has nothing to do with AI specifically, for the same reason that debris from a bank being constructed would tell you nothing about how banks normally impact a community. There’s a famous New York Times headline that comes up in most conversations of AI and water use:

But the reason their taps ran dry (which the article itself says) was entirely because of sediment buildup in groundwater from construction. It had nothing to do with the data center’s normal operations (it hadn’t begun operating yet, and doesn’t even draw from local groundwater). The residents were wronged by Meta here and deserve compensation, but this is not an example of a data center’s water demand harming a local population.

Basically every single news story that’s broken about this has been misleading for similar really simple reasons it’s easy to cross check and verify. I’ve written up my issues with most major news coverage of AI water below. You don’t have to take my word for it, you can look at each one and see if I’m right or wrong for yourself.

The Georgia data center is only using ~2% of the county’s water. For comparison, a pharmaceutical manufacturing plant is using ~4% of the county’s water. A construction plant for Rivian cars is using about the same amount of water as Meta’s data center. The data center is functioning like any other normal industry in the county.

No matter where you look, whether it’s the place with the highest percentage of local water going to data centers (the Dalles, Oregon) or the place with the most water in total going to data centers (Loudoun County, Virginia) or the place with the highest water stress where lots of new data centers are being built (Maricopa County, Arizona), data centers are not negatively impacting local’s freshwater access , because they are behaving like any normal private industry. You can follow each link in the parentheses for a breakdown of how that county uses water and how data centers affect them.

The only difference is that data centers contribute way way way more tax revenue per unit of water used than most other industries. Take Maricopa County in Arizona. The county is home to Phoenix, and is in a desert where water is pumped in from elsewhere. It’s also one of the places in the country where the most new data centers are being built.

Circle of Blue, a nonprofit research organization that seems generally trusted, estimates that all data centers in Arizona will use 905 million gallons of water in 2025. For context, Maricopa County golf courses use 29 billion gallons of water each year. In total, the county uses 2.13 billion gallons of water every day, or 777 billion gallons every year. Data centers make up 0.12% of the county’s water use. Golf courses make up 3.8%.

Data centers are so much more efficient with their water that they generate 50x as much tax revenue per unit of water used than golf courses in the county:

So even though data centers are using 30x less water than golf courses, they bring in more total tax revenue:3

Some people see this, and react with something like “Well I don’t think golf courses OR data centers should be built in the desert.” At some point this becomes an argument against anyone living in deserts in the first place. If you want to have a gigantic city in the desert, like Phoenix, that city needs some way of supporting itself with taxes, and giving jobs to the people who live there. Most industries use significant amounts of water. If Phoenix is going to exist, it’s going to need private industries built around it that are using some water. We have two options here:

• Build industries that generate huge amounts of tax revenue relative to the water they use. Data centers fall into this category (though they don’t provide many jobs).

• Do not build cities in the desert in the first place.

Arguments against data centers existing in the desert because they harm water systems there also often apply to building cities in the desert in the first place. It’s fine and consistent to say that Phoenix shouldn’t exist because it’s unnaturally pumping water from hundreds of miles away, but it’s inconsistent to say that Phoenix should exist, that its water bills should be kept as low as possible, but also that no industries that use any water should be built there.

Just to drive this home, suppose Arizona announced “We are not going to increase the water use in the state at all, but we are going to close all golf courses and replace them with data centers that use the exact same amount of water.” Doing this at the rate data centers generate tax revenue per gallon would lead to an additional 42 billion dollars in tax revenue each year. That’s almost double Arizona’s total state tax revenue in 2023 ($28 billion). The most important thing for desert regions that still want taxable industry is to replace very wasteful, water inefficient industries with very water efficient, revenue-generating industries. It makes sense to worry about adding more total water demand in these regions, but it does not make sense to single out specific industries because they’re a new use of water if they’re competing with and hopefully pricing out less water efficient industries. Data centers seem preferable as a way of generating tax revenue in deserts compared to most other large industries and commercial buildings.

In low water scarcity areas, adding a large new buyer to a water system doesn’t necessarily harm other customers. Water utilities are natural monopolies with large fixed costs, and economic research shows significant economies of scale in water production — a 1% increase in production volume is associated with a 0.16% decrease in unit production costs. Whether this translates into lower retail rates for households depends on how the utility structures its rates across customer classes. In practice, most U.S. utilities use cost-of-service rate structures that separate residential and commercial/industrial customer classes, so a data center’s water costs are generally allocated to the commercial class rather than passed through to households. And where data centers have negotiated development agreements, they’ve often funded infrastructure upgrades that benefit the broader system.

In high water scarcity areas, city and state leaders have already thought a lot about water management. They can regulate data centers the same ways they regulate any other industries. Here water is more zero sum, but data centers just end up raising the cost of water for other private businesses, not for homes. Data centers are subject to the economics of water in high scarcity areas, and often rely more on air cooling rather than water cooling because the ratio of electric costs to water costs is lower.

This seems fine if we think of data centers as any other industry. Lots of industries in America use water. AI is using a tiny fraction compared to most, and generating way, way more revenue per gallon of water consumed than most. Where water is scarce, AI data centers should be able to bid against other commercial and industrial businesses for it. So far, I haven’t seen any arguments against building data centers in high water stress areas that aren’t basically saying “we shouldn’t have any industries at all in places with high water stress” which seems wrong. People still choose to live in places like Phoenix and expect to have strong local governments that need a big tax base to function well. If you’re against industry in high water stress areas period, you need to be against people living in Phoenix in the first place, which means their water bills should probably rise anyway.

There are many cases of data centers being built, providing lots of tax revenue for the town and water utility, and the locals benefiting from improved water systems. Critics often read this as “buying off” local communities, but there are many instances where these water upgrades just would not have happened otherwise. It’s hard not to see it as a net improvement for the community. If you believe it’s possible for large companies using water to just make reasonable deals with local governments to mutually benefit, these all look like positive-sum trades for everyone involved.

Here are specific examples:

• The Dalles, Oregon - Fees paid by Google fund essential upgrades to water system.

• Council Bluffs, Iowa - Google pays for expanded water treatment plant.

• Quincy, Washington - Quincy and Microsoft built the Quincy Water Reuse Utility (QWRU) to recycle cooling water, reducing reliance on local potable groundwater; Microsoft contributed major funding (about $31 million) and guaranteed project financing via loans/bonds repaid through rates. These improvements increase regional water resilience beyond the data center itself.

• Goodyear, Arizona - In siting its data centers, Microsoft agreed to invest roughly $40–42 million to expand the city’s wastewater capacity—utility infrastructure the city highlights as part of the development agreement and that increases system capacity for the community.

• Umatilla/Hermiston, Oregon - Working with local leaders, AWS helped stand up pipelines and practices to reuse data-center cooling water for agriculture, returning up to ~96% of cooling water to local farmers at no charge.

I could go on like this for a while. Maybe you think every one of these is some trick by big tech to buy off communities, but all I’m seeing here is an improvement in local water systems without any examples of equivalent harm elsewhere

What about water pollution?

AI data centers are not a notable source of water quality pollution in their host communities. Their cooling water is typically kept in closed loops, any periodic blowdown is routed to a sanitary sewer for treatment or discharged under numeric permit limits, and an increasing share of facilities use highly treated recycled water that would otherwise be released by wastewater plants. By contrast, the largest water quality problems in the United States come from sectors like agriculture and construction.

The EPA’s national assessments repeatedly identify agriculture as the leading source of impairment for rivers and streams due to nutrient and sediment runoff, with continued nitrogen and phosphorus problems that affect drinking water and coastal ecosystems. Construction activity is also a well documented source of sediment discharges if not controlled. Data centers are not flagged by EPA in these national problem lists, and they do not handle the kinds of process chemicals or waste streams that typify the industrial categories with effluent guidelines.

This makes sense when you think about it. Data centers are just big computers. Water just runs through them to cool them, in the same way your laptop needs to be cooled by a fan. Why would using water to cool a big computer significantly pollute the water? You want the water interfering with the physical material of the big computer as little as possible. There are some

How should local communities think about new data centers being built near them?

Data centers have an impact on local water systems, just like any other private industry. They shouldn’t just randomly be built anywhere. Local communities should consider the costs and benefits. But in doing this, they need to consider the actual amounts of water data centers will use compared to other normal industries, not compared to individual lifestyles.

Obviously, questions about electricity or pollution are real and should be considered separately, but at least in terms of water, I’ve found only one example (Newton County, Georgia) where data center expansion appears to have partially contributed to rising water costs, and even there it’s one factor among several. In contrast, there are many places where data centers have benefited local water access, and a universal pattern of huge tax revenues.

Personal

I think a lot of people don’t realize how much water we each use every day. Altogether, the average person’s daily water footprint is 422 gallons, or 1600 liters. This is mostly from agriculture to grow our food, manufacturing products we use, and generating electricity. Only a small fraction is the water we use in our homes.

Our best current data on AI prompts’ water use from a thorough study by Google, which says that each prompt might only use ~2 mL of water if you include the water used in the data center as well as the offsite water used to generate the electricity.

This means that every single day, the average American uses enough water for 800,000 chatbot prompts. Each dot in this image represents one prompt’s worth of water. All the dots together represent how much water you use in one day in your everyday life (you’ll have to really zoom in to see them, each of those rectangles is 10,000 dots):

However, that 2 mL of water is mostly the water used in the normal power plants the data center draws from. The prompt itself only uses about 0.3 mL, so if you’re mainly worried about the water data centers use per prompt, you use about 300,000 times as much every day in your normal life. That’s the same water your local power plant uses to generate a watt-hour of energy, enough to use your laptop for about 2 minutes. So every hour that you use your laptop, you’re using up 30 chatbot prompt’s worth of water in a nearby power plant.

Have you ever worried about how much water things you did online used before AI? Probably not, because data centers use barely any water compared to most other things we do. Even manufacturing most regular objects requires lots of water. Here’s a list of common objects you might own, and how many chatbot prompt’s worth of water they used to make (all from this list, and using the onsite + offsite water value):

• Leather Shoes - 4,000,000 prompts’ worth of water

• Smartphone - 6,400,000 prompts

• Jeans - 5,400,000 prompts

• T-shirt - 1,300,000 prompts

• A single piece of paper - 2550 prompts

• A 400 page book - 1,000,000 prompts

If you want to send 2500 ChatGPT prompts and feel bad about it, you can simply not buy a single additional piece of paper. If you want to save a lifetime supply’s worth of chatbot prompts, just don’t buy a single additional pair of jeans.

Because generating electricity in America often involves water, anything that you do that uses electricity often also uses water. The average water used per kWh of electricity in America is 4.35 L/kWh, according to the Lawrence Berkeley National Laboratory. This has a few weird assumptions (explained here), so to be conservative I’ll divide it in half to 2 L/kWh. This means that every kWh of electricity you use evaporates the same amount of water as 1000 chatbot prompts (including both onsite + offsite water cost). Conveniently, 1 prompt’s water per Watt-hour.

Here are some common ways you might use electricity, and how many AI prompts’ worth of water the electricity used took to generate:

• Playing a PS5 for an hour - 200 prompts’ worth of water

• Using a laptop for an hour - 50 prompts’ worth of water

• An LED light bulb on for an hour - 6 prompts

• A digital clock for an hour - 1 prompt

• Heating a kettle of water - 125 prompts’ worth of water (the kettle itself has enough water for ~500 prompts)

• Heating a bath of warm water - 5000 prompts (the bathtub itself has enough water for 80,000 prompts)

If you want to reduce your water footprint, avoiding AI will never make a dent. These numbers are so incredibly small it’s hard to find things to compare them to. If you send 10,000 chatbot prompts per year, the water used in AI data centers themselves adds up to 1/300,000th of your total water footprint. If your annual water footprint were a mile, 10,000 chatbot prompts would be 0.2 inches.

Even if AI is completely useless, and all the water you use on it is “wasted,” literally everything else you do in life involves larger amounts of wasted water, even if that activity is really valuable for you. If you boil water to make healthy food, you could make the exact same amount of food if you took a water dropper and extracted a few milliliters from the pot. A medium-sized pot can hold ~10 liters of water. That’s enough for 5000 chatbot prompts. If you took a dropper and removed 1/5000th of the water from the pot, you would still be able to use it to make whatever food you want, so that extra amount is “wasted” but it’s so small that it doesn’t matter at all. If you saw someone filling a pot like this:

and then using this dropper to remove tiny amounts of water at a time to “save” as much water as they could until the pot only contained exactly as much water as they needed so that they didn’t waste a drop,

you would correctly say that was a huge waste of time. In no other places do we think it’s reasonable to worry about “wasting” such a tiny amount of water in our personal lives. Even if you think all water used on AI is completely wasted, it still shouldn’t bother you any more than the fact that people don’t spend the time to remove tiny drops of water from the pots they boil to make food.

How big of a deal is it that data centers use potable water?

One concern about AI is that unlike a lot of the other industries I listed, it’s using potable water: water treated to specific quality standards such that it’s safe for people to drink. Crops basically never use potable water—irrigation relies on untreated surface water and groundwater, not municipal drinking water supplies. Maybe potable water is so much more valuable and scarce that, even though AI isn’t using much, there’s not too much potable water anyway, so AI can still create a significant problem for water access?

According to the USGS’s most recent comprehensive water use report, America’s total water withdrawals in 2015 were about 322 billion gallons per day, of which 281 billion were freshwater. Of that freshwater, about 39 billion gallons per day were withdrawn by public supply systems—utilities that treat water to potable standards and deliver it to homes and businesses. That 39 billion is what consumers buy from local water utilities.

How difficult is it to turn those roughly 242 billion gallons of non-potable freshwater into potable water? If it’s extremely easy, the distinction doesn’t matter much. It would be kind of like having money in two different checking accounts where it’s easy to move it back and forth, but saying we should worry a lot about one checking account being small regardless of the size of the other.

After looking into this, I’ve found that not only is it straightforward to turn freshwater into potable drinking water using conventional treatment processes, but also that the water utility sector exhibits well-documented economies of scale: larger water systems produce, treat, and deliver water at lower unit costs (dollars per gallon) than smaller systems because of efficiencies in the use of labor, raw water supply, and water treatment. A National Academies report on water utility privatization found that as water utilities face higher treatment costs, the desirability of achieving economies of scale becomes more pronounced, and that technological improvements have made larger regional water systems increasingly cost-effective. Research on water treatment costs has confirmed that unit treatment costs are higher for smaller municipalities and decline as facility size increases. Studies of water utility pricing have consistently found that larger utilities serving more customers tend to have lower per-unit prices, reflecting these scale economies.

What this means is: if any region has abundant freshwater but limited potable water, the best way to make potable water more available and cheaper is to introduce a new large, consistent buyer, which will give the local utility enough revenue to upgrade and expand their treatment facilities. Saying that my data is misleading because AI “only uses valuable potable water” actually gets the issue backwards: In regions with abundant freshwater, the potable water 'bottleneck' is treatment and delivery infrastructure, not the raw resource. Research consistently finds economies of scale in water systems, larger communities tend to have lower per-unit water costs because fixed infrastructure costs are spread across more users. This suggests that adding a large, consistent buyer to a water system in a water-abundant area is more likely to support infrastructure investment than to strain the system, though outcomes depend on local rate structures and capacity.

To use a metaphor, imagine that freshwater supply in a region is like jewels in a mine, and potable water is like polished, carved jewels. We can imagine two cases:

Case 1: There is a near-infinite amount of jewels in the mine, but very few are being mined, polished, and carved.

Case 2: There are not many jewels in the mine.

Imagine that demand for polished jewels suddenly skyrockets in each case.

In the first case, companies realize they can make a lot of money if they send a lot more miners in and build a lot of infrastructure to polish and carve the jewels. After a year, 100 times as many miners, carvers, and polishers are working to produce jewels. In this situation, would you expect the price of polished carved jewels to rise or fall? It seems obvious that here the price would fall, due to economies of scale. There are enough jewels in the mine that their supply doesn’t affect price, and all the revenue from jewel sales can be invested in more efficient systems of producing them. You would expect as a buyer to have to pay less and less for a jewel. In the same way, where freshwater is plentiful, treated water becomes less expensive as more people buy it.

In the second case, the supply of jewels in the mine drops off quickly. Raw, uncarved jewels become more and more expensive to obtain. Even if the cost of carving them is low, the total cost of the jewels will rise more and more due to the supply going down. Here, more demand for jewels causes the price to go up, but only because the raw jewels themselves are getting more expensive. In the same way, where freshwater is scarce, treated water becomes more expensive as more people buy it, but only because the freshwater itself is becoming more expensive and scarce.

So even before looking at the numbers, people complaining that AI requires precious rare drinking water are actually getting the issue completely backwards. This comes up very often. Karen Hao, who wrote Empire of AI, regularly emphasizes in interviews (and the book itself) that data centers are not just using water, but are “tapping directly into cities’ drinking water supplies.” Here’s a quote from a Democracy Now interview:

From a freshwater perspective, these data centers need to be trained on freshwater. They cannot be trained on any other type of water, because it can corrode the equipment, it can lead to bacterial growth. And most of the time, it actually taps directly into a public drinking water supply, because that is the infrastructure that has been laid to deliver this clean freshwater to different businesses, to different homes. And Bloomberg recently had an analysis where they looked at the expansion of these data centers around the world, and two-thirds of them are being placed in water-scarce areas. So they’re being placed in communities that do not have access to freshwater. So, it’s not just the total amount of freshwater that we need to be concerned about, but actually the distribution of this infrastructure around the world.

A few things worth noting about this quote. First, the claim that data centers “cannot” use any water other than freshwater is wrong. Data centers can and increasingly do use reclaimed wastewater for cooling. AWS has announced that 20 of its data centers cool with purified wastewater, and Google uses reclaimed or non-potable water at over 25% of its data center campuses. It’s true that reclaimed water can cause more corrosion, scaling, and microbiological growth than potable water, which is why most data centers prefer potable water—but it is not technically impossible, and the industry is actively moving toward non-potable sources. Second, the Bloomberg analysis that Hao references did find that about two-thirds of new data centers built or in development since 2022 are in places with high water stress—but Bloomberg’s own reporting attributes this to data centers chasing cheap energy and favorable tax policies, not water availability, noting that “water is often the last consideration when making siting decisions.”

But the larger point here is this: it’s actually preferable for data centers to buy water from municipal water systems, because that means they’re paying into the system itself and providing utility revenue that on net benefits other buyers. One industry analysis estimated that around 57% of data centers’ direct water use comes from potable water supplies, with the rest from alternative sources. When data centers connect to public water systems rather than drilling private wells or drawing directly from rivers, they are customers of the utility, generating revenue that can fund infrastructure upgrades. It would be much worse if they weren’t using the municipal water system, because that would mean they would still be taking freshwater and competing with the municipal water system if water were limited, but not providing any revenue to the utility.

So I think this on its own makes my case extremely strong that we need to look exclusively at the freshwater withdrawals being used on AI, not the potable water. The fact that data centers predominantly use potable water either doesn’t matter (in freshwater-abundant areas) or helps my case (because it means data centers are paying into municipal systems). But let’s say this effect of economies of scale somehow doesn’t happen. What would this mean for AI?

How much do we actually pay to make water potable?

In the U.S., “potable” basically means “meets EPA’s National Primary Drinking Water Regulations”: a set of contaminant limits and required treatment techniques for things like microbes, nitrates, arsenic, and so on. EPA’s NPDWR table spells out the specific numerical limits. There’s nothing metaphysically special about this water. It’s normal freshwater that has gone through coagulation, filtration, disinfection, and sometimes a few extra polishing steps to hit those standards. EPA’s Surface Water Treatment Rules are written on the assumption that your source is an ordinary river, lake, or reservoir.

One way to see how cheap this conversion is is to look at what utilities already charge. EPA’s WaterSense program compiles national rate data and estimates that, in 2023, the average U.S. residential water price was about $6.64 per 1,000 gallons, with wastewater service adding another $8.57 per 1,000 gallons. Commercial customers paid about $5.56 for water and $6.86 for wastewater per 1,000 gallons. EPA summarizes these averages here.

That combined ~$15 per 1,000 gallons is not just the cost to flip freshwater into potable water. It also includes:

• Treatment

• Distribution pumping and storage

• Maintaining and replacing pipes

• Billing, customer service, and regulatory compliance

If you look at simpler wholesale or regional numbers, you get a cleaner sense of the underlying production cost. For example:

• Ottawa County, Ohio reports that its regional water plant charges about $2.77 per 1,000 gallons to its distribution customers to pump raw lake water, treat it, and deliver finished water to local systems. That figure explicitly includes operations, maintenance, debt service, and admin. The county posts the breakdown on its FAQ page.

• A Texas feasibility study for Brownsville’s conventional water treatment plant (separate from a proposed desal project) uses an internal treatment cost of about $1.34 per 1,000 gallons of potable water as the baseline. That appears in the Texas Water Development Board’s report tables.

These are “all-in” utility costs to take raw freshwater, treat it to drinking standards, and push it into the local system. They’re on the order of a couple of dollars per 1,000 gallons.

Nationally, EPA’s own example water bill assumes a retail volumetric charge of about $0.00295 per gallon, which is $2.95 per 1,000 gallons, for a typical surface-water system, and that again includes treatment plus distribution and overhead. EPA walks through that math in its “Understanding your water bill” example.

So as a very rough rule of thumb:

• Turning raw freshwater into potable water plus delivering it is usually in the $2–$7 per 1,000 gallons range, depending on the city.

• The incremental treatment-only cost (before pipes and billing) is typically well under a couple of dollars per 1,000 gallons, and often closer to $1.

How much potable water specifically will AI use? How much would it cost to make all that water potable?

The US public water supply uses ~40 billion gallons per day, all of this is potable. Data centers used 50 million gallons per day onsite in 2023. So their potable water usage was 0.13% of the public water supply. I want to beat the drum again that data centers more broadly are being used by all Americans for half their waking lives (because they support the internet), so the fact that as an industry they’re using a little bit of the total potable drinkable water in the country shouldn’t surprise us on its own. Assuming again that AI is ~20% of what happens in data centers (10 million gallons per day), AI in 2023 used ~0.03% of all drinkable water in America. If we assume an aggressive 10x growth of AI in data centers by 2030, and no changes to how it uses water, this will rise to 0.3%: 100 million gallons per day.

What about the water used offsite for power generation? This is often brought up as a “hidden” water cost of AI, and many estimates of AI’s total water use include it. This estimates often make AI’s total water use 5-10x bigger. However, effectively none of the water power plants use is potable. If you see an estimate of AI’s total water use, don’t immediately assume that all that water is drinkable. Check and see if the study is including the offsite water (they often do). If you’re worried about potable water specifically, only look at the water the data center itself is using.

Taking our estimate that making freshwater potable usually costs about $1 per thousand gallons, this implies that the US would be spending an additional $100,000 per day to generate enough potable water for AI data centers. If water markets and policies are working well, this cost should mainly be born by the AI companies themselves. This cost is about 0.01% of Google’s revenue alone. But let’s say that this doesn’t happen, and somehow the normal split between commercial and household rates breaks, and the cost is instead exclusively borne by American household consumers. How much would this increase water bills?

The average American family of four’s water bill is $2.60 per day. This implies American households spend in total $221,000,000 per day on water. The total increase in demand for potable water from data centers in the extreme case where they grow 10x by 2030, source all their water from municipal water systems, and the cost is somehow exclusively borne by households rather than AI companies and other commercial buyers, would together increase total American household spending on drinkable water by 0.04%.

Okay, well we know data centers aren’t evenly spread around the country. What if instead that 10x increase in water use happens exclusively in top 4 counties in the US with the most data centers are:

• Loudoun County Virginia - population 446,530

• Prince William County Virginia - population 484,625

• Maricopa County Arizona - population 4,726,247

• Licking County Ohio - population 180,311

The cost of turning enough additional water potable for data centers, if it were borne exclusively by the people living in these four counties, would increase their total spending on water by roughly 3% by 2030. This is the absolute most the cost of making freshwater potable could contribute people’s difficulty of accessing freshwater, under the insane assumptions that:

• The 10x growth in AI activity will happen exclusively in 4 of America’s 3000 counties.

• All costs of making more water potable are somehow only borne by households rather than companies.

The fact that AI requires potable water just isn’t an issue. The shocking headlines about AI “using up our precious drinking water” get the problem backwards. It’s good that we have new large buyers paying utilities to upgrade their water systems, because the main issue with water costs in America comes from the maintenance of old treatment and delivery systems that don’t have enough revenue to upgrade. People sometimes talk as if drinkable water is this incredibly rare naturally occurring thing that we stumble on, rather than something that we make. I blame Poland Spring ads.

A type of crisis that doesn’t happen

My impression after reading a lot about water issues in America is that the vast majority are caused by one of two things happening:

• An area running out of enough raw freshwater to treat and deliver to people.

• Old infrastructure not having enough buyers to afford upgrades. This is how you get towns that still have lead pipes: there isn’t enough funding to support the infrastructure project of replacing them.

The idea that it matters a lot that AI uses potable water and not just freshwater seems to imply that there’s a third type of crisis possible:

• A place with otherwise abundant but untreated water having too much commercial demand put on its water utility, causing prices to rise for residential users.

But as best I can tell, this seems to almost never happen, especially when compared to the first two. I can’t find many examples like this at all. This seems to hint that, again, what matters is total available freshwater, not potable water.

Do data centers poison water supplies?

It’s very common that I’m told that data centers poison the water around them. This idea seems to be basically entirely 100% fake, but very popular and caught on because of misreadings of two specific popular news stories.

There are three ways data centers can make water more polluted:

1. Sediment runoff from construction. This is a problem with any large building. A data center or a Wal Mart super center or an Amazon fulfillment center or a hospital being built near you pose equal risk of doing this.

2. Evaporating water contaminated with pollution, like nitrates. The contamination doesn’t get evaporated, and instead is filtered out of the data center. This increases the proportion of contaminants in the water without the data center actually adding any new contaminants itself. As I’ll show below, this isn’t really a problem even in regions where a lot of focus has gone to it.

3. Treatment chemicals data centers use to prevent corrosion, and biological growth in their cooling systems. These chemicals (typically biocides, corrosion inhibitors, and scale inhibitors) are discharged in what’s called “blowdown water” when cooling towers cycle out concentrated water. The water is either sent to either a municipal sewer system, where it goes to a wastewater treatment plant before reaching any natural waterway, or to a nearby river, stream, or lake directly. Either way, Data center blowdown is regulated under the Clean Water Act via NPDES discharge permits, and the chemicals involved are the same ones used in any industrial cooling system, including those at hospitals, universities, and office buildings.

There are no reported cases anywhere of data center blowdown causing any specific issues for water quality. There are places where “residents have concerns” about them, but there has not been any measurable change in water supply pollution due to data centers.

One of the most popular news stories about data centers was caused by #1, where construction from a data center in Georgia may have caused sediment to build up in a nearby groundwater supply and dry up the wells. That led to this New York Times story profiling a couple who was impacted:

And a lot of other scary headlines. This one from the BBC on the same couple has been especially popular:

Most people who saw these stories seem to only have read the headlines, and gotten the impression that the data center itself drew water or polluted the water during its normal operations, and that this caused the water problem. It’s not helped that both headlines contain wildly misleading terms. The NYT’s says “tech giants are building data centers that guzzle up water. This has led to problems for people who live nearby.” But the story itself is NOT about a data center guzzling up water, and the problems for the residents nearby was entirely caused by construction. Normal data center operations “guzzling up water” have so far not created significant issues for water access anywhere. They definitely have not meaningfully polluted water anywhere. But you wouldn’t know it from these headlines. The BBC headline would make much more sense if it said “I can’t drink the water - life next to a US large building.” So many people have seen these two headlines and now believe data centers suck water dry and poison and pollute the water around them. Neither is true, and I know that because the content of the articles themselves don’t make that claim at all! I unpack this more here.

There is only one story where #2, data centers increasing concentration of pollutants by evaporating already polluted water, has happened and gotten attention. There was a big story a few months ago about an Oregon data center “giving people cancer” by polluting the water. When you read the story itself, you discover that all the pollutants in the water giving people cancer are actually nitrates coming from farmers who aren’t being careful about spillover. What’s happening is that the data center is using a tiny amount of the region’s nitrate-poisoned water, evaporating it (which leaves the nitrates behind in the remaining water), and sending the used water to a waste treatment plant, where the concentration of nitrates is higher than it would have otherwise been because some of the water was evaporated. The nitrates in the wastewater plant seem to be leaking into nearby groundwater, causing an increase in cancer rates.

How much is the data center itself contributing to that increase? It seems like no one except me has taken a shot at actually estimating that. I do that below, my best guess is that the data center is contributing to significantly less than 1% of the total increase in nitrate concentration in the groundwater. It’s effectively like Amazon is running a few acres of the normal farmland in the area, and using the same amount of nitrates as all the other much larger farms around them. And yet news articles about this regularly featured scary images like Jeff Bezos superimposed in sickly colors in front of cancer cells:

Beyond the specific numbers here, it seems like Amazon was doing everything right. They were just buying water from a local utility and sending all wastewater back to a treatment plant. Why is it considered the job of private business to track the fact that the water they’re buying is actually polluted, so they’re returning more polluted water after? This is entirely the responsibility of the people in charge of managing the water, and the people breaking rules governing how much nitrates they can add to water.

But wait, you may have heard that Amazon agreed to pay $20 million to settle a lawsuit brought against them for the pollution here. There were a lot of stories about this, like this one:

Surely this means they’re guilty right? Nope. As usual this story is being wildly misrepresented.

In the settlement documents, Amazon states explicitly that it is entering the agreement "solely to avoid the burdens and expense of litigation," and denies "each and every one of the allegations of wrongful conduct.” They say their data centers "don't add nitrates to that water, and the water we return represents a very small fraction of the region's overall system." If you do the math, it seems impossible not to agree with Amazon's statement here. If you disagree, I’d be interested if you could show me how this was physically possible.

Amazon is one of 17 defendants in this case. The others include the large farms and food processing companies (the groups that actually added the nitrates in the first place), and the wastewater facility that's been failing to deal with leakage into the groundwater. They've been building up the nitrate concentration in the water since the early 90s, way before the data center was opened in 2011.

Amazon was the first to settle because its marginal contribution to the problem is the easiest to pay off and move on from! The agricultural defendants are still litigating because their part in this is way larger.

But as usual people are entirely misreading this and now think that 1) A data center poisoned people and gave them cancer, and 2) It's so clear this happened that Amazon won't even fight it in court. Ridiculous! This story is a great example of how actual water villains, in this case big ag and the poorly run wastewater management facility, can use data centers as convenient scapegoats. Freaking out about a fake problem often lets real villains off the hook. For a good overview of the actual causes of the problem in this case, I’d recommend this great local reporting on the failures of the wastewater facility.

And that’s it! Those are the only two places in the country where data centers seem to have polluted water at all, yet I very regularly get told by everyday people that data centers “poison water.” This is the result of two very popular but wildly misleading headlines.

Using AI can save way more water than is used in data centers

None of what I’m going to list here is generative AI like ChatGPT, but other machine learning tools have been heavily benefited by spillover effects from the general economic success of generative AI tools. These tools include things like computer vision systems for manufacturing quality control, predictive maintenance algorithms for industrial equipment, recommendation engines for e-commerce, fraud detection systems for financial services, and optimization algorithms for logistics and supply chains. While these applications have existed for years, the massive investment flowing into AI infrastructure (data centers, chip manufacturing, training talent) has made them cheaper, faster, and more accessible to deploy at scale.

AI is a way to get a machine to build its own soup of internal heuristics for how to handle complex situations. This soup of heuristics both succeeds and fails in surprising ways. We don’t design the heuristics themselves, and don’t even know what they are or how they work, in the same way we don’t know how a lot of the heuristics in our own brains work. There are a lot of situations where deploying this type of machine can help us optimize water use, because there are lots of places in America and around the world where huge amounts of water is wasted. Almost 20% of U.S. drinking water is lost to leaking pipes before reaching consumers.

• AI leak detection saves 3 billion gallons of water over a few years in New Jersey.

• AI detecting a single leak in a local suburb saved 350,000 gallons of water per day, saving a municipality $213,000 per year.

• Another leak detection company saves 25,000,000 gallons of water per year.

• A single AI-based irrigation optimization tool saved South American farmers 19 billion gallons of water over 2 years. This is 2 times as much water as all American AI data centers used over the same period.

This was the result of a few minutes of Googling. I could go on way longer with more examples of ways simple AI tools are saving towns huge amounts of water.

It’s okay to use water on a digital product

Many people’s background aversion to using water on data centers is that it’s a physical resource being spent on a digital product. Shouldn’t we only spend physical goods on other physical goods?

We already use a lot of water on the internet, and digital goods in general. Most of the ways we generate electricity uses a lot of water, so most of the time, when you’re using a computer, TV, or phone, you’re also using water. Internet data centers have always relied on water cooling, so we were always using water to access and share digital information.

Information is valuable. The real value of a book is almost entirely in the words it contains, not the physical quantity of ink and paper that make it up. We think it’s valuable to spend lots and lots of physical resources each year making books. Books and newspapers use 153 billion gallons of water annually. This is almost entirely in the service of delivering information. If it’s okay to spend water on creating and distributing books, it’s okay to spend water on other sources of valuable information. The water used to deliver digital information is orders of magnitude lower than physical books.

You might object that AI does not deliver anything like the value of books. My point isn’t to make a claim about how much valuable information AI provides, only that it isn’t inherently bad to spend a physical resource to deliver information. Ultimately if you believe AI is entirely valueless, than any water used on it is wasted regardless of whether AI’s output is physical or digital. But the fact that it’s digital on its own shouldn’t factor into whether you think it’s valuable or not.

The social value or harm of a tool isn’t the final word on how harmful it is to the environment

A very common point that comes up in conversations about AI and water use is that no matter how little water AI uses, AI is either useless or actively harmful, so all that water is being used on something bad. This makes it inherently worse than using any large amounts of water on good things. For example, when I share these stats from before:

Here’s a list of common objects you might own, and how many chatbot prompt’s worth of water they used to make (all from this list, and using the onsite + offsite water value):

• Leather Shoes - 4,000,000 prompts’ worth of water

• Smartphone - 6,400,000 prompts

• Jeans - 5,400,000 prompts

• T-shirt - 1,300,000 prompts

• A single piece of paper - 2550 prompts

• A 400 page book - 1,000,000 prompts

I often get the response that all of these things have social value, whereas AI has no value, so AI is worse for water than all these things, even though it’s using tiny tiny tiny amounts of water compared to each.

It seems like some people measure how wasteful something is with water is by a simple (Value to society / Water) equation where no matter how tiny something’s water use is, if it’s negative value, it’s always worse than something okay with huge water use.

This doesn’t make sense as a way of thinking about conserving water, for the same reason that it’s not a good way of thinking about saving money. If I were doing a fun activity that cost $40,000, and something useless or bad that cost $0.01, even though the $0.01 thing was bad, cutting it just would never ever be as promising or urgent as finding ways to reduce the cost of the $40,000 thing, or to just go without it.

Driving somewhere I want to be is much worse for the environment than riding a bike in the wrong direction. I agree that we need to factor in the value somehow, but it can’t just be “Anything socially bad is always worse for the environment than anything socially good.” AI water is often hundreds of thousands of times as small as many other ways we use water.

Talking about the social harm of a tool and adding “And it uses a few drops of water!” basically always dilutes the point you’re trying to make. I’m not exactly consistently pro AI. There’s a lot I’m worried about. But I find it distasteful when people effectively say “This far-right authoritarian government is using powerful AI systems to surveil people!… and also, every time they use it, a few drops of water are evaporated!” This just so obviously dilutes and trivializes the much more important point that I’d really rather it not be brought up. Manufacturing a gun uses at minimum 10,000 times as much water as an AI prompt in a data center, but if authoritarians are bearing down on people, I’m not going to add “And it cost a glass of water each to make their guns!'“

There’s a trade-off between water and energy for data center cooling systems. For the climate, water’s often preferable

Data centers don’t have to use water for cooling, they can also circulate cold air. They do this much more often when they’re built in deserts, because water’s more expensive and solar power’s cheaper and more abundant. As with any industry, they respond to the costs of goods and adjust how they use them accordingly.

But replacing water with air cooling systems means a lot more energy is used on cooling. Circulating cool air is more energy intensive than circulating water. Because water has much higher heat capacity and thermal conductivity than air, it can absorb and transfer heat more efficiently. Air cooling (especially in hot climates) requires stronger fans, chillers, compressors, and mechanical systems to push cooled air throughout the facility.

One study found that replacing air cooling with liquid cooling reduces a data center’s total power usage by 10%. This is a big deal, because electricity demand is a much more serious problem for data centers. Using 10% less energy also means roughly 10% less CO2 emissions. If water usage isn’t an issue, it seems like the main effect of water cooling is preventing a significant amount of CO2 emissions and electricity demand.

What about all those news stories about AI harming local water access?

Every popular article about how AI’s water use is bad for the environment in the last year has had a wildly misleading framing

The Washington Post: Every email written using ChatGPT uses a whole bottle of water

This is the single most influential article ever written about ChatGPT and the environment. I still to this day regularly bump into people who think that ChatGPT uses a whole bottle of water every time you prompt it.

Friend of the blog SE Gyges has written the best thorough explanation of why this article looks likely to be an intentional lie. I’d recommend the entire thing. The article concludes that the only way this number could possibly be real is if you make every one of the following assumptions:

For a worst-case estimate using the paper’s assumptions, if

• you query ChatGPT 10 times per email,

• you include water used to generate electricity,

• the datacenter hosting it is in the state of Washington,

• the datacenter uses the public power grid or something close to it,

• water evaporated from hydroelectric power reservoirs could otherwise have been used productively for something other than power generation,

• and LLMs were not more efficient when they were being sold for profit in 2024 than they were in 2020 when they had never been used by the public

then it is true that an LLM uses up 500 or more milliliters of water per email.

You can reach a similar estimate by different methods, since they break out the water use per state differently. For example, if the datacenter hosting ChatGPT is not in Washington, it will have a higher carbon footprint but a lower water footprint and you will have to query it 30 or 50 times to use up an entire bottle of water. This is not what anyone imagines when they hear “write a 100-word email”.

That study’s authors are well aware that none of these assumptions are realistic. Information about how efficient LLMs are when they are served to users is publicly available. People do not generally query an LLM fifty times to write a one hundred word email.

It is completely normal to publish, in an academic context, a worst-case estimate based on limited information or to pick assumptions which make it easy to form an estimate. In this setting your audience has all the detail necessary to determine if your worst-case guess seems accurate, and how to use it well.

Publishing a pessimistic estimate that makes this many incorrect assumptions in a newspaper of record with no further detail is just lying to readers.

The Economic Times: Texans are showering less because of AI

Take this one from the Economic Times, it circulated a lot:

• Texas AI centers guzzle 463 million gallons, now residents are asked to cut back on showers

The article clarifies that this is 463 million gallons of water spread over 2 years, or 640,000 gallons of water per day. Texas consumes 13 billion gallons of waters per day. So all data centers added 0.005% to Texas’s water demands.

0.005% of Texas’s population is 1,600. Imagine a headline that said “1,600 people moved to Texas. Now, residents are being asked to take shorter showers.”

Many iterations of the same article appeared:

• San Antonio data centers guzzled 463 million gallons of water as area faced drought

• Data Centers in Texas Used 463 Million Gallons of Water, Residents Told to Take Shorter Showers

One article corrected for the much larger uptick of data centers in 2025:

• Texas Data Centers Use 50 Billion Gallons of Water as State Faces Drought

50 billion gallons per year is a lot more! That’s more like 1.1% of Texas’s water use. Nowhere in this article does it share that proportion. It seems pretty normal for a state as large as Texas to have a 1% fluctuation in its water demand.

The New York Times: Data centers are guzzling up water and preventing home building

From the New York Times:

• Their Water Taps Ran Dry When Meta Built Next Door

The subtitle says: “In the race to develop artificial intelligence, tech giants are building data centers that guzzle up water. That has led to problems for people who live nearby.”

Reading that subtitle, you would have to assume that the main data center in the story is guzzling up the local water in the way other data centers use water.

In the article, residents describe how their wells dried up because residue from the construction of the data center added sediment to the local water system. The data center had not been turned on yet. Water was not being used to cool the chips. This was a construction problem that could have happened with any large building. It had nothing to do with the data center draining the water to cool its chips. The data center was not even built to draw groundwater at all, it relies on the local municipal water system.

The residents were clearly wronged by Meta here and deserve compensation. But this is not an example of a data center’s water demand harming a local population. While the article itself is relatively clear on this, the subtitle says otherwise!

The rest of the article is also full of statistics that seem somewhat misleading when you look at them closely.

Water troubles similar to Newton County’s are also playing out in other data center hot spots, including Texas, Arizona, Louisiana and the United Arab Emirates. Around Phoenix, some homebuilders have paused construction because of droughts exacerbated by data centers.

The term “exacerbated” is doing a lot of work here. If there is a drought happening, and a data center is using literally any water, then in some very technical sense that data center is “exacerbating” the drought. But in no single one of these cases did data centers seem to actually raise the local cost of water at all. We already saw in Phoenix that data centers were only using 0.12% of the county water. It would be odd if that was what caused home builders to pause.

The article goes on with predictions about Georgia’s water use around the data center. Residential water-sewer bills in Newton County have in fact risen — from about $73/month in 2015–2022 to $85/month by 2025, with the NCWSA’s own budget documents citing the second Meta campus as one factor accelerating capital improvements. However, the NCWSA also cites inflation and housing market uncertainty as drivers, and Meta is only using ~2% of the county’s water, less than the local pharmaceutical plant (~4%) and comparable to Rivian’s factory. So data center expansion is one contributing factor among several, not the primary cause — but the article’s concern wasn’t baseless.

I think the main story being an issue with construction, but the title associating it with some issue specific to data centers, seems pretty similar to a news story reporting on loud sounds from construction of a building that happens to be a bank, and the title saying “Many banks are known for their incredible noise pollution. Some residents found out the hard way.” This would leave you with an incorrect understanding of banks.

Contra the subtitle, data centers “guzzling up water” in the sense of “using the water for cooling” has not led to significant problems for the people who live nearby. Newton County is a partial exception. Data center expansion is one factor among several driving rate increases there, but even that case is far from the picture the subtitle paints of data centers causing water crises wherever they’re built.

CNET’s long very vague report on AI and water

This same story was later referenced by a long article on AI water use at CNET, here with a wildly misleading framing:

The developer, 1778 Rich Pike, is hoping to build a 34-building data center campus on 1,000 acres that spans Clifton and Covington townships, according to Ejk and local reports. That 1,000 acres includes two watersheds, the Lehigh River and the Roaring Brook, Ejk says, adding that the developer’s attorney has said each building would have its own well to supply the water neededEverybody in Clifton is on a well, so the concern was the drain of their water aquifers, because if there’s that kind of demand for 34 more wells, you’re going to drain everybody’s wells,” Ejk says. “And then what do they do?”

Ejk, a retired school principal and former Clifton Township supervisor, says her top concerns regarding the data center campus include environmental factors, impacts on water quality or water depletion in the area, and negative effects on the residents who live there.

Her fears are in line with what others who live near data centers have reported experiencing. According to a New York Times article in July, after construction kicked off on a Meta data center in Social Circle, Georgia, neighbors said wells began to dry up, disrupting their water source.

There’s no mention anywhere in the article that the data center in Georgia was not using the well water for normal operations.

Bloomberg: AI is draining water from areas that need it most

Here’s a popular Bloomberg story from May. It shows this graphic:

Red dots indicate data centers built in areas with higher or extremely high water stress. My first thought as someone who lives in Washington DC was “Sorry, what?”

Northern Virginia is a high water stress area?

I cannot find any information online about Northern Virginia being a high water stress area. It seems to be considered low to medium. Correct me if I’m wrong. Best I could do was this quote from the Financial Times:

Virginia has suffered several record breaking dry-spells in recent years, as well as a “high impact” drought in 2023, according to the U.S. National Integrated Drought Information System. Much of the state, including the northern area where the four counties are located, is suffering from abnormally dry conditions, according to the U.S. Drought Monitor. But following recent rain, the Virginia Department of Environmental Quality on Friday lifted drought advisories across much of the state, though drought warnings and watches are still in effect for some regions.

Back to the map. There were some numbers shared in a related article by one of the same authors. But readers were left without a sense of proportion of what percentage of our water all these data centers are using.

AI’s total consumptive water use is equal to the water consumption of the lifestyles of everyone in Paterson, New Jersey. This graphic is effectively spreading the water costs of the population of Paterson across the whole country, and drawing a lot of scary red dots. The dots are each where a relatively tiny, tiny amount of water is being used, and they’re only red where the regions are struggling with water. This could be done with anything that uses water at all and doesn’t give you any useful information about how much of a problem they are for the region’s water access.

Even the title chart can send the wrong message.

I think for a lot of people, stories about AI are their first time hearing about data centers. But the vast majority of data centers exist to support the internet in general, not AI.

Simply showing the number of data centers doesn’t show the impact of AI specifically, or how much power data centers are drawing. Power roughly correlates with water, because the more energy is used in data center computers, the more they need to be cooled, and the more water is needed to do that. Here’s a graph showing the power demand of all data centers, and how much of that demand AI makes up.

Obviously there’s been a big uptick on power draw since 2019, but AI is still a small fraction of total data center power draw. I think Goldman Sachs underestimated AI’s power draw here, experts think it’s more like ~15% of total power used in data centers, but it’s important to understand that the vast majority of that original scary red data center graph isn’t AI specifically.

AI is going to be large part of the very large data center buildout that’s currently underway, but it’s important to understand that up until this point most of those data centers on the graph were just the buildout of the internet.

One more note, circling back again to Maricopa County.

The county is a gigantic city built in the middle of a desert. For as long as it’s existed, it’s been under high water stress. Everyone living there is aware of this. The entire region is (I say this approvingly) a monument to man’s arrogance.

The only reason anyone can live in Phoenix in the first place is that we have done lots of ridiculous massive projects to move huge amounts of water to the area from elsewhere.

This is an area where environmentalism and equity come apart. I’d like residents of Phoenix to have access to reliable water supplies, but I don’t think this the most environmentalist move. I think the most environmentalist move would probably be to encourage people to leave the Phoenix area in the first place and live somewhere that doesn’t need to spend over two times as much energy as the country on average pumping water. I have to bite the bullet here and say that between environmentalism and equity, I’d rather choose equity and not raise people’s water prices much, even though they’ve chosen to live in the middle of a desert.

It seems inconsistent to think that it’s wrong for environmentalist reasons to build data centers near Phoenix that increase the city’s water use by 0.1%, but it’s not wrong for Phoenix to exist in the first place. If it’s bad for the environment to build data centers in the area at all, Phoenix’s low water bills themselves seem definitionally bad for the environment too. I think you can be on team “Keep Phoenix’s water bills low, and build data centers there” or team “Neither the data centers nor Phoenix should be built there, we need to raise residents’ water bills to reflect this fact” but those are the only options. I’m on team build the data centers and help out the residents of Phoenix.

More Perfect Union

More Perfect Union is one of the single largest sources of misleading ideas about data center water usage anywhere. They very regularly put out wildly misleading videos and headlines. There are so many that I’ve written a long separate post on them here. They are maybe the single most deceptive media organizations in the data center conversation relative to their reach.

Empire of AI

Karen Hao’s book Empire of AI was very popular and influenced the AI/environment conversation. It includes a chapter called “Plundering the Earth” that covers AI water and environmental issues. Unfortunately this section is one of the most misleading pieces of coverage of AI water issues that I’ve read. Within 20 pages, Hao manages to:

• Claim that a data center is using 1000x as much water as a city of 88,000 people, where it’s actually using about 0.22x as much water as the city, and only 3% of the municipal water system the city relies on. She’s off by a factor of 4500.

• Imply that AI data centers will consume 1.7 trillion gallons of drinkable water by 2027, while the study she’s pulling from says that only 3% of that will be drinkable water, and 90% will not be consumed, and instead returned to the source unaffected.

• Paint a picture of AI data centers harming water access in America, where they don’t seem to have caused any harm at all.

• Frame Uruguay as using an unacceptable amount of water on industry and farming, where it actually seems to use the same ratio as any other country.

• Frame the Uruguay proposed data center as using a huge portion of the region’s water where it would actually use ~0.3% of the municipal water system, without providing any clear numbers.

These are all the significant mentions of data centers using water in the book.

There’s too much to say on this one for this section, so I made it its own post. Hao very quickly commented on my post and was very gracious in correcting some of the big errors I spotted in more recent editions of the book:

An Oregon Data Center is giving people cancer

This story by Rolling Stone (repeated in Futurism and The Verge) implies that a data center in Oregon is increasing the rates of cancer and other health problems people in a small town develop because it’s increasing the concentration of nitrates in the water. But the story breaks down as soon as you look at the numbers.

This article itself makes it clear that the water pollution linked to cancer and miscarriages is coming entirely from nitrate pollution from farms. In general, data centers don’t pollute water.

It says the data centers are drawing water already polluted with nitrates, evaporating some of it, and returning it to a wastewater facility. Because nitrates don’t evaporate, the water returned has a higher concentration of nitrates. This is why they say the data centers aren’t “polluting” the water, they’re “worsening” a pollution problem. But even this is silly. The Port of Morrow (that takes in the contaminated water) handles 3.5 billion gallons of wastewater annually. It’s not clear how much the data center uses, the article just says “The data centers suck up tens of millions of gallons of water from the aquifer each year.” Let’s assume that’s the absolute most “tens of millions of gallons” can be: 99 million per year, or 3% of the port’s water.

The article notes:

The massive inputs of fertilizer to grow crops and feed for the animals came at a price: the contamination of the Lower Umatilla Basin. In 1992, DEQ measured an average nitrate concentration of 9.2 ppm across a cluster of wells pulling from the basin. By 2015, that average had risen 46 percent, to 15.3 ppm. For some wells, DEQ found nitrate levels nearly as high as 73 ppm, more than 10 times the state limit of 7 ppm.

How much of this is the data center responsible for? The article notes:

As the underground aquifer became tainted with more nitrates, even the ostensibly clean water that the Port pulled from the aquifer’s deepest wells — which it used to service its large industrial customers like Amazon — became polluted. Soon, Amazon was using water to cool its data warehouses with nitrates as high as 13 ppm — above the federal and state limits.

When that tainted water moves through the data centers to absorb heat from the server systems, some of the water is evaporated, but the nitrates remain, increasing the concentration. That means that when the polluted water has moved through the data centers and back into the wastewater system, it’s even more contaminated, sometimes averaging as high as 56 ppm, eight times Oregon’s safety limit.

An issue here is that I don’t know how much of the water the data center withdraws is evaporated. We know the data center takes in ~3% of the water used by the port, but not how much it actually delivers to the port after.

I’m going to make a few attempts at estimating how much this data center could be increasing the concentration of nitrates.

I’ll take the very most extreme case:

• The concentration of nitrates in the other wastewater is the same as the aquifer (~13 ppm). This seems unlikely, because it’s waste. It’s likely polluted by the farmers.

• The data center evaporates 100% of the water it uses. Again unlikely.

• The data center is using the very maximum amount of water implied by Rolling Stone’s wording: 99 million gallons per year.

In this case, the reservoir is dealing with 3.5 billion gallons of water from other sources, all at 13 ppm nitrate. 13 millionths x 3.5 billion gallons = 45500 units of nitrate in the water. The data center evaporates 99 million gallons of water at 13 ppm, leaving 1287 units of nitrate to add to the wastewater without adding any additional water. So the new concentration of nitrates in the data center is (45500+1287)/(3,500,000,000) = 13.37 ppm. So in the absolute most extreme case, the data center evaporating water is responsible for an additional 0.37 ppm of nitrate in the waste water. This is about 6% of the observed change in nitrate concentration. This seems equivalent to Amazon operating 6% of the farmland in the region, and this is the absolute worst harm the data center could be causing.

But the actual contribution to groundwater contamination is probably way less, for a few reasons:

• The wastewater mostly flows back to farms so the soils can capture the nitrates in the water. It’s mentioned here as a small fraction of the cause of the groundwater nitrate contamination, the others being septic tanks, CAFOs, explosives, and food processing waste.

• If the data center uses water at 13 ppm and evaporates it until it’s at a maximum of 56 ppm, then the absolute most it ever evaporates is 77 ppm.

• The wastewater the port deals with is likely to already have a significantly higher nitrate concentration because it’s dealing with agricultural runoff. I can’t find a good source on this, but it’s likely higher than the 13 ppm recorded in the aquifer.

• The water draw of the data center itself is probably significantly lower.

All of these together make me suspect that Amazon is almost definitely responsible for significantly less than 1% of the increase in nitrates. I’ll circle back when I find better numbers.

Beyond the specific numbers here, it seems like Amazon was doing everything right. They were just buying water from a local utility and sending all wastewater back to a treatment plant. Why is it considered the job of private business to track the fact that the water they’re buying is actually polluted, so they’re returning more polluted water after? This is entirely the responsibility of the people in charge of managing the water, and the people breaking rules governing how much nitrates they can add to water.

You may have seen more recently that Amazon agreed to pay $20 million to settle a lawsuit brought against them for the pollution here. There were a lot of stories about this, like this one:

Surely this means they’re guilty right? Nope. As usual this story is being wildly misrepresented.

In the settlement documents, Amazon states explicitly that it is entering the agreement “solely to avoid the burdens and expense of litigation,” and denies “each and every one of the allegations of wrongful conduct.” They say their data centers “don’t add nitrates to that water, and the water we return represents a very small fraction of the region’s overall system.” If you do the math, it seems impossible not to agree with Amazon’s statement here. If you disagree, I’d be interested if you could show me how this was physically possible.

Amazon is one of 17 defendants in this case. The others include the large farms and food processing companies (the groups that actually added the nitrates in the first place), and the wastewater facility that’s been failing to deal with leakage into the groundwater. They’ve been building up the nitrate concentration in the water since the early 90s, way before the data center was opened in 2011.

Amazon was the first to settle because its marginal contribution to the problem is the easiest to pay off and move on from! The agricultural defendants are still litigating because their part in this is way larger.

But as usual people are entirely misreading this and now think that 1) A data center poisoned people and gave them cancer, and 2) It’s so clear this happened that Amazon won’t even fight it in court. Ridiculous! This story is a great example of how actual water villains, in this case big ag and the poorly run wastewater management facility, can use data centers as convenient scapegoats. Freaking out about a fake problem often lets real villains off the hook. For a good overview of the actual causes of the problem in this case, I’d recommend this great local reporting on the failures of the wastewater facility.

5 common misleading ways of reporting AI water usage statistics

Comparing AI to households without clarifying how small a part of our individual water footprint our households are

Many articles choose to report AI’s water use this way:

“AI is now using as much as (large number) of homes.”

Take this quote from Newsweek:

In 2025, data centers across the state are projected to use 49 billion gallons of water, enough to supply millions of households, primarily for cooling massive banks of servers that power generative AI and cloud computing.

That sounds bad! The water to supply millions of homes sounds like a significant chunk of the total water used in America.

The vast majority (~93%) of our individual total consumption of freshwater resources does not happen in our homes, it happens mainly in the production of the food we eat, and industry and commercial buildings. Experts seem to disagree on exactly what percentage of our freshwater consumption happens in our homes, but it’s pretty small. Most estimates seem to land around 1%. So if you just look at the tiny tiny part of our water footprint that we use in our homes, data centers use a lot of those tiny amounts. But if you look at the average American’s total consumptive water footprint of ~1600 L/day, 49 billion gallons per year is about 300,000 people’s worth of water. That’s about 1% of the population of Texas. The entire data center industry (both for AI and the internet) using as much water as 1% of its population just doesn’t seem as shocking.

Referencing the “hidden, true water costs” that AI companies are not telling you, without sharing what those very easily accessible costs are

A move that I complained about in my last post is that a lot of articles will imply that AI companies are hiding the “true, real” water costs of data centers by only reporting the “onsite” water use (the water used by the data center) and not the “offsite” water use (the water used in nearby power plants to generate the electricity). Reporting both onsite and offsite water costs has become standard in reporting AI’s total water impact.

Many authors leave their readers hanging about what these “true costs” are. They’ll report a minuscule amount of water used in a data center, and it’s obvious to the reader that it’s too small to care about, but then the author will add “but the true cost is much higher” and leaves the reader hanging, to infer that the true cost might matter.

We actually have a pretty simple way of estimating what the additional water cost of offsite generation is. Data centers on average use 0.48 L of water to cool their systems for every kWh of energy they use, and the power plants that provide data centers energy average 4.52 L/kWh. So to get a rough estimate:

• If you know the onsite water used in the data center, multiply it by 10.4 to get the onsite + offsite water.

• If you know the onsite energy used, multiply it by 5.00 L/kWh to get the onsite + offsite water used.

Obviously scaling up a number by a factor of 10 is a lot, but it often still isn’t very much in absolute terms. Going from 5 drops for a prompt to 50 drops of water is a lot relatively, but in absolute terms it’s a change from 0.00004% of your daily water footprint to 0.0004%. Journalists should make these magnitudes clear instead of leaving their readers hanging.

This talking point can be doubly deceptive if you only look at the proportion

Let’s say there are 2 data centers in a town (I’ll call them Poseidon and Enki) drawing from the same power source. The local town’s electricity costs 4 L of water per kWh to generate.

The Poseidon data center is pretty wasteful with its cooling water. It spends 2 L of water on cooling for every kWh it uses on computing, way above the national average of 0.48L/kWh. So if you add the onsite and offsite water usage, Poseidon uses 6 L of water per kWh.

The Enki data center finds a trick to be way more efficient with its cooling water. It drops its water use down to 0.1L/kWh. Well below the national average. So if you add its onsite and offsite water usage, it uses 4.1 L per kWh without using any more energy.

Obviously, the Enki data center is much better for the local water supply.

Both data centers are asked by the town to release a report on how much water they’re using. They both choose to only report on the water they’re actually using in the data center itself.

Suddenly, a local newspaper shares an expose: both data centers are secretly using more water than they reported, but Enki’s secret, real water use is 41x its reported water costs.

While Poseidon’s is only 3x its reported water costs:

Here, Enki looks much more dishonest than Poseidon. If readers only saw this proportion, they would probably be left thinking that Enki is much worse for the local water supply. But this is wrong! Enki’s much better. The reason the proportions are so different is that Enki’s managed to make its use of water so efficient compared to the nearby power plant.

I think something like this often happens with data center water reporting.

When I wrote about a case of Google’s “secret, real water cost” actually not being very much water, a lot of people messaged me to say Google still looks really dishonest here, because the secret cost is 10x its stated water costs once you add the offsite costs. A way of reframing this is to say that Google’s made its AI models so energy efficient that they’re now only using 1/10th as much water in their data centers per kWh as the water required to generate that energy. This seems good! We should frame this as Google solidly optimizing its water use.

Take this quote from a recent article titled “Tech companies rarely reveal exactly how much water their data centers use, research shows”:

Sustainability reports offer a valuable glimpse into data center water use. But because the reports are voluntary, different companies report different statistics in ways that make them hard to combine or compare. Importantly, these disclosures do not consistently include the indirect water consumption from their electricity use, which the Lawrence Berkeley Lab estimated was 12 times greater than the direct use for cooling in 2023. Our estimates highlighting specific water consumption reports are all related to cooling.

The article should have mentioned that this means data centers have made their water use so efficient that basically the only water they’re using at all is in the nearby power plant, not in the data centers themselves. But framing it in the original way way make it look like the AI labs are hiding a massive secret cost from local communities, which I guess is a more exciting story.

Vague gestures at data centers “straining local water systems” or “exacerbating drought” without clarifying what the actual harms are

If you use literally any water in any area with a drought, you’re in some sense “straining the local water system” and “exacerbating the drought.” Both of these tell us basically nothing meaningful about how bad a data center is for a local water system. If an article doesn’t come with any clarification at all about what the actual expected harms are, I would be extremely wary of this language. In basically every example I can find where it’s used, the data centers are adding minuscule amounts of water demand to the point that they’re probably not changing the behavior of any individuals or businesses in the area.

Simply listing very large numbers without any comparisons to similar industries and processes

This is the great singular sin of bad climate communication. The second you see it, you should assume it’s misleading. Simply reporting “millions of gallons of water” without context gives you no information. The power our digital clocks draw use millions of gallons of water, but digital clocks aren’t causing a water crisis.

Take this example:

7.2 million gallons per year! Sounds like a ton. How much is that? This data center would represent about 0.02% of nearby El Paso’s water usage. Probably not nearly as much as this tweet is trying to get across.

Whenever you see an article cite a huge amount of water with no comparisons at all to anything to give you a proper sense of proportion, ask a chatbot to contextualize the number for you.

Take this excerpt from “Are data centers depleting the Southwest’s water and energy resources?”

Meta’s data centers, meanwhile, withdrew 1.3 billion gallons of water in 2021, 367 million of which were from areas with high or extremely high water stress. Total global water consumption from Meta’s data centers was over 635 million gallons, equivalent to about 6,697 U.S. households. It’s not clear how much of this water withdrawal occurs in the United States, although that’s where most of Meta’s data centers are located. Neither report reveals the specific water use of the company’s Arizona data center.

I’m going to rewrite this, but using my town of 16,000 people (Webster, Massachusetts) as a unit to measure Meta’s data center withdrawal instead of individual households. Webster’s utility delivers 1.4 million gallons of water per day to the citizens and businesses there (511 million gallons per year).

Meta’s data centers, meanwhile, withdrew as much water as 3 Massachusetts small towns in 2021. Two thirds of a single one of those small towns was in areas with high or extremely high water stress. Total global water consumption from Meta’s data centers was a little more than a single Massachusetts small town. It’s not clear how much of this water withdrawal occurs in the United States, although that’s where most of Meta’s data centers are located. Neither report reveals the specific water use of the company’s Arizona data center.

This all seems silly when you consider Meta’s one of the single largest internet companies.

Reporting the maximum upper bound for water a data center uses as the number it will actually use

Many articles about current or future AI data centers report the number in the water permit they apply for as the amount of water they actually use day to day. But this almost never happens.

When a data center is being built, the company needs to obtain water use permits from local authorities before construction. At this stage, they have to estimate their maximum possible water consumption under worst-case scenarios:

• All cooling systems running at full capacity

• Peak summer temperatures

• Maximum IT load (every server rack filled and running)

• Minimal efficiency from cooling systems

The permit needs to cover this theoretical maximum because regulators want to ensure the local water infrastructure can handle the demand and that there’s enough water supply for everyone. It’s easier to get a higher permit upfront than to come back later and request more, so data centers are incentivized to aim high.

Actual water usage is always significantly lower than what the permits allow, because they’re designed with the absolute worst conditions in mind. But many popular articles about how much water data centers use give the number on the water permit, not how much the data center actually uses.

Here’s one of many examples:

Duff is the only city council member to vote no on a recently approved $800 million data center - rumored to be for Facebook - after discovering the facility would eventually use 1.75 million gallons of water every day for cooling their rows of servers once fully operational.

Some examples of great news coverage

After complaining so much, I’d like to share some examples of where the AI water issue has been covered especially well.

First is I think the best single popular summary that’s been written of the issue: Thirsty for power and water, AI-crunching data centers sprout across the West. This piece doesn’t break any of my rules that I share above and gives readers a great overview of where data centers fit into the broader water picture in the places they’re built.

Second is Hank Green’s recent coverage. A few places I disagree, but again fantastically comprehensive and explicitly clarifies some of the places where this is very confusing for readers.

These are both stand outs. I’ll continue to collect really good reporting here as I catch it.

Further reading

• The Lawrence Berkeley National Laboratory’s report on data center energy and water use in 2024 is the most comprehensive document we have on AI and water right now.

• Brian Potter’s recent piece on water and update on data center water use.

• Hannah Ritchie has some recent great stuff on data centers and chatbots

• Matt Yglesias’s piece on data centers and water

• Friend of the blog SE Gyges has a great breakdown of the single most popular statistic about ChatGPT that’s also a lie: it uses a bottle of water per email.

• More from me:

  • What a data center is

  • More Perfect Union videos are wildly deceptive on data center water use

  • Contra the UK government, please don’t delete your old photos and emails to save water

  • Using ChatGPT is not bad for the environment - a cheat sheet

  • Computing is efficient

This jump correlates with the massive new data center buildout, which started around 2021. This has led many to conclude that data centers are the main culprit. If data centers raised electricity bills by 35%, that seems like a huge disaster!

But that’s wrong. If you adjust for inflation, American household electricity costs are still historically low, and have only rebounded to where they were in 2017:

Sometime around 2022, American electricity prices started to outpace inflation. Before 2022, electricity prices closely tracked inflation. After 2022, they went ahead of it a bit.

Here’s a broader picture of inflation-adjusted household electricity rates over time. The recent spike begins in 2022, it’s this spike the red arrow’s pointing at that needs an explanation.

Did data centers cause that jump? No. A Lawrence Berkeley National Laboratory study found that state-level load growth from 2019 to 2024 — including growth driven by data centers, manufacturing, and electrification — was associated with lower average retail electricity prices, not higher ones. States with the highest load growth saw inflation-adjusted price reductions, while states with shrinking loads generally saw prices rise. The researchers attribute this largely to fixed infrastructure costs being spread across more kilowatt-hours of demand.

Two important caveats. First, the LBNL researchers note that this negative relationship between load growth and prices “was smaller and lost statistical significance when analyzing residential prices alone,” the clearest effect shows up in overall average retail prices, not necessarily household bills specifically. Second, the researchers caution that this historical relationship may not hold going forward, noting that “in some cases, spikes in load growth can result in significant, near-term retail price increases.” The PJM capacity price spikes beginning in 2025, driven substantially by projected data center growth, are an example of exactly this kind of near-term price impact.

To summarize their explanation of the price increase:

• Prices rose mainly because utility costs rose. On a nominal basis, the average retail price increased about four point eight percent per year from 2019 to 2023. After inflation, the national average was mostly flat over that window, though residential prices climbed faster than inflation. 

• Fuel and purchased power costs were the single biggest mover. Utilities’ fuel and purchased power costs surged in step with the natural gas spike, accounting for about half of the total expense increase. Gas prices then eased, but fuel and purchased power costs in 2023 were still about thirty percent above 2019. 

• The rest of the cost growth came from operations and maintenance, depreciation, taxes, and other line items.

This explains why the jump happens when it did: Russia invaded Ukraine, Europe stopped buying Russian gas, America was suddenly competing with European countries more for natural gas from other places, and natural gas prices spiked for a bit. Here’s a graph of natural gas prices:

That red arrow I added to the graph of electricity prices is on the day when Russia invaded Ukraine.

There are a few additional reasons to think data centers didn’t cause this spike.

First, the rise in data center electricity demand in the US started a few years before the spike. It’s a pretty straight line up, and markets have a lot of heads-up when a new large buyer like a data center is coming online. It would be weird if halfway along this basically straight line in 2022, energy markets were suddenly taken by surprise:

Second, American electricity demand is rising for the first time since 2007 or so (with data centers responsible for about 40% of the growth), but it’s still rising at a slower rate than it had for most of the 20th Century. Total electricity consumption is driven by demand, so a consumption graph is a good proxy for a demand graph. Importantly, during the period from 1985 to 2005, we sustained a higher rate of increasing electricity demand, but inflation-adjusted household electricity prices fell the entire time.

So it’s possible in principle anyway for electricity supply to keep up with demand and keep prices low, for the same reason a grocery store doesn’t double its prices when it opens a second location. Notice that there was a period from 2020 to 2022 when electricity demand was rising but prices were flat. Another hint that this is more of a problem with supply.

This is the first conclusion of my deep dive: Data centers have (so far) not affected national average household electricity bills. Many of the graphs that get shared about this are showing the effects of inflation, Russia’s invasion of Ukraine, and a series of other supply-side issues with normal electricity costs that aren’t the result of responding to rising demand.

This is not to say that data centers won’t affect national prices in the future. Just that as of 2025, the average American’s electricity bill has not increased as a result of data center demand.

Brian Potter concluded his excellent deep dive on US electricity prices by noting:

At a high level, trends in wholesale electricity prices seem to mostly match trends we see in consumer prices. Until 2020, wholesale prices were flat or declining; since 2020, they’ve risen substantially, faster than consumer electricity prices. And in most places, transmission capacity appears to be an increasing bottleneck: we’re having a harder and harder time accessing the least expensive power. Probing the specific relationship between wholesale electricity and consumer electricity prices would be complex, and beyond the scope of this essay, but it seems likely to me that rising wholesale prices are a major factor.

The main question left to answer is how data centers affect electric state and local bills electric rates, not national rates.

Unfortunately, locals who live near data centers can’t tell what portions of their electric bill increases were caused by inflation, Russia, or supply-side issues, and what was caused by data centers. The data center buildout started in 2021, at the exact same time American inflation began to skyrocket.

Between 2020 and 2025, the average Virginia resident’s electricity rate went up by 28.1%. Virginia has both the highest percentage of electricity used by data centers, and some of the fastest data center growth in the country.

If you were a resident of Virginia, you might look around, notice that data centers in your state are using 33% more electricity than they were 5 years ago, and that your electric bill went up by 28.1%. You might conclude that there was a connection, and even go on a More Perfect Union video to talk about how data centers raised your electricity prices.

But you might not know that most of that 28.1% growth was due to inflation, or more importantly that Virginia’s electricity prices actually increased less than the national average:

Most American states don’t have anywhere near as many new data centers coming online as Virginia. Maine barely has any, but its rates shot up faster than any other state. If data centers were having a really massive impact on electrical bills, wouldn’t you expect your bill to grow faster than states without data centers?

Maybe Virginia’s a fluke, and the other states with huge data center builds have seen skyrocketing electric rates. Michael Giberson at RSA looked into this. The national average electricity price increase for all sectors was $4.27. He looked at each state with the most electricity going to data centers and added numbers to the graph to show how state prices grew more (green) or less (red) than the national average of +4.27 cents per kWh.

11 out of 15 of the states with the most data center buildouts have seen lower than average increases in electricity rates. Further, the red numbers are all way larger here than the green numbers (except New Jersey). Even the states that grew more than average were basically right at the average line. Weird! Here are these numbers on a graph:

So far, this does not look like data centers are uniquely disastrous for electricity bills.

But things get murkier. 30-60% of a data center’s operating costs are its electricity bills. This means data center companies mostly opt for states where electricity costs are low and there’s a lot of pre-existing grid capacity. These cheap, abundant grids can handle massive new loads, but this does not mean prices don’t rise at all as a result of data centers, they just may rise less than if data centers had built elsewhere. Companies have an incentive to build where they will raise electricity costs the least, that’s where they’ll also be paying the least for electricity! So this leaves us with an unfortunate paradox where data centers may be raising local prices, but they’re often in the places we can detect that the least. Also, there might be very local effects that don’t show up at the state leve.

We know shockingly little about exactly how much data centers are affecting household electricity prices at the state and local level. We do know they have made household electricity more expensive than it would have otherwise been in specific locations, because they have been cited by authorities in Virginia, Arizona, Delaware, Oregon, and other places as one of several causes of rising costs.

Complicating all this still is that there is no clear longterm correlation between how much total electricity is used in your area and your electric bills, for the same reason larger grocery stores don’t necessarily charge higher prices. Once supply catches up to demand, prices balance. If higher local electricity demand led to permanent higher prices, we would expect urban electricity prices to be higher than rural ones. In reality, urban and rural rates are basically the same.

Something odd in the debate is that some people are implying that data centers are making electric prices permanently higher by drawing more power. This is obviously not how electricity markets work. The thing that spikes your electricity bill is demand outpacing supply in the short term, not just a total increase in demand over the long term.

But also, data centers are forecast to just keep being built, so the grid might be in a permanent state of catch-up for a while, keeping prices higher before they have a chance to adjust.

And all this is in the background of America needing to do a huge buildout of clean energy to transition away from fossil fuels, and all that infrastructure would cost consumers more anyway. Also, whose responsibility is it to keep rates low? The private consumers of electricity, or the utilities? Does it make sense to both say that the utility’s responsible for keeping rates low but also there’s just not enough capacity to do that in the short term so regardless data centers will raise costs? Maybe we could redistribute all that tax money data centers provide to pay for consumer electric bills, or… or what about…

Every time I try to address one specific part of this topic, I come away thinking people can only really understand it if they understand five others as well. The complexity builds. I think you can’t really understand the issue of data centers and household electricity costs without also knowing things like:

• Data centers bring in more tax revenue per watt-hour than any other industry,

• The differences in tax cuts local governments provide them to entice them to build there

• How this all really depends on whether the local grid can build out enough new clean electricity infrastructure.

• How to think about who’s responsible for rising prices between consumers or producers.

• How data center builds might be good for the green energy transition because they’re forcing these huge new investments in the grid we need to make anyway

• How electrifying everything in the economy is a massive project necessary for green energy, so electricity demand is going to rapidly shoot up with or without data centers.

• How AI usage is expected to accelerate or harm the green energy transition

and so on.

I decided that I did need some kind of basic understanding of all this stuff, so I set out to gather what I think is the important info here. I’m just a hobbyist and far from an expert on any of this. This series is more a summary of what I’ve learned after doing a bunch of deep dives on these topics rather than “This is the definitive final story of data centers and electricity.” I do think I’ve climbed to reach some vistas in this mess of a problem where I can see things more clearly, so this series will aim to get you there too.

A lot of tiny efficient computers

Suppose I want to run my own tiny AI model. Not one a lab made, just my own model on a personal device in my home. I go out and buy a second very small computer to run it.

I use it a lot. Each day, I ask 100 questions to my mini AI model. Each prompt uses about ten times as much energy as a Google search, but a Google search is so tiny that the prompt also uses a tiny amount of energy. All together, my 100 prompts use the same energy as running a microwave for 4 minutes, or playing a video game for about 10 minutes.

Sending 100 prompts to this AI model every single day adds 1/1000th to my daily emissions.

Is what I’m doing wrong?

I think the average person would say no. This is such a tiny addition that I personally should be able to decide whether it’s worthwhile for me. We don’t go around policing whether people have used microwaves a few seconds too long, or played a video game for a few minutes too long. Why try to decide whether it’s evil for me to spend a tiny fraction of my daily energy on a computer program I personally think is valuable?

Let’s say that I learn that my neighbor is also using an AI model, and has bought her own second tiny computer for it too. It also only adds 1/1000th to her emissions as well. She also uses it 100 times per day, indicating that she’s getting some value out of it.

Is this wrong?

If it’s not wrong for me to use my personal tiny computer, it shouldn’t be wrong for her either. That would be treating me as somehow special or different from her. If it’s not bad for me to raise my emissions by 1/1000th doing something that I personally think is valuable, it’s not wrong for her either.

There are now 4 people in my neighborhood using a tiny computer for AI. It’s not wrong for each of them to do this. It’s like if they each started playing a new video game for a few minutes a day, or using their microwave a little bit more. The AI app is barely adding anything as a percentage of the neighborhood’s emissions. Just like it’s 1/1000th of my personal emissions, it’s also 1/1000th of the combined emissions of the 4 people who use it.

The 4 people in my neighborhood realize that they can combine their AI tools together, into a single more efficient computer. They wire them up into a centralized location between their homes. This saves some electricity and cooling costs:

Would this be wrong?

No. It’s actually using less electricity than it otherwise would, so this causes the emissions to drop. The climate only responds to total emissions, it doesn’t respond to where specifically energy is used.

What if more people wanted to do this? Lots of people have their own little mini AI computers. They want to add them to this growing centralized pile of mini AI computers. Each one only adds 1/1000th to their daily emissions if they send 100 prompts per day.

Once we got to 100,000 people, we would have a really gigantic centralized pile of tiny, hyper-efficient computers serving a population of a whole city. Each individual computer would still be efficient, each individual person would still only be adding a tiny tiny amount to their emissions, and the only new thing is that all these tiny computers would be in a single central location. They would probably be made even more energy-efficient due to economies of scale, and new more efficient systems for cooling and power delivery that could be used with all of them in one place.

If 100,000 people around the world were running these tiny computers from their personal homes, no one would take any issue with it, for the same reason no one has a problem with people playing gaming consoles for a few extra minutes a day. But because they are all suddenly in one place, they look large compared to other buildings, and so some people get confused and assume that because the new building is so large and energy-hungry, it must also be inefficient or wasteful. This is a huge mistake, and one that’s made all the time in conversations about AI and the environment.

All these tiny hyper-optimized computers in a single central location, serving hundreds of thousands of people at once, is what a data center is. Data centers are concentrations of hyper-efficient computer processes that no one would have any issue with at all if they were in the homes of the individual people using them. If you wouldn’t have a problem with these tiny computers, you shouldn’t have a problem with data centers either. The only difference is the physical location of the computers themselves, and the fact that these tiny computers are actually combined into larger ones to serve multiple people at once. The only reason they stand out is that these processes are concentrated in one building, which makes them look large if you don’t consider how many people are using them. If instead you see data centers as what they really are, building-sized hyper-efficient computers that hundreds of thousands of people are using at any given moment, they stop looking bad for the environment. In fact, they are the most energy-efficient way to do large scale computing, and computing is already very energy efficient. The larger the data center, the more energy-efficient it is. “Power usage effectiveness” is the ratio of how much energy the data center delivers to computers vs. how much it takes in. A perfectly energy-efficient data center would have a PUE of 1. The more energy a data center uses, the closer its PUE approaches zero energy waste:

Scale isn’t wasteful

There are two coffee shops, Skeet’s and Novabucks.

Skeet’s is 5 times as energy efficient as Novabucks. If you buy a coffee from them, the emissions you cause will be just 1/5th what they would be at Novabucks.

Skeet’s is much more popular than Novabucks. They have 1000 coffee shops. Novabucks just has 5.

Which one is worse for the environment?

If you want a coffee, and want it to be environmentally friendly, which one should you buy from? The answer seems obvious here. Skeet’s only emits a 5th as much per coffee, so you should buy from them.

But Skeet’s has so many more coffee shops that even though they’re individually more efficient, their total emissions are 40 times as high as Novabucks. They emit a 5th as much, but they have 200 times as many stores. (1/5) x (200) = 40. Someone could show you this graph, and say “You should really stop buying from Skeet’s, look at how much more CO2 they emit! Buy from Novabucks instead.”

This is obviously wrong. That giant Skeet’s line is actually made up of a ton of super eco-friendly coffees. The smaller Novabucks line is made up of way less efficient coffees. The difference is that Skeet’s is providing way more individual efficient coffees than Novabucks is providing inefficient ones.

So Skeet’s is just emitting more because it’s more popular, not that each coffee is wasteful. Anyone who wants a coffee and wants to reduce their emissions should always go to Skeet’s. Skeet’s is better for the environment than Novabucks.

Whether you buy a coffee at Skeet’s or Novabucks is basically exactly like whether you run a computer program in a data center or your home device.

The data center is highly energy optimized for dealing with huge numbers of computer programs from lots of people at once, much more optimized than any home computer. The reason it’s using so much more energy is that hundreds of thousands of people are using it at once, not that it’s wasteful with the energy it uses. For the same reason that Skeet’s isn’t bad for the environment, data centers aren’t either. The scale of something like a data center or a coffee business tells you little about how bad it is for the environment. What matters is how much energy it uses per task and per user interacting with it. Data centers are maybe the single most energy efficient buildings we have created in terms of how much energy they use relative to how many people are interacting with them at once. Data centers are like huge public computers tens of thousands of people can play with at once. Each individual person’s energy usage in a data center is way less than it would be if they ran the same task on their personal computers, but the data center seems bad only because it concentrates these tasks in a single physical location.

This is why I think it’s such a terrible idea for people interested in the environment to attack data centers specifically. Each data center is a miracle of efficiency and uses way less energy per user than any other physical structure we have ever been able to build. Concentrating such huge demands in a specific location can create problems at a local level, but the issue is only that the surrounding grid wasn’t built for these huge demands, not that the data centers themselves are somehow wasteful with the energy they use.

In general, if people are doing tiny individual tasks that don’t add much to their emissions, I think they should be left alone and not condemned for those tiny tasks, regardless of whether we think the tasks themselves are valuable. Data centers just concentrate these individually tiny tasks into a single location. That’s all they do.

Even data centers used to train AI models look tiny if you divide the cost of training by the number of people who will use the model and the number of prompts they will send. It’s like making an up front investment of about 50 Watt-hours per person who will eventually interact with the model. It’s an investment per person who will use the model of the energy it takes to run a laptop for an hour. Training an AI model is kind of like if someone came over to your house and designed a new app on your computer you could use, and it took them about an hour to finish, and that just happening for all the billion people who use chatbots. These buildings are miraculously energy efficient, it’s just that they’re also being used by hundreds of millions of people. We should see that as an environmental achievement.

To use a more blunt analogy, the reason I’ve been driven a little crazy by the recent debate about data centers is that it feels as if everyone has suddenly started condemning bicycles as being bad for the environment because in total, manufacturing and disposing of bicycles emits more CO2 than Hummers each year. This obviously ignores that bicycles are extremely low energy per mile ridden compared to Hummers. Once you understand just how wildly energy (and water) efficient data centers are per user, it is much harder to condemn them purely for being large.

Does the data bear this out?

What would we expect to see if data centers are so hyper-efficient? They should be a tiny, tiny part of our emissions relative to how much we use them, but concentrate that efficient energy usage in specific places and strain local grids. I think of this as the environmental paradox of data centers: they both put uniquely large concentrated demand on local grids, but are also tiny efficient parts of the global energy grid. This creates confusion in discussing their environmental impacts.

Globally, the average person interacts with the internet for 7 hours per day. The entire time they’re interacting with the internet, they’re effectively using a data center as they would a personal computer. Data centers host and organize and store the internet. When you use the internet, you’re using a data center.

All data centers worldwide emitted roughly 180 million tonnes of CO₂ in 2024, about 0.5% of the world’s annual emissions.

At first, this may seem high. Half a percentage of global emissions is still higher than some countries, because there are ~200 countries. Even if each country were the same size, each would be half a percentage of the world’s emissions. People sometimes point out that data centers globally are emitting more CO2 than countries like the Philippines.

But if everyone on Earth were spending half their waking lives every single day interacting with the Philippines, it makes sense that the country’s emissions would at least double. It would probably skyrocket way higher than that! It should be physically impossible to build a ton of buildings that everyone on Earth uses for half the day every day that would somehow only emit as much as the Philippines, that would only be 0.5% of all emissions, and yet we’ve done it! That’s a miracle!

America is currently going through a buildout of new AI data centers. 51% of all hyperscale AI facilities in the world are in America, and 60% of new ones will be built here. These data centers are supporting users around the world, not just in America. We’re making building-sized computers that people around the world can log into and use to access AI tools. About a billion people use AI. The last year we have great data on America’s AI energy usage is 2023, when it used about 40 TWh. If we guess AI energy increased by 20% in 2024, and that 500 million people were interacting with AI in 2024 (the actual number is probably much higher, because AI is built into a lot of what we do online now), then for each person using an American AI data center, the data center used about 96 kWh of energy in 2024 per user, 0.5% of the average global citizen’s annual energy usage, and just 0.2% of the average American’s. This is a rough estimate, but if you actually include a rough guess of how many people globally are interacting with AI and how many of those people are interacting with US data centers, AI data centers in America still appear remarkably energy efficient.

What about water? Well the water that was actually used onsite in American data centers was only 50 million gallons per day in 2023, the rest was used to generate electricity offsite. Only 0.04% of America’s freshwater in 2023 was consumed inside data centers themselves. Data centers consumed just 3% of the water used by the American golf industry in the same year.

Forecasts imply that American data center electricity usage could triple by 2030. Because water use is approximately proportionate to electricity usage, data centers may consume 150 million gallons of water per day onsite, 0.12% of America’s current consumptive freshwater use.

So the water all American data centers will consume onsite in 2030 is equivalent to:

• 8% of the water currently consumed by the US golf industry.

• 8% of the water consumed by US steel production.

• The water usage of 260 square miles of irrigated corn farms.

The total projected onsite water consumption of all American data centers is incredibly small by the standards of other ways water is used. Again, consider that these small amounts of water are being used on something we all spend half our lives using.

The real environmental problem with data centers is that they concentrate so much specific electricity demand into such small specific places that the surrounding grid often cannot keep up, and needs to rely more on fossil fuel plants, or onsite gas turbines. This is serious, but to think seriously about this problem, people need to understand that this high demand is not coming from any wasted energy in the data center itself, it’s that there are hundreds of thousands of people invisible to us who are interacting with any one of them at a given time. Concentrating this demand actually makes them more energy efficient, due to economies of scale. The larger the data center, the more efficient it tends to be. So all else being equal, if you want a computer application to use the least energy possible, you should prefer that it be in the largest data center possible, even though this data center will put more strain on the surrounding grid.

On top of all this, anything that uses electricity is going to be much easier to decarbonize than most parts of the economy. The total electricity sector is only responsible for 25% of global emissions. The much bigger challenge is going to be vehicles and industrial processes that use fossil fuels directly. If we transition to a clean grid, the vast majority of the environmental problems with data centers will be solved.

When I hear people ominously intone that American data centers may triple their energy usage by 2030, I agree that this will be a struggle for the American energy grid, but on a per capita basis where we consider all global users of American data centers, I always have in the back of my mind that this means going from this:

To this:

Resources on data centers

This post was a mashup of four previous posts I’ve written to get the idea across more clearly. I expand on a lot of the points in this post in each of these:

• For the climate, little things don’t add up

• When making cuts for the climate, always consider the value you’re getting per unit energy, not just the total energy used

• Computing is efficient

• Do data centers only seem bad for the environment because we can see them?

One possible objection to this piece is that Jevons Paradox implies that making computing more efficient might make it use more resources in general. I’ve tried to respond to this objection at this link.

Some great overviews of data centers:

• How to Build an AI Data Center

• The Datacenter as a Computer: An Introduction to the Design of Warehouse-Scale Machines, Second Edition

• Hannah Ritchie on forecasts for how data centers will affect the climate into 2030

• An older article from before the AI boom addressing worries about internet data centers. This is a longstanding source of confusion.

• A long podcast overview of data centers

Unfortunately, MPU’s presentation of a few key data center issues is wildly deceptive. Viewers are left with a much worse understanding of the issue. Many millions of viewers are being tricked.

Here I’ll focus on a core false claim they reinforce over and over again: normal data center operations use so much water that they limit water access for regular people. They very regularly present locals saying this, here are some examples I gathered from different videos:

There is never any indication from MPU that American data center water usage has not raised the household cost of water at all anywhere, limited household water access in any way, or depleted local water supplies. Data centers aren’t using that much water. They behave like any other normal industry, and America overall is relatively good at water economics. In many areas, they actually significantly improve local water access by giving the utility enough revenue to make necessary upgrades. If you’re skeptical, you can read my deep dive on this here. The numbers are clear and straightforward.

I’ll run through a ton of examples of MPU being wildly deceptive on this. This is not just a selection of the worst clips, it’s almost all their presentations of data center water usage. I link to each full video so you can see if I’m taking anything out of context.

They key confusion MPU often relies on is that there have been a few places where the construction (not the operation) of new data centers has polluted local groundwater. The most prominent example is Meta’s large data center in Georgia, which seems to have added sediment to local groundwater and caused neighborhood wells to try up. This was the subject of this New York Times story, which many misunderstood to mean the data center itself used up water. These issues were all caused during construction. The data center hadn’t begun operations, and doesn’t draw groundwater. I have a breakdown here of ways the piece implied (incorrectly) that data center operations have raised water costs elsewhere. To be clear, Meta wronged these people, and stricter regulations should have prevented this, but it’s entirely a construction issue that could happen with any large building. It’s not an issue unique to data centers. There are very few other cases where water was polluted at all of the ~5,400 data centers in the US.

MPU videos edit together extreme false claims from locals about far-off data centers using water. This gives MPU themselves cover to push false ideas. They get to say “We’re just sharing what everyday people on the ground believe!” even though these people don’t have any experience with data centers themselves. At the end of this post, I share a long interaction I had with the head of MPU which seems to confirm this strategy.

Examples

A first example with a long deep dive on the claim

Take this clip from MPU’s coverage of a proposed AWS data center in Arizona:

This implies the Meta data center in Georgia didn’t just add sediment to nearby groundwater during its construction, its operational water use “bled the state dry.” The lead up to this clip is entirely about how much water the data center will use, not construction. This leaves the viewer with the impression that Georgia’s water access has been severely harmed by the normal operation of the data center.

The claim that Meta “bled Georgia dry” is wildly false. But the real picture has some nuance. The data center in Georgia began operating in 2021. Residential water-sewer bills in Newton County remained nearly flat from 2015–2022 (~$73/month), then rose in 2022 due to economic issues caused by the pandemic making electricity and other inputs more expensive. By 2025, the average bill had risen to $85.10/month, with projections to $111.08/month by 2029.

The NCWSA’s own budget documents cite the second Meta datacenter campus (Baymare) as giving “impetus to maintaining the capital improvement program in high gear,” so Meta’s expansion is one factor in the accelerated capital spending driving part of the rate increase. However, there are strong reasons to think Meta is far from the primary cause:

• The NCWSA also cites uncertainty in the housing market and high inflation rates as factors impacting costs.

• Newton County expects an industrial wastewater reclamation facility to come online in 2026 which will on net reduce total industrial use of freshwater in the county. The county is anticipating needing to provide less, not more, water to data centers, and yet the rate increased anyway.

• Meta’s data centers are only using ~2% of the county’s water. For comparison, a pharmaceutical manufacturing plant is using ~4% of the county’s water. A construction plant for Rivian cars is using about the same amount of water as Meta’s data center. The data center is one of several large industrial users driving infrastructure investment, not the sole or even primary cause.

Of the main industries using water in Newton County, Meta’s data centers are providing way more tax revenue per unit water back to the local county. The county received ~$1,900,000 from Meta, ~$520,000 from Rivian, and ~$2,400,000 from the pharmaceutical factory. Dividing by the water each industry uses, Meta provides $0.01 of tax revenue per gallon used, Rivian provides $0.005, and the pharmaceutical factory provides $0.009. In terms of tax revenue provided to the community, the data centers are the most water-efficient. You would not know this from watching this video.

So the framing that Meta “bled Georgia dry” is wildly misleading — Meta is using a small fraction of county water, contributing more tax revenue per gallon than comparable industries, and the rate increases have multiple causes. But it’s fair to say that data center expansion is one contributing factor to the capital improvement spending that’s raising rates in Newton County. This is the single example I’ve been able to find of data centers partially contributing to a rise in water costs, and even here the data center is one driver among several. MPU’s framing is deceptive, but the picture isn’t as simple as no impact whatsoever.

This claim is presented in this video without any question or pushback, and the activists saying it are very clearly framed as the good guys. The subtitles make sure to grab a “Busted!” jeer at the AWS representative from someone in the crowd. The viewer is left to conclude there’s some truth to the claim that the normal operation of the Meta data center harmed Georgia’s water access.

With this context, circle back and watch this full clip of the hearing (I’ve had this video skip to the hearing section). Think about what image this video paints in viewer’s minds about what’s happening in Georgia:

Let’s expand out to the general claim that data center operations harm local water access. Here are a bunch of individual places where MPU strongly implies data centers tend to harm local water access where they are built.

Many more examples

Keep in mind throughout all of these that there are no instances anywhere of data center operations raising water costs or limiting residents’ access to groundwater. Try to ask yourself what message viewers are taking from these videos, and whether they will leave open to the truth that data centers haven’t harmed water access. I’ve tried to be comprehensive and select most parts of their videos where they explicitly mention water usage, and what they say.

A common move is to share a contextless large number. For example, this tweet makes this sound like a huge amount of water:

7.2 million gallons per year! Sounds like a ton. How much is that? This data center would represent about 0.02% of nearby El Paso’s water usage. MPU videos also regularly share large numbers like this to shock viewers who don’t know how much water everyday industries use.

This video description seems to strongly imply that the normal operation of the data center is causing water issues for a local community:

In the video itself, it’s made clear that this is another example of data center construction altering the water in a very specific community, not the normal operation of a data center. The company is not “using up the drinking water.” Throughout the whole video, the data center is still being constructed and not using water at all!

Next, this title:

For context, a million gallons of water per day is exactly as much water as the maximum normal demand of that pharmaceutical plant in Georgia. A single large Anheuser-Busch brewery in Fort Collins uses 2.5x as much water. This is definitely a large amount, but it’s comparable to other large industrial uses of water.

More importantly, MPU is using a trick to maximize clicks here:

In the video itself, a farmer worries about data centers using up all the local water, something that hasn’t happened anywhere in the country:

And later in the same video, the claim is made that this data center threatens the community’s water supply:

This is an odd framing. “Threatening a local community’s water supply” sounds like AI data centers are somehow more harmful than any other normal private industry. It’s hard to imagine the host saying something about a brewery or electric car factory “threatening community water.”

It may be that both of these examples actually exclusively mean “The construction of the data center might contaminate very local groundwater” but this is never made clear to the viewer, who is left to infer from the title that the main problem with the data center will be its “1 million gallons of water per day.” In reality, the normal operation of data centers has not “threatened community water supplies” anywhere.

The thumbnail on this next video definitely implies that data centers are limiting everyday people’s water access:

The title of this video is “Amazon’s Secret Plot to Build a Data Center in the Desert.” This seems to imply that data centers in deserts are shockingly bad. In reality, Arizona already has large numbers of data centers throughout the state. They collectively use about 1/50th as much water as just the golf courses in Arizona, but generate more tax revenue than the golf industry. Building data centers in the desert has been normal for a while and doesn’t harm water access. Here’s a comparison1 of water usage for data centers and golf courses in the county in Arizona with the most data water draw:

The introduction to this video (up to 0:25) implies that this will threaten water access:

And the longer interview with Victoria here does too:

The full video’s worth watching, it’s very deceptive throughout. It leaves you without an understanding that

1. Arizona already has a large booming data center industry.

2. All data centers in Arizona are using a tiny fraction of the state’s water use.

3. Water costs in Arizona haven’t risen at all due to data center demand.

Golf is using 50 times as much water as all data centers in Arizona, and providing way less local tax revenue.

Here’s a clip from a short here where they announce that a community won against big tech “trying to tap into the community’s water” as if this poses a problem, or is somehow different from all the other normal industries using way more water in Arizona:

Here’s a more blatant example of the interviewer fishing for an interviewee to share the water use angle. There are almost no examples of data centers polluting water at all, except for a few bad cases, one in Georgia and another in Indiana:

Here’s an example from this video where they have a local farmer share a concern about something that hasn’t happened anywhere, with zero pushback:

Here’s a six minute video called “Google threatens the water supplies of a drought stricken town"

The town this video is focused on, The Dalles in Oregon, is already the location with the highest percentage of water going to a local data center. 29% of all city water goes to a local Google data center. Water costs stayed flat until 2025, then rose a bit to fund a general upgrade to the water system. Some assumed this upgrade was purely to support the data center, but the city government responded to the question of whether the upgrade was motived by the data center with this:

No. The replacement of the water treatment plant and water transmission lines, the two largest and most expensive projects, are needed because the existing facilities are at the end of their useful life. It is now more expensive to rehabilitate and expand these facilities than it is to replace them. The water treatment plant is 75 years old and has experienced significant deterioration of its concrete structures. The water transmission pipelines, which are about 80 years old and made of unlined steel, are deteriorated and leaking. These projects would be needed even if the City did not have Google as a water customer.

The county used tax revenue from Google to fund the majority of the upgrade. They estimated that without the revenue from the data center, the cost of the upgrade would raise local water bills by 23%. Instead, households only paid 7.3% more for the upgrade. So in the place with the highest percentage of water going to data centers, the data center slightly lowered water bills.

You would not know any of this from watching this video. There’s a strong implication throughout that the data center is merely extracting water that people could otherwise use.

At one point in the video, one of the interviewees says “These corporations are the ones that need to pay something because they’re utilizing a resource and they’re not paying. But they’re keeping profits for their shareholders and themselves.” Google’s data centers in The Dalles provide more than 15% of the entire county’s tax revenue. The idea that they’re “not paying for the resources they’re using” is a lie.Separate from taxes, the Google data centers still pay normal utility bills for the water. If the local water utility decides that they are using too much, they can raise commercial water rates to reflect the scarcity of the resource. This is a common strange move in MPU presentation of data centers’ relationships to local governments. They talk as if the data center is not providing anything to the area, and the government has decided to allow it to come in and make everyone’s lives worse for no reason. In reality, there are usually a lot of benefits to a town’s public services in having a data center built, for the same reason towns are benefited by the presence of any other well-functioning private industry: they pay a lot in taxes, and the water utility gets a lot of additional revenue to fund upgrades.

The main water issue in American small towns isn’t the supply of water, it’s aging water infrastructure that doesn’t serve a large or rich enough tax base to get the money to upgrade. Old infrastructure makes water more expensive. It can also be dangerous (lead pipes etc.). Small town water costs are often higher, not lower, than cities, due to economies of scale in water production. This wouldn’t happen if water costs simply rose when more water is used. In several well-documented cases, data centers moving into smaller communities have provided utilities with revenue to fund infrastructure upgrades — The Dalles is a classic example of this dynamic playing out. Here are a few others:

• Council Bluffs, Iowa: Google pays for expanded water treatment plant.

• Quincy, Washington: Quincy and Microsoft built the Quincy Water Reuse Utility (QWRU) to recycle cooling water, reducing reliance on local potable groundwater; Microsoft contributed major funding (about $31 million) and guaranteed project financing via loans/bonds repaid through rates.

• Goodyear, Arizona: In siting its data centers, Microsoft agreed to invest roughly $40–42 million to expand the city’s wastewater capacity.

• Umatilla/Hermiston, Oregon:Working with local leaders, AWS helped set up pipelines and practices to reuse data-center cooling water for agriculture, returning up to ~96% of cooling water to local farmers at no charge.

What’s happening in a lot of areas like The Dalles is an otherwise poor community is having a very rich, well-run private industry move in, pay a ton of taxes, buy a lot from local water utilities. Not only can I not find any places where data centers raised water prices at all, there are lots of places where having a new big customer revitalized a poor community’s water delivery system. MPU leaves viewers with the sense that the opposite is happening.

There are no examples of MPU presenting the “data centers don’t harm local water access” side fairly

Every now and then MPU interviews someone who supports a data center build, but the editing makes it clear that they’re the bad guy in the video.

Here’s an example where the camera focuses a lot on a pro data center council member’s lavish house and personal water pool. The interviewer says among other things “This is a battle between tech billionaires and people - working class folks, who want to protect their water. Do you see this as a fair fight?” and later says “Supervisor Scott says he doesn’t think the people should be allowed to vote on this” and then the camera pans back to a chanting crowd. Funny enough, the clip opens with another org denying MPU an interview because they’re worried about biased media coverage. They’re right to be! MPU seems pretty biased here! It’s very funny that they cut straight from this note to a wildly biased interview where they make it clear to frame their subject as the bad guy.

Conclusion

It’s pretty easy to find a lot of examples where a viewer will leave an MPU video with a strong sense that data centers are limiting everyday people’s access to water, and that anyone saying otherwise is a shill for the AI companies, is actively lying to you, and ignoring the voices of everyday people on the ground. This sense of harm and guilt really cements the idea in viewers’ minds. In reality, this has not happened anywhere. One of the most popular sources of information about data centers online is giving people straightforwardly false ideas about how they work and affect local water supplies. Regardless of your opinions about other issues with data centers (pollution, electricity costs, or issues with AI more broadly) we should all agree that this is unacceptable for an outlet with such a large audience.

My run-in with the head of MPU

This will be a long account of a twitter spat I had with the director of MPU. I’m normally don’t want to mix Twitter drama with my blog, but MPU’s director’s bad behavior here led me to update pretty negatively on how MPU approaches simple facts. Most importantly, neither he, his staff, or any of his 100,000 followers could provide me with an example of where normal data center operations limited water access at all.

It started when I posted this:

The director of MPU quoted it with this:

I felt pretty confident that I was correct here, so I replied, and it led to this exchange:

This was the end of this thread. Faiz and I had a few other interactions, but he could never justify the widespread implications throughout MPU videos that data centers harm local water access. Here are a few other interactions we had:

Didn’t get an answer on that one. Faiz’s strategy here seems to be interviewing locals about their opinions of far-off data centers they’ve been misinformed about. The magic here is that MPU isn’t factually responsible for what locals say, so they can film people saying and implying anything that fits their narrative, and then get away with filling videos with false claims. MPU then gets to declare “Hey, we’re just showing what people are saying.” Further, they get to cast these false beliefs as coming from the voices of the oppressed, not their own directors and editors trying to push a very specific story.

Here’s an instance where he swooped in later, still dodging the central question of whether data centers have raised water prices at all, anywhere:

I’ve unfortunately come to the conclusion that MPU does not seem interested in the facts of this question, and are more excited about running with a general anti data center line that gets views, even if they actively mislead viewers about pretty important information. I think it’s a shame that they’re having such a big effect on the data center debate and would encourage people to look elsewhere for news and resources on AI data centers.

Something seems wrong here.

I think this is actually just another way to phrase the statement “Tiny things we do may not seem like much, but if everyone does them, they have a big impact on the climate.” Both seem equally silly to me, but the second is very common in conversations about climate change.

Tiny parts of our personal emissions are tiny parts of global emissions

I’m going to make an obvious point that I worry is glanced over in these conversations.

Suppose this is a single person’s emissions. The red represents the emissions of a tiny activity they do.

Some might say that this small red square will add up if a lot of people do it. What would several people doing this look like?

We can combine them together into a single clump of total emissions.

Gathering the red dots together:

Notice that the ratio of the red activity to all the other emissions is the same for the group as the individual person.

This obviously makes simple mathematical sense. 1/100 = 4/400. We’re just multiplying numerators and denominators by the same constant. The ratios themselves stay the same.

But I worry that this basic intuition is lost on a lot of people when they say things like “tiny things add up.” Tiny things add up absolutely, but they don’t add up relatively. They often remain the same tiny percentage of total emissions as they are of individual emissions. An activity that’s 1/10,000th of your personal emissions is likely to also be 1/10,000th of global emissions. It’s only as promising a way to reduce global emissions as it is to reduce your personal emissions. If you’re trying to rally everyone to stop an activity that wouldn’t meaningfully raise your personal emissions at all, this is a sign that you’re wasting your and others’ time attention and effort on a drastically ineffective thing to do in the name of stopping climate change.

AI fits this pattern perfectly. All data centers emitted 180 Mt CO2 in 2024. AI likely used about 15% of data center electricity, so it emitted around 30 Mt. Most of the emissions from AI come from electricity used in data centers, not from physical construction of AI hardware, so its total emissions weren’t much more than 30 Mt.

Globally, the world emitted 41,600 Mt CO2 in 2024. AI was responsible for 0.07% of emissions last year. This includes every single instance of AI. Not just chatbots, but literally every instance of deep learning (there are a lot of these). Prompting chatbots 100 times during the day would add about 0.07% to your daily emissions. The reason global AI is 10x higher as a percentage is a combination of the fact that you interact with many other AI applications throughout the day without realizing it, Americans emit way more than average, and the average global citizen isn’t interacting with AI (or other things) as much. You’ll notice that when you see AI compared to total emissions, its proportion of total emissions looks very similar whether you’re looking at an individual person using chatbots a lot or the world as a whole.

This shouldn’t surprise us, because tiny things on an individual level don’t add up to huge things on a global level, they stay the same tiny proportion globally as they are individually. The world is really only individuals putting different demands on the broader energy system, so this makes sense. A behavior change is only as promising a solution for climate at the global level as it is for your personal carbon footprint. Tiny things have equally tiny potential for actually moving the needle on how much we emit.

Making the list

What if you could make a list of demands for every behavior change you’d want everyone to make? The list would be ordered by which actions would help the climate the most if everyone did them. You’d want everyone to know about the first few and act on them. Maybe you could get the first 5 across to a wide audience. After that it starts to get more uncertain.

What would be at the top of the list? What would you prioritize? For me, we’d start with these three:

1. Find any ways you can push for converting the grid to green energy (I’d recommend Clean Air Task Force to get up to speed on where you can be most useful).

2. Donate to well-researched high impact climate charities.

3. Vote for politicians who will take significant positive actions on climate.

There’s a big power gap between these and any specific lifestyle changes you could make. Lifestyle changes are nothing in comparison to big systematic changes to our energy grids. The differences are not small. They’re gigantic. Here’s an example I used before to illustrate the point:

Suppose there are 3 people who each want to have an impact on the climate: Kate, Bob, and Freddy. They each independently choose their own ways of impacting the climate. All seem really noble and self-sacrificial. Kate joins a committee of 500 people working for a year to keep a nuclear power plant open for another 10 years. Any impact they have will be divided by 500 people. Bob goes vegan for a year. Freddy has a debilitating ChatGPT addiction and prompts it 2,000 times per day. About one prompt every 20 seconds every waking hour. He quits for a year.

Try to form an idea in your head of how their climate impact compares to each other. It’s not immediately obvious.

After 1 year, Kate has prevented 70,000 tonnes of CO2 from being emitted. Bob has prevented 0.4 tonnes. Freddy prevented 0.2.

This means that Kate had as much effect as 175,000 Bobs, or 350,000 Freddys.

Here’s how many people would need to go vegan for a year to match Kate’s impact:

I worry that when people sneer at quantifying climate interventions, they don’t realize how gigantic the differences are in what we can do.

To use a very blunt analogy, if we’re in a war against climate change, individual contributions to changes to the energy grid are like nuclear bombs, and individual lifestyle changes are like sticks of dynamite. It often seems ridiculous to me to even mention them in the same breath.

But let’s say you want to get those most important lifestyle changes across to people anyway. What would be most important? I’d push for relatively simple changes that aren’t especially disruptive to people’s lives but still cut huge amounts of carbon; the things most people can stick to that will actually make some kind of noticeable dent.

4. Buying green energy from your grid or installing a private solar panel.

5. Buying an electric car.

6. Skipping a flight if there’s another way of traveling.

7. Doing a deep one-time retrofit of your home.

Try making your own list like this. Think about what 20 interventions for the climate you would most want to communicate to the average person. Try to think about how many the average person could actually be convinced to follow. Where would you get the most climate impact for your effort?

Where would AI fit into this?

These are things the average American does every day that emit more than prompting ChatGPT 100 times (28 g CO2, and yes, including training and the embodied emissions of hardware):

• Making coffee

• Running an AC for 10 minutes

• Running a fridge normally

• Taking a 1 minute hot shower

• Using hot water for 1 minute to rinse dishes

• Ironing clothes

• Microwaving food

• Cooking on an electric stovetop for 5 minutes

• Leaving 8 LED bulbs on for an hour

• Running a humidifier for a few hours

All of these things would be way, way, way down my list of the most important things for Americans to cut for the climate. We have a limited number of ways the average person will be willing to adjust their lives. A limited amount of messaging we can get to them before they focus on other things. Only if I knew that I could literally get a majority of people to change their lives in hundreds of other ways for the climate would I then consider suggesting “being mindful of your AI prompts.” As of right now, the vast majority of Americans wouldn’t consider paying more than an additional $40/month in taxes to combat climate change. They are not ready to make hundreds of additional lifestyle changes, and if we get to a future of abundant green energy, we won’t need to ask for these changes anyway.

Worse, if an activity is hundreds of lines down this list, unless you’re sure that the person has also cut every single thing before it, it is very likely that if they cut the activity, they’ll replace it with something higher up on the list that emits more and is less useful to them. So going too far down this list puts you in more and more danger that you recommendation increases rather than decreases their emissions. Not only is it distracting and ineffectual, it might even raise the average person’s emissions.

Not many people are actually thinking about climate change much

One of the reasons it’s so important to get climate messaging right is that for a majority of people who hear it, it will be one of the only times they think about climate change at all. Only 23% of Americans say they have thought about climate change a lot.

65% say they rarely or never discuss climate change with friends

A majority of Americans report that their friends and family don’t believe it’s especially important to put effort into stopping climate change.

And 44% don’t believe the actions of individuals will have any effect on the climate.

Unfortunately, this does make it seem like climate messaging is somewhat zero sum: one message people hear may crowd out others. Each message will often be one of the only times the listener considers climate change. This means we need to communicate especially important information when we do recommend lifestyle changes. Sacrificing this rare important opportunity on a message that will cut literally millions of times fewer emissions than other possible lifestyle changes seems like a decision to crowd a very limited field with something that won’t help at all. Advertising extremely tiny cuts as significant is effectively a form of kneecapping real climate action. I see people going around warning about the impacts of personal AI use as irresponsible, in the same way I think it’s irresponsible to imply that you can have significant climate impact if you randomly unplug your fridge every now and then. Messages like this are crowding out serious climate conversations for the sake of something that will never be a significant part of our emissions. The competition to reach everyday people is already extremely crowded, and people waste valuable opportunities when they throw out innumerate recommendations that won’t save us from the worst effects of climate even if all 8 billion people alive ruthlessly followed them.

Collective action

Climate is a collective action problem.

We all have strong incentives to use more fossil fuels. Fossil fuels are, unfortunately, miraculous. It’s a miracle there’s a liquid you can spend a few minutes filling your thousand pound metal vehicle with that will singlehandedly blast it hundreds of miles. There’s no reason these had to exist, but they do, and because they’re so miraculous it’s very hard to make green replacements for them, so those replacements cost more money. Life’s easier if you personally don’t worry about climate change and just do what you want.

Your individual emissions will, on their own, literally never matter for the climate. You could fly in a private jet every day of your life, and your emissions wouldn’t make a dent in the amount that’s actually going to cause tipping points for the climate. No one in a hundred years will notice any difference in the weather based on your entire lifetime emissions.

However, if a lot of people don’t change their emissions, climate change will make life significantly worse. Larger and larger catastrophes will happen.

This is a classic collective action problem: Everyone would be much better off if we all took some action, but everyone individually has a strong incentive not to, and each person’s contribution won’t matter unless a lot of people do it. Thus, the only way we can really encourage people to cut their emissions is if we reliably show that many other people are doing the same, so there will in fact be collective payoff for individual sacrifices.

Because climate is a collective action problem, we should keep a few things in mind:

• People have strong reasons to emit. Asking them to change their behavior and make sacrifices for the climate is already a lot. We can’t presume they’ll have infinite willingness to change. If they did, we would have solved climate change already.

• We need to find simple rules that everyone can know everyone else is trying to follow. If they’re too complex, citizens will lose hope that others are doing their part of the collective action problem, and they themselves will be less likely to play along. These simple rules also need to be things that will actually reliably reduce emissions a lot. The rule “pepper your daily life with random impulses of guilt about some tiny emissions but not others. Cut based on ominous news stories rather than actual numbers” seems like a terrible simple rule to give people as a way of solving the collective action problem of climate, yet that seems to be a very popular message in a lot of spaces. We should discourage this.

For more on collective action problems I’d review the prisoner’s dilemma. We need to remember in climate communication that we are not merely building big cool coalitions of the virtuous in-the-know people, we are asking everyone to take scary leaps into potentially more difficult lives to solve a problem that will only really be solved if many other people take the same leap. This is hard! It’s why solving climate change is so difficult. We need to think about this as threading a needle. People who run around hyping people up about tiny emissions merely because they “add up” to still very tiny proportions of global emissions are not taking this problem seriously.

If we want to limit warming to 2 degrees by 2100, this is a bad time to do that.

Let’s try to figure out what it is.

The full cost of a prompt

We’ll start with the agreed-on best rough guess for the median ChatGPT prompt: 0.3 Wh of energy. This includes the energy costs of cooling in the data center, idling chips, and data center overhead.

What is this missing?

Training the model

Well it doesn’t factor in the cost of training the model. There’s no agreed-upon ratio of how much energy went into training GPT-5 or how many prompts it will receive, so we can’t just divide the energy by the prompts. We do know that ~40% of all energy used on AI in the US is used on training new models right now, so we can roughly approximate that the 0.30 Wh represents ~60% of the true cost when you include training. So training + inference gets us to 0.5 Wh / prompt.

Emissions

Of course, we don’t actually care about the energy used. We care about the emissions. The average American power plant emits 0.37 g CO2 per Wh, but data centers in America use energy that’s 48% more carbon intensive than average. Combining these:

(0.5 Wh / prompt) x (0.37 g CO2 / Wh) x 1.48 = 0.27 g CO2 / prompt.1

Transmission from the data center to your computer

Transmitting information from data centers to your personal device uses minuscule amounts of energy. 1 GB of data transfer across the internet costs ~10 Wh. A standard ChatGPT response is 1–10 kB of text. This means that transmitting each prompt response to your computer uses at most 0.00001 Wh. 0.002% of the cost of a prompt. It’s negligible.

Embodied carbon of the AI chips

What about the cost of physically making the AI hardware? There’s a new study that’s the most comprehensive attempt to estimate the “embodied cost” of actual AI hardware, and says that only 4% of the emissions associated with an AI chip used for GPT-4 come from producing it. 96% of the chip’s emissions happen when it’s used for training and inference. This makes sense. AI chips are designed to have energy running through them 24/7. It’s kind of like if you made a wire, and measured all the carbon emitted when generating all the electricity that ran through the wire over its entire lifespan as part of the wire’s emissions. You’d expect most of the wire’s emissions to come from this electricity, not from producing the wire itself. In the same way, the actual use of the chip should be pretty energy intensive relative to building it.

Conclusion

So if the embodied emissions are 4% of the cost, 0.27 g CO2 is 96% of the cost of the prompt. The full carbon cost is ~0.28 g CO2 / prompt.

This is the same amount of carbon emissions as:

• Running a microwave for 3 seconds

• Using a laptop for 1 minute

• Running a clothes dryer for 1 second

• Driving a sedan 5 feet

• Playing a PS5 gaming console for 15 seconds

It’s 0.0005% of the average American’s daily emissions. An American emits ~200,000 times as much every day. You have to prompt ChatGPT 2000 times in a day to increase your emissions by 1%, but prompting it that much would take up a lot of time and probably prevent you from doing other high emissions activities like driving, so unless you were constantly prompting ChatGPT every moment while multitasking with everything you do, it would be very very hard to actually raise your emissions at all using ChatGPT. It seems like, even with all the hidden costs included, ChatGPT is much more likely to lower your emissions, because it emits so much less than the average ways Americans spend their time.

What about more abstract costs?

Normalizing AI

Maybe using AI is bad because it “normalizes AI” and causes other people to use it.

For the most contagious strains of COVID, the average unvaccinated person could expect to infect ~8-10 other unvaccinated people.

Let’s say using AI is as “contagious” as the most contagious strains of COVID. If you start using ChatGPT, 10 of your friends start using it too. Maybe in this case you’re responsible for all their emissions too.

So every prompt you send causes 10 other prompts from your friends. Any one prompt is actually 10x as bad. The cost changes from 0.28 g CO2 to 2.8 g CO2.

Here, each prompt is still just a 20,000th of the average American’s daily emissions. Here are 20,000 dots:

You would have to prompt ChatGPT 200 times just to raise your emissions by 1%.

So “normalizing AI” doesn’t raise ChatGPT prompts up to a level where you should be concerned about them.

Digital infrastructure

Maybe we should include the full carbon costs of the digital infrastructure and supply chains required to support AI, not just the data centers themselves. By using AI, you’re complicit in the rapid buildout of data centers and global electronics supply chains that support it, not just the individual energy cost of your prompt.

Something strange about comparisons of the full infrastructure costs of AI to other things we do is that they often fail to account for the infrastructure costs of those other things.

A recent popular blog post was titled “Why Saying “AI Uses the Energy of Nine Seconds of Television” is Like Spraying Dispersant Over an Oil Slick.” The author’s main point is that each individual AI prompt is able to use so little energy only because of this vast and expanding background buildout of AI infrastructure, so just reporting (as I do) that an AI prompt only uses as much energy as a few seconds of a microwave is hiding the more ominous reason why it’s able to be so cheap in the first place. By using AI, you’re complicit in some way in that infrastructure buildout.

This criticism would make more sense to me if everything else in society didn’t also have a vast sprawling physical infrastructure supporting it. “9 seconds of TV” has huge networks of electronics systems supporting it, as well as crazy amounts of money and people-hours going into making the most entertaining TV, lavish (often wasteful) lifestyles enabled by the profits from TV. Obviously, TV advertising also encourages people to buy more stuff from other complex supply chains.

If you make a comparison like this:

One off cost of an AI prompt + the full infrastructure supporting AI ←→ 9 seconds of TV

then it’s easy to make AI seem like the bigger problem, but if you make what I think is the correct comparison instead:

One off cost of an AI prompt + the full infrastructure supporting AI ←→ 9 seconds of TV + the full infrastructure supporting TV

Then I suspect the infrastructure costs of AI and TV will roughly cancel each other out, and you might as well just make the original comparison:

One off cost of an AI prompt ←→ 9 seconds of TV

This is why I think it’s reasonable to make this comparison.

We don’t really hold this standard for anything else we talk about. I can say “Your sedan emits about 320 g of CO2 for every mile you drive” and I don’t think that’s deceptive, even though the sedan is relying on a vast road infrastructure that costs 69 million tonnes of CO2 each year in America alone just to maintain, the sedan itself has “embodied carbon costs” from manufacturing it that I’m not including, and driving a car normalizes the behavior for other people. I think people understand that these additional costs exist when they talk about how much cars emit per mile, and I think they also understand these costs exist when we talk about how much AI emits per prompt.

AI infrastructure is being built out faster than most other infrastructure, but the rate of change of growth of something’s emissions doesn’t on its own tell us much about how bad it is for the climate. The carbon emissions from the global supply chain of Labubus recently began to rapidly increase, but Labubus are going to remain such a tiny part of global emissions that this doesn’t matter. What matters is the total amount of emissions and how much value we’re getting from them.

AI and electronics will obviously emit way way more in total than Labubus. However, the IEA expects the data center buildout to, on net, significantly decrease emissions overall, because AI will be optimizing so many other processes in society and making green energy and smart grid tech more viable. They project that AI alone will prevent 4 g of carbon emissions for every 1 g all global data centers emit (for both the internet and AI). So I could end up saying something strange like “Every ChatGPT prompt encourages the data center buildout, which is good because forecasts imply that the buildout will on net lower global emissions, so every ChatGPT prompt you send decreases emissions by x amount.” We’ve ascended into a level of abstraction that I think is goofy. Things get goofy when you try to cram the responsibility for every possible outcome of AI into the individual cost of an AI prompt. Your predictions about this “total, hidden” cost will mostly depend on the decisions people make in the future about how and where to use AI, not on ChatGPT’s individual in-the-moment impact itself. Seems goofy!

This goofiness is why I think it makes more sense to just limit what we say about the climate impacts of individual prompts to those prompts themselves, and leave broader forecasts of AI’s total climate impact in the future for separate conversations.

Using ChatGPT is not bad for the environment

This will be my final word for now on the costs of individual ChatGPT prompts. If someone approaches you and says that AI has “hidden environmental costs” that are big enough that your individual prompts add in any meaningful way to your daily carbon emissions, you should politely but firmly tell them that they’re wrong. Specifically, push them to share what the actual cost is and how it actually compares to other regular things we do. It shouldn’t be so easy to just vaguely gesture at this “real cost” without having some rough idea of what it actually is, and yet people do it all the time with no pushback.

There’s a separate conversation to be had about the net environmental impacts of AI overall, or of data centers built in specific places. That’s very different than the individual impact of individual AI prompts. Both conversations are important, but we need to be clear which one we’re having. And when it comes to the cost of individual prompts, the numbers are clear. In general, I don’t think the average person should focus much on their personal carbon footprint compared to systematic changes to the energy grid, so I’d really rather we just have the other conversation anyway.

The climate does not react to which industry emits or which specific buildings emit. It only reacts to the total CO2 in the air.

The national microwave

If data centers didn’t exist, we would need to rely on our personal computers to run everything we do online. When you logged into YouTube, you’d run YouTube software on your own computer like it was a video game, and save all videos you upload on your own computer. Every time someone else wanted to watch your videos, they would need to connect to your computer like other players on a video game can connect to a game you’re running. The magic of data centers is that by piling huge amounts of computing in one place, you can make all that computing more energy efficient than it would be otherwise, and more accessible for everyone involved.

Other things don’t work this way. If I want to microwave something, I have to own my own microwave. I can’t send things off to some centralized microwave somewhere else.

What if I could?

Let’s say there were a single national microwave. When you needed food heated up, you would teleport the food to the national microwave, in the same way you effectively teleport information from your computer to data centers.

This microwave would need to be massive to fit all microwaved meals Americans are making at any given time.

I estimate that all microwaves in America use ~25 GWh of electricity every day. This means that the big national microwave would be using as much electricity every day as Seattle. An entire new city’s worth of electricity demand added to the grid, just for heating food!

What about the water? Well the average water consumption per kWh in American power plants is 4.35 L, so this microwave would be consuming 110,000,000 liters of water every single day. 45 Olympic pools of water every day.

I think if the big national microwave existed, it would be drawing a lot of scorn. “This seems so wasteful. What’s wrong with just using an oven?” People would see it as sapping a whole city’s worth of energy. There would be articles about how it’s sapping the local grid and using as much energy as 800,000 households.

But this microwave does exist. It’s just hidden, broken up into hundreds of millions of little pieces of itself scattered across America. The impacts are just as real as if it existed in one place. The climate does not care about where emissions happen. It just responds to total emissions. All microwaves in America are emitting just as much as the big national microwave would. But they don’t exist in one place as a single evil building people can get mad at. The big national microwave is invisible.

Suppose someone says “Tech bros just reinvented the oven, only this time it’s destroying the planet.” Would we want people to stop using the big microwave and switch to their ovens?

Well, ovens use way more energy to cook the same food. Getting people to stop using the big microwave and switch to ovens would actually cause way more emissions in total, probably around 10 times as much. Those emissions would just be dispersed across the country, so people would feel better about them, because they couldn’t see them. People really like when they can’t actually see the effects they’re having on the climate, but those effects are still real. Shutting down the big national microwave would be a huge environmental mistake, even though it looks much more evil than everyone’s individual ovens.

These are three mistakes I think people would make if we had the big national microwave:

• They would only see it as a single, inert building using a ton of energy, and wouldn’t compare its benefits per unit emissions to other buildings. They wouldn’t consider how many more people were interacting with it, or divide the energy cost by the number of people using it. It would make sense that a building managing all microwave needs for all America would be using more energy than a nearby toy store, because it would be serving a lot more people. The toy store is actually way more inefficient with its energy, but comparing it to the toy store would be ridiculous. If it sounds ridiculous to say “All American microwaves are using more energy than this one toy store, people should all throw away their microwaves!” then it should sound equally ridiculous to condemn the big national microwave for the same reason.

• They would want to shut it down and have switch to things that seem more “normal” like using an oven, even though those normal things are mostly way worse for the climate. The normal things don’t have a big evil building of their own to represent how much energy they’re really using, so unlike the microwave, their emissions are invisible, even though they’re much greater in total.

• They would only see the building as using a lot of energy relative to everyday people (800,000 households!) and wouldn’t consider it in the context of climate overall and how big of a problem it is compared to most other things we do. They would prioritize optimizing the microwave over much more easy climate wins, like shutting down coal plants or electrifying cars.

People also make these three mistakes when thinking about data centers. When you see climate arguments against data centers, check and see if the same arguments would apply to the national microwave, and wouldn’t apply to regular microwaves. That tells you that the other person is just mad that emissions are visible, not that they’re a lot.

Data centers as the national microwave

Data centers are weird buildings. They concentrate more power demand in a smaller space than basically any other type of building. They also, at any one moment, have hundreds of thousands of people from around the world interacting with them.

From the outside, data centers look boring and inert.

But in some sense they’re the most active physical objects we’ve ever built. They’re building-sized computers constantly running gigantic amounts of equations. Microsoft’s 2023 GPT-4 training cluster reportedly had ~25,000 A100 GPUs in a single build. This amount of chips could perform around 10¹⁸ (million trillion) addition or multiplication calculations per second. All humans who have ever lived, doing one calculation per second for 24 hours per day, could together do that many calculations in about 100 days.

Each data center is more of a massive very general tool than a normal building.

Like the national microwave, hundreds of thousands of people interact with a data center at once. People effectively teleport computer tasks away for some far off building to deal with it much more efficiently and reliably. When you read stories about data centers using huge amounts of power, you should think “This is because tens of thousands of people are interacting with the building at once and using it as a tool, like a gigantic national calculator.” You might think the way the data center’s being used is bad or wasteful, but you should still hold in your head that this is just a very very concentrated result of 10s of thousands of small individual decisions, like the microwave is.

Data centers are uniquely energy efficient. Basically all the energy they take in is delivered directly to incredibly optimized computer tasks. They’re the most energy efficient way to do a lot of computing. Since computing can give us valuable information about the world, it seems good to be able to build these huge clusters where computing can happen more efficiently than anywhere else.

Individual data centers can grow so large that they take on the energy demands of whole cities, but globally all data centers on Earth (supporting the entire internet) use 1.5% of global electricity, and only 0.23% of global energy. Data centers are managing and hosting the global internet. The average person on Earth spends 6 hours and 40 minutes every day on the internet, so they’re spending 40% of their waking lives interacting with data centers. Data center efficiency means that something we all spend almost half our lives interacting with only uses 0.23% of our global energy. I’m not aware of any other service that comes close to this.

All global data centers were responsible for 0.48% of emissions in 2024. On this chart of where global emissions were happening (in 2016, more recent data’s hard to find) all data centers would be about the size of the “rail” line.

Obviously this is not nothing, but take a minute to look through everything else in the infographic. If you hadn’t heard of data centers before, and saw them listed as a tiny stripe (they would be a small sliver of the “commercial” stripe), would something emitting 0.48% of global emissions stand out to you as something that required special attention? How many conversations in the last year have you seen about the emissions of cement, or landfills, or paper & pulp production?

I think a part of the reason you haven’t seen more discussions of the latter is that they don’t have big evil buildings of their own. There’s no big centralized paper production mill that’s using huge amounts of energy and water people can get mad about, even though paper production overall is emitting significantly more than data centers right now. People are less upset about those emissions because they’re invisible.

Identifying climate bad guys by just looking at what seems like it’s a bad guy is a terrible way to think about the problem. It makes it easy to become tricked by industries hiding or normalizing their emissions.

Of course, unlike other industries, data center energy use is growing.

The IEA forecasts that data center emissions will peak in 2030. This is what their emissions look like now, and will look like in 2030, compared to other industries and emissions overall. I’ve split them up into the IEA’s implied forecasts for AI specifically and non-AI applications:

Again, not nothing, but the 2030 numbers are the very most emissions data centers are forecasted to ever cause in one year.

Unlike most other things on this list, data centers can help us in basically every other area of climate. Having very general tools to do immense numbers of calculations and optimization tasks is useful for tackling other problems. Of all the areas I’d want to cut for the climate, having giant warehouse-sized calculators for these tasks seems like one of the last places to cut. Even though data centers are newer, they’ll probably be much more useful for solving the climate crisis than most other industries.

But these other industries don’t have big evil buildings of their own. They have much more effectively hidden their emissions from everyday people, or normalized them so people treat them as permanent unchangeable background parts of the world. If we want to think seriously about climate, we need to be willing to slice through those illusions and think of the whole world as more malleable. When we do that, data centers stop standing out. They take their place among other very energy-optimized industries that serve a lot of people and emit relatively tiny amounts per user, and the main climate villains (things like cars, livestock, and bad energy management in buildings) become more apparent.

More than anything else, it’s the energy grid itself that needs to change, rather than the relatively small individual ways we use energy. 1.5% of electricity in the world is being used on computing in data centers. The problem isn’t that this number should move from 1.5 to 0% (computing is incredibly valuable!), it’s that 100% of electricity should be green in the first place.

None of this is to say that data center emissions don’t matter at all, only that they’re taking up a suspiciously large part of the climate conversation relative to how much they actually emit, or how beneficial it would be to stop building them.

If cars or livestock or energy waste had their own big evil buildings, it would become much more obvious how big of a climate disaster each is. It’s our job as people thinking about climate to make the invisible visible.

A lot of people in debates about AI and the environment are talking about Jevons’ Paradox: Making something more energy efficient sometimes causes it to use more energy in total.

The way this works is simple. Making something more efficient often makes it cheaper (energy costs money), so more people use it. If it becomes popular and enough people use it, it adds up to consuming more total energy than if it stayed inefficient and unpopular. Technically, Jevons’ Paradox applies to any resource, not just energy and emissions, but we’ll stick to climate-relevant stuff for now.

Jevons’ Paradox doesn’t always happen. There are plenty of cases where making things more efficient was good for the climate and didn’t lead to more emissions. It’s just a dynamic we need to watch out for.

An example

Imagine all cars are super inefficient, and only get 1 mile per gallon. The gas to power them would be too expensive for the average person. Maybe a few rich people would use them, but that’s it.

Someone concerned about the environment says “Wow, those cars rich people drive are burning so much gas. If I make them more efficient, that will be good for the environment!”

This person finds a way to optimize cars to raise their efficiency from 1 mile per gallon to 30 miles per gallon. They sit back and think “I’ve done such a great thing for the climate, so much gas will be saved. So much emissions prevented.”

Now that the cars get so many miles per gallon, everyone can afford them. Most people buy cars, and the total emissions of all cars skyrockets, becoming one of the main sources of all greenhouse gases on Earth.

Oops.

Jevons’ Paradox and AI

There’s growing worry that AI is facing the same issue. As labs make their models more energy efficient, they’re able to offer them for less money in more places. This means that overall, AI is getting used so much more that the efficiency gains are causing it to use more energy (and emitting more) in total.

A lot of conversations frame Jevons’ Paradox as a negative. Here’s an example from a recent article:

Google claims to have significantly improved the energy efficiency of a Gemini text prompt between May 2024 and May 2025, achieving a 33x reduction in electricity consumption per prompt. The company says that the carbon footprint of a median prompt fell by 44x over the same time period. Those gains also explain why Google’s estimates are far lower now than studies from previous years.

Zoom out, however, and the real picture is more grim. Efficiency gains can still lead to more pollution and more resources being used overall — an unfortunate phenomenon known as Jevons paradox. Google’s so-called “ambitions-based carbon emissions” grew 11 percent last year and 51 percent since 2019 as the company continues to aggressively pursue AI, according to its latest sustainability report.

That sounds bad. Here’s an even more extreme take, from a blog post that’s been popular recently: Why Saying “AI Uses the Energy of Nine Seconds of Television” is Like Spraying Dispersant Over an Oil Slick

TLDR: the smaller the number, the more you should be worried

Claims about the energy footprint of AI are suddenly everywhere. “AI search summaries use 10 times more energy than a traditional search”. “Each chatbot session consumes 500 liters of water.” Recently, Google issued its own report, full of long awaited technical detail that specialists are poring over, but the headlines were simpler: “Each prompt of Gemini uses as much energy as watching nine seconds of television.”

…

Now imagine that I made money on each prompt (or persuaded investors that I did, even though I don’t). That gives me every reason to get better at building the infrastructure to serve all of those prompts. I add data centers, each more efficient than the last. With each new round of building, the wattage of my data centers goes up and up. The racks get denser and denser, and I need more and more of them, all over the world. $600 billion worth of them, as a down payment.

In other words, the only way my per-prompt energy cost goes down is if my overall energy footprint goes up.

This is the counterintuitive reality. The lower the energy used per prompt, the more concerned you should be.

This isn’t exactly Jevons’ Paradox (the causality is flipped), but it implies a similar dynamic. The author doesn’t mention Jevons’ Paradox by name, but I assume they’d agree it’s relevant here. The piece isn’t just about Jevons’ Paradox, it’s also about how society’s general digital infrastructure is wasteful and uses a lot of energy. To put it mildly, I disagree, but that’s a separate question.

I think there’s a lot of good-faith concern about Jevons’ Paradox, but I don’t think this level of concern makes sense when you think about what’s actually happening.

A lot of framings I see of Jevons’ Paradox are actually pretty terrible, and don’t get across the subtleties of the idea. Many conversations about AI and the environment include a very simplistic and limited understanding of the trade-offs involved here.

Let’s look at those trade-offs.

Jevons’ Paradox in an industry is only bad if it leads to more total global emissions, not just more emissions in the specific industry

Solar panels

The cost of building new solar panels has been rapidly falling since the 70s.

As a result, solar power has exploded way faster than energy experts predicted.

There are a lot of reasons for this, but one is that solar panels have become more efficient and less resource-intensive to produce (using less energy per panel to produce), so they’ve become cheaper, and as a result way more people are buying them (they use way more resources and energy in total). This has been a wonderful surprise for those of us who have been following climate for a while. In the 2000s, it seemed unreasonable to predict such rapid growth.

The emissions we spend on producing solar panels are completely cancelled out by the emissions the solar panels prevent, and the panels go on to prevent way more. Exact estimates vary a lot, but let’s say they save 10x the carbon creating them emits.

The climate only responds to total emissions. If making a solar panel adds 1 unit of carbon, and using it subtracts 10, the solar panel is exactly equivalent to something that adds 0 units of carbon and subtracts 9. We should want to make more solar panels for the sake of the climate for the same reason you should want to keep making trades where you give someone $1 and they give you $10. You’d really want to max out those trades. Your bank account doesn’t care about the individual $1 bills you lost, it just responds to the net total at the end, which will be $9x(the number of trades). The climate doesn’t respond to the fact that solar panels emit 1 unit of carbon, it only really responds to the fact that solar panels on net remove 9 units of carbon from the atmosphere. We should mash the “build solar panel” button a lot, just like we should mash the "trade $1 for $10” button.

Imagine that you went back in time to 1980 and showed someone this table comparing solar panel dollar and carbon costs, and the total emissions of all solar panel construction:

Their reaction is:

Oh no! This is terrible. Solar panels have fallen victim to Jevons’ Paradox. Making them cheaper has only made their total emissions explode. We should work hard to keep them expensive and inefficient. We need to stop this from happening!

How would you respond? What would you tell them?

I would tell them this:

Jevons’ Paradox is only bad for the climate if the product will, on net, increase total CO2 emissions, not just the emissions of the product itself. Jevons’ Paradox is actually good news for the climate if the product being sold will, on net, prevent more emissions than the carbon cost of creating and using it.

It’s good that solar panels were affected by Jevons’ Paradox. In some sense, it’s a good sign that they’re emitting more now, for the same reason that it would be a good sign if you spent a lot more $1 bills in the trades that get you $10 each time. If you found out that Jevons’ Paradox had occurred, and now every $10 trade only cost $0.50, and as a result even more trades were happening, that would be a good situation for you.

Obviously, in an ideal world, we’d want constructing the panels to emit less, just like you’d rather pay $0.50 for each $10, but the fact that panels are emitting so much more in total is, if anything, a sign that way, way more emissions are being prevented in total as well. Jevons’ dynamic in solar panels is a way to get society to make way more +1 -10 carbon tradeoffs.

There are industries where Jevons’ Paradox has been very bad for the climate. The most obvious is cars. Cars becoming more affordable and energy efficient has been uniquely bad for the climate, because cars emit a lot and don’t prevent as much emissions elsewhere.

These examples show that it’s only bad for the climate if an industry experiencing the Jevons’ Paradox dynamic is expected to, on net, cause more total emissions than it prevents. If it’s expected to prevent more emissions than it causes (especially if it’s expected to prevent more and more as it grows, like solar panels do) then it’s neutral or even good that the Jevons’ dynamic is happening.

AI

At first, it might seem crazy to compare AI to solar panels. Solar panels are a miraculous climate technology. AI is a very broad general tool that’s emitting a lot.

I’m not asking you to think of AI as miraculous climate tech. All I’m asking is to consider the question “Will AI, on net, prevent more emissions than it causes?”

What experts predict

It is really really really hard to predict what is going to happen with AI. Expert predictions are useful here, but we should take them with a grain of salt. Here’s a compilation of expert predictions. I’m trying to paint a complete picture here. If you think I’m leaving any important predictions out from generally trustworthy experts, let me know and I’ll add them:

• 2025

  • IEA: Energy & AI: If widely adopted, existing AI applications could cut emissions by ~5% of energy‑related CO₂ in 2035; warns that rebound (e.g., shifts to autonomous vehicles) could undercut benefits. IEA’s “widespread adoption” case shows use‑phase savings that exceed data center emissions in 2035, but that outcome is not guaranteed. The IEA separately predicted that data centers’ total greenhouse gas savings in 2030 will be 4x-10x as much as much as AI’s total emissions in data centers.

  • Grantham Research Institute & Systemiq: Green and intelligent: the role of AI in the climate transition: With effective application in power, transport, and food consumption, AI could reduce 3.2–5.4 GtCO₂e/year by 2035 (≈ single‑digit % of global GHGs), and authors argue these savings would outweigh AI‑driven power demand.

• 2024

  • Nature Communications: Potential of artificial intelligence in reducing energy and carbon emissions of commercial buildings at scale: AI adoption in U.S. commercial buildings could cut energy use and CO₂ by ~8–19% by 2050. Energy in commercial buildings is responsible for ~26% of global emissions, so this one AI application alone would reduce emissions by at least 2%. Globally, this alone would more than cancel out all data center emissions.

• 2022:

  • IPCC: AR6 WGIII: Digitalization (incl. AI) can enable mitigation but may increase energy demand and wipe out efficiency gains without proper governance; highlights rebound effects as a material risk. (No single net number.)

• 2021:

  • Brockway et al: Energy efficiency and economy-wide rebound effects: A review of the evidence and its implications: Median economy‑wide rebound ≈50–60% of expected energy savings (meaning on net AI will still save significantly more energy); in some cases ≥100% (i.e., savings erased).

A general pattern emerges where it seems like in expectation, reports are either finding that AI will net out to 0 or be pretty net positive for the environment. The IEA separately wrote that they expect the emissions prevented by AI to be 3-4x higher than all global data center emissions (including all data centers that support the internet).

I lean in the direction that it seems likely that AI will, on net, reduce emissions rather than increase them, and the more it’s used and deployed the more net emissions it’s likely to reduce. In this scenario, the Jevons’ dynamic with AI is leading us to a world where AI reduces total emissions more on net, so we should be cautiously optimistic about the fact that the Jevons’ dynamic is happening, in the same way we’re optimistic about the Jevons’ dynamic with solar panels (though with AI we have much more uncertainty).

Why would AI reduce emissions?

Most of AI’s effects on emissions will probably happen in the way AI’s used, not in the data centers that run it.

Consider Amazon. There’s an ongoing debate about Amazon’s net effects on climate change. Amazon optimizes package deliveries, but maybe causes people to consume more in total, and there’s a debate about how the emissions of the supply chains Amazon uses compare to emissions of other things we buy.

But no one in the environmental debate about Amazon has said “The most important question is how much energy its website is using in data centers.”

That’s because the energy data centers use is tiny relative to the website’s effects on economic and personal behavior.

Another example of a way the internet has an outsized effect on the physical world is Google Maps. Maps uses hardly any energy in data centers, and has optimized billions of car trips to be as physically short as possible. It seems certain that the emissions Maps has saved completely dwarf the emissions it’s caused by using energy in data centers. Google reported that just five of its apps (including maps) prevented ~26 million tonnes CO₂e in 2024. All global data centers (AI + the internet) combined emitted 180 Mt CO2 in 2024, so 5 Google apps alone might have prevented 14% of all current data center emissions.

This shouldn’t surprise us. Cars emit a lot and aren’t especially optimized. Finding ways to optimize driving (which both the internet and AI are useful for) should yield a lot of low-hanging fruit for reducing emissions.

AI is obviously using way, way more energy in data centers than Amazon or Maps, but as I’ve written about before, this is mainly a result of AI being used so much, not each individual prompt using a significant amount of energy. AI has the potential to have large effects on the physical world per unit energy it uses in data centers. The more energy efficient each prompt is, the more potential there is for the effects on the physical world to be much greater than the energy costs in data centers.

The IEA expects data center emissions to peak in 2030, at about 320 Mt CO2 that year.

Data center energy use will continue to grow after, but will be balanced out by much greener energy sources, so emissions will (very slowly) fall.

For context, 320 Mt will (under their most aggressive path to net zero emissions) be ~1.3% of global emissions that year. Obviously not nothing. But what this means is that if all AI and all the global internet combined are optimizing processes in society enough to reduce global emissions by just 1.3%, they will completely pay for themselves, and anything after that will be net good for the climate. I think the internet is probably already doing this anyway, and AI will accelerate that to save even more emissions.

A lot of people’s exposure to deep learning is ChatGPT, which can often be goofy and give wrong answers, and is used to cheat or skip out on work. I think this is causing people to assume that deep learning more broadly also isn’t useful. This is wrong.

Deep learning more broadly seems incredibly useful, especially for climate. It’s useful wherever there are situations where we know roughly what we want computers to do or respond to, but we can’t give them step-by-step instructions for how to do that. It’s almost as if computers themselves have been invented a second time, this time to solve a huge new host of problems traditional computing could never approach.

Here are just a few fields where deep learning shows incredible promise to prevent emissions. These broad areas adopting deep learning each individually has the potential to reduce global annual emissions by 1%:

• Materials science: The discovery of better materials to use in renewable energy tech like solar panels and batteries). Even small efficiency gains here have huge climate payoffs. I think materials science might be the main field critics of AI’s environmental impacts aren’t paying enough attention to.

• Energy grid optimization. Smart grid technology (where the energy grid optimizes itself and moves and stores energy based on sudden changes in supply and demand) is going to be critical for building green energy, since many renewable sources (especially solar and wind) are very intermittent and benefit from a grid that can respond rapidly and intelligently to sudden changes in energy supplies. Deep learning has a lot of potential to benefit smart grid technology. AI can provide better weather forecasting smart grids can respond to.

• Building energy use optimization: Energy used in buildings is responsible for 26% of all greenhouse gas emissions. Even slightly more optimal heating and cooling systems would help a lot.

• Transit: deep learning. For flights, deep-learning guided contrail avoidance can significantly reduce plane emissions. Urban traffic signal optimization can reduce local air pollution for pedestrians as well as emissions.

The Jevons’ Paradox in what AI optimizes

You might have noticed that a lot of the above examples themselves involve making other industries more efficient, so they risk inducing the same Jevons’ dynamic there. This is far from guaranteed, but there are specific cases where AI reducing energy costs seem likely to create a Jevons’ dynamic for another industry. The most likely place for this to happen is autonomous vehicles. It seems likely that autonomous vehicles will be extremely optimized compared to most drivers. They’ll be able to constantly monitor traffic patterns and other AV behavior better than humans (not to mention that a human won’t need to spend time driving them). This could reduce costs so much that on net way more people drive (as well as people who couldn’t drive before, like kids). This seems likely to raise emissions.

The question is whether new demand from cheap goods AI provides will on net lead to higher emissions. As of writing, it seems like experts are unsure but lean in the direction that AI is more likely than not to reduce net emissions. A lot of this will depend on good policy and practices around AI use that environmentalists should push for. It seems like AI’s energy in data centers will be a small part of its climate impact.

My argument in a nutshell

Deep learning seems poised to prevent more emissions than it causes. To do this, it probably needs to be deployed widely across a lot of society (having a deep learning model optimizing all city traffic lights to minimize stops, as an example), which will require a lot of energy, but on net this energy will “pay for itself” in the same way solar panels “pay for themselves.” If the Jevons’ Paradox is happening to a product like this, that seems net good for the environment. The more AI takes up our energy grid, the more likely it is that way more emissions are being prevented elsewhere. This isn’t perfectly linear (obviously people use AI for a lot of goofy things) but it seems likely that a world where AI permeates the economy more and optimizes lots of complex problems (and uses more energy as a result) is also a world where on net more stupid inefficiencies are being ironed out, new discoveries in green energy tech happen faster, and more emissions are prevented in total. We should have a lot of uncertainty about this, but expert predictions seem to lean in the direction that AI will prevent more emissions than it causes, and this will scale with how much it’s used in specific areas. This already happened with the internet.

I wouldn’t put all my money on AI being net good for the environment, but I also wouldn’t automatically assume it’s bad that Jevons’ dynamic is happening here.

I disagree for 2 reasons.

First, I’d written them hoping people would spend less time worrying about their individual lifestyles and more time thinking about how to actually systematically change our energy grid. I don’t want people to spend much time worried about their personal carbon footprints, and I definitely don’t want them to think that “does ChatGPT use much energy?” is the only relevant question in the AI/environment debate.

Second, the AI/environment issue is massive, and each individual question involves an incredible amount of information and nuance. Just making the case that using ChatGPT is not bad for the environment took me ~20,000 words. This is just one tiny variable in the huge debate. It’s not possible to write anything that both covers all relevant environmental questions about AI and goes into sufficient detail to do each full justice, unless it’s a full book. Saying that an article that doesn’t address every single environmental issue with AI is “meant to distract from the other harms” is basically saying we can’t do any specific deep dives on anything relevant to AI’s environmental impact at all.

I want to make it clear which parts of the debate I’ve tried to address so far, and which I haven’t. I also want to show why it’s so unreasonable to expect any one blog post to address every single issue (I’ll be attempting that soon regardless). I thought it’d be useful to map the debate.

Trying to map the debate

Emissions

Even just focusing on the climate impacts of AI can be difficult. I’m going to try to write out the main parts of the debate as different variables in a big equation for AI’s total harm to the climate.

A point I’ve been trying to get across is that the reason AI is using a lot of energy is that it’s being used a lot, not that each individual use burns a lot of energy.

We can write this as a simple equation:

My two big posts on individual chatbot prompts were exclusively focused on the (average energy/prompt) part of this equation. I haven’t posted anything about the total energy AI is using (~62 TWh in 2024, about 0.2% of global electricity), or the number of prompts and tasks all AI is handling. This post is a good overview of how rapidly AI’s taken off, it makes it clear that the total number of prompts is so large that even though the energy per prompt is small, the total energy used is high.

Of course, when thinking about the climate, we don’t actually care directly about the energy used. We care about the emissions. Total emissions will be equal to the average emissions per unit energy used by AI.

We can swap total energy used for the right side of the original equation to get:

Of course, the total emissions of AI aren’t only determined by the emissions per prompt. They’re also determined by the emissions caused by training the AI models, and the full embodied emissions of building and recycling AI chips. Each of these adds to the total emissions:

Already getting pretty messy…

The emissions caused by AI probably won’t mainly come from the energy they directly use in data centers. It’s likely that the vast majority of AI’s impact on emissions will come from how it’s used, not the impacts of the data centers themselves. AI is likely to both cause and prevent a huge amount of emissions. It could cause more emissions by making autonomous vehicles so readily available and cheap that the average person ends up driving much more. It could reduce emissions by advancing materials science that improves the efficiency of green tech (which would have way, way more impact on emissions than all data centers combined). So we need to add even more terms:

I’ve made the point before that we can’t just look at the total emissions of a process, we need to make some decisions about how useful or useless (or bad) it is. A children’s hospital emits way more than a Hummer, but I’d rather get rid of the Hummer than the children’s hospital. AI’s overall value is a thorny question. I’ve made some points about it here and here, but it seems hard to score and add to this equation. We could try to score it by adding a “value score” from 0 to 1. 0 is “AI is completely useless” and 1 is “AI is so useful that no amount of emissions we cause for it will be a problem.” We could then multiply the total emissions by (1 - (value score)) to get a sense of how bad AI is overall. This seems kind of silly, so I won’t write out the full equation. My point is that a lot of this debate is going to hinge on AI’s actual value as a service, not just its net emissions.

This equation is also limited in that it ignores how prompting AI incentivizes more AI use, or really any predictions about the future of AI, which are all over the place.

When you write all this out, it becomes clear that expecting every post about AI and the environment to address every single variable is ridiculous. There’s a crazy amount to say about each one.

Here are some relevant articles on each individual variable:

• # of prompts: “AI is the most rapidly adopted technology in history.”

• Average energy/prompt: Google’s recent report, Epoch, My posts here and here.

• Emissions/energy used: MIT Technology Review’s coverage, where they find data centers use energy that’s 48% more carbon intensive than average.

• Emissions from training models: I think everyone thinking about AI and the environment should make sure they understand the scaling laws AI labs are working under. Labs have a pretty strong incentive to make models larger and larger, because this consistently yields better and better performance. Here’s a report from Epoch on the gargantuan energy costs of training new models.

• Embodied emissions of AI hardware: Here’s a recent report trying to estimate AI’s embodied emissions.

• Emissions caused by how AI’s used: Jevons’ Paradox in general, and lots of specific examples like this: “A report published by the University of Michigan Center for Sustainable Systems highlighted that AV systems could increase vehicle primary energy use and GHG emissions by 3% to 20% due to more power consumption, weight, drag, and data transmission.”

• Emissions prevented by how AI’s used: “Net zero needs AI — five actions to realize its promise” and “AI Could Be Harnessed to Cut More Emissions Than It Creates.”

• Value we’re getting out of it: Here’s a very pro and very anti AI article. Here’s a very short and very long take from me. This debate alone is consuming so much of people’s reading time that it seems kind of hopeless to reach any significant agreement anytime soon.

• The future of AI’s energy use: Ridiculously hard to predict. The IEA has a decent report here.

Here are a ton of other questions about AI and the environment. I cannot possibly fit these all into a single equation:

Water

• Local water prices

• Local water ecology (which sometimes competes with local water prices. Phoenix residents who want maximally cheap water are not exactly friends of the local water ecology. Trade-offs need to be made between equity and local environments).

• National and global access to freshwater

• Water pollution

• Ways AI can optimize water use

• Water economics in general

Air pollution

• How bad is air pollution from AI data centers?

• How does it compare to other industries?

• How can the ways AI is used prevent or cause air pollution?

Electricity costs

• Rising electricity rates in different areas.

• Household vs. commercial electricity rates.

• How much should we focus on rationing the energy we have vs. building abundant green energy?

Mining materials

• Mineral mining’s harm to local environments.

Geopolitics

• Does AI increase or decrease the likelihood of great power war? Nuclear war would be bad for the environment…

Misc.

• Ways local regions could use tax revenue from data centers for the environment

• Will AI be an existential risk? The whole reason we work on environmental issues is for humanity and Earth to survive and do well. If there’s a decent likelihood AI itself could be an existential risk, this needs to be included in our calculations too.

Clarifying some of my beliefs on these questions

People have noticed that I’m not especially worried about AI’s overall environmental impacts, and some have inferred that this is creating a bias in my writing, or that I’m actively trying to present an incomplete picture to lure people into ignoring AI’s environmental harms. I want to clear up a few of my takes on these broader questions:

• The idea that individual AI prompts are a significant part of the way you might harm the environment in your day to day life is straightforwardly wrong, and anyone telling you otherwise is either deeply confused or lying to you. This is not a place where reasonable disagreement is really possible.

• I suspect that AI overall will on net be good rather than bad for the environment. This is a place I think reasonable people can radically disagree. I’m pretty unsure here. People much smarter and more knowledgeable than me have come to very different conclusions. A few reasons I suspect AI is probably on net going to be good for the environment:

  • AI is ultimately reliant on computers, and computing is basically the most resource-efficient thing we do as a society. Most physical activities humans do (flying, driving, shipping, farming, smelting steel) require orders of magnitude more energy per unit of output than running computations. The more AI uses our global energy budget, the more it’s likely to be optimizing lots of other processes that would have used significantly more energy without it. Consider Google Maps. It seems pretty likely to me that Maps has decreased people’s total time on the road by at least 1%. This saves way more emissions than Maps creates from the energy it uses in data centers. Computing gives us a lot of useful info for saving energy while spending very little energy to get that info.

  • The IEA expects all global data centers to reach peak emissions in the 2030s at around 320 million tonnes (Mt CO₂) per year. This is a lot, but to put this in context, it’s about 0.86% of global emissions right now. So if you believe that the entire internet, and entirety of AI use in data centers together, could by 2030 be optimizing science and society to the point that our net emissions are just 1% lower than they would be if the entire internet and all large AI tools together didn’t exist, all data centers will have “paid for themselves” and will be net good for the climate. The climate only responds to total emissions. If AI prevents more than it adds, it’s by definition good for the climate. Google already estimates that maps + an AI model that optimizes traffic in cities to reduce emissions prevented ~26 million tons of CO2e in 2024 alone, so AI and the internet might already be “paying for” 1/12th of their maximum ever global emissions just through a few specific Google apps. We shouldn’t be too surprised by this, because cars emit a ton, and haven’t really been optimized in the ways they could be. Think of all the red lights you’ve waited at with no cars coming, while your car just idles and spews out emissions, because the traffic light can’t detect that there’s no reason not to switch to green. AI and the internet emit very little relative to how often we use them, and there a ton of very simple ways to use them to optimize wasteful processes elsewhere.

  • Deep learning in general seems to be turning out to be a very powerful tool in a few key scientific areas, especially materials science. Despite the naysayers, deep learning just seems super useful for a lot of climate tech. Read more about that here.

  • This is not a fringe position. The International Energy Agency says:

    • The adoption of existing AI applications in end-use sectors could lead to 1400 Mt of CO₂ emissions reductions in 2035 in the Widespread Adoption Case. This does not include any breakthrough discoveries that may emerge thanks to AI in the next decade. These potential emissions reductions, if realized, would be three times larger than the total data centre emissions in the Lift-off Case, and four times larger than those in the Base Case.

    So in a world where we adopt AI in more parts of the economy faster, it looks like on net it will prevent more total emissions. If we achieve the high adoption situation the IEA describes, it will be on net reduce global emissions by 1100 Mt of CO2 each year. That’s the current total yearly emissions of all global shipping and aviation. Seems good!

I don’t want people to get these 2 beliefs above mixed up. I am certain that your personal chatbot prompts don’t meaningfully harm the environment or increase your personal carbon footprint. I’m pretty uncertain about AI’s environmental impacts, but most of what I do know about it pushes me in the direction of thinking it won’t be a major problem, and will likely be net good for the climate. I’m not trying to mix these two beliefs. You can (and should) agree with me on the first, without agreeing with me at all on the second. There are so many other variables and open questions that the first doesn’t at all imply the second.

Conclusion

You should be wary of anyone saying that an article going into details on one specific environmental issue with AI is “painting an incomplete picture” of AI’s environmental impact, because painting that full picture would require crazy amounts of very detailed writing. We need to be able to do deep dives on each of these issues individually without being told off for not including every other possible environmental issue along with it.

“Google says a typical AI text prompt only uses 5 drops of water — experts say that’s misleading”

Stop and think about what this headline conjures up for you. For me, it sounds like Google is being shifty with numbers to present its AI model as not using much water, and in reality it’s using enough water to at least consider as a possible problem.

A lot of people are under the impression that prompting chatbots adds a significant amount of water to their personal environmental footprint. This title would vaguely reaffirm that belief.

The article’s paywalled, so I’ll quote from it sparingly:

Google estimates that a median Gemini text prompt uses up about five drops of water, or 0.26 milliliters, and about as much electricity as watching TV for less than nine seconds, roughly 0.24 watt-hours (Wh)

This is from Google’s latest study, the most comprehensive report we’ve received on individual impacts of individual AI prompts on the environment.

But Google also left out key data points in its study, leading to an incomplete understanding of Gemini’s environmental impact, experts tell The Verge.

“They’re just hiding the critical information,” says Shaolei Ren, an associate professor of electrical and computer engineering at the University of California, Riverside. “This really spreads the wrong message to the world.” Ren has studied the water consumption and air pollution associated with AI, and is one of the authors of a paper Google mentions in its Gemini study.

Okay, this sounds bad. “They’re hiding the information” makes it sound like Google is actively choosing to withhold important information about how much water they’re really using. It’s not just a passive mistake, and this accusation is being shared by someone who seems to have expertise on AI and the environment.

When it comes to calculating water consumption, Google says its finding of .26ml of water per text prompt is “orders of magnitude less than previous estimates” that reached as high as 50ml in Ren’s research. That’s a misleading comparison, Ren contends, again because the paper Ren co-authored takes into account a data center’s total direct and indirect water consumption.

The article concludes:

If you look at the total numbers that Google is posting, it’s actually really bad,” de Vries-Gao says. When it comes to the estimates it released today on Gemini, “this is not telling the complete story.”

Nowhere this article is an actual number shared for how much water Gemini is using once you consider the offsite water. Google is “hiding the true water cost,” but so is the reporter.

Consider what your takeaways from this article would be. For me, I would think “Gemini’s water use is significant” and “Google is lying about how much water Gemini is really using. It must be bad!” There was a reference to estimates “going as high as 50 mL” so maybe it’s actually around there. That’s not 5 drops, that’s 1000!

Let’s see what happens when we look at the numbers.

The numbers

Calculating “offsite” water use for the power in data centers is easy. This is the water used in nearby power plants to generate electricity. Lawrence Berkeley National Laboratory’s comprehensive report on data centers finds that the power plants American data centers draw energy from use an average of 4.52 liters of water per kWh of energy generated. The Google paper found that their model uses ~0.3 Wh per prompt, and included the full energy cost in the data center per prompt. 0.3 Wh = 0.0003 kWh. So if the power plant uses 4.52 L of water per kWh, and the data center used 0.0003 kWh per prompt, then the power plant is using (4.52 L / kWh) x (0.0003 kWh) = 0.0014 L = 1.36 mL of water per prompt.

Adding this to Google’s original estimate for the water used in the data center itself, we get 1.62 mL of water per prompt.

So instead of using 5 drops of water, it’s using about 32.

Putting this number in context, the average American’s lifestyle uses 1600 liters of water per day when you include our food production, energy generation, and at home use (this consumptive use of specifically surface and groundwater, or “blue water”, where the water is not returned to its original source).

This means that Google’s original claim was that each Gemini prompt uses 0.000015% of your daily water footprint, and the correct value (that Google is quote “hiding” from you) is 0.00010%. Instead of 1 in 6.7 million of your daily water use, a Gemini prompt is actually as much as 1 in 1 million. Google originally misled you by announcing that it takes 650 Gemini prompt’s worth of water for 1 second of your shower. In reality, it only takes 97.

Do you think the average reader would leave the article with an accurate understanding of the magnitudes here? Is it fair to say Google is “Hiding” this information?

Not reporting offsite water doesn’t seem inherently deceptive, as long as you make it clear that you’re only reporting onsite water

I can see the argument for both reporting and not reporting offsite water, but I don’t think it’s especially deceptive to only include onsite water.

Obviously people are right to ask for the full water impacts of data centers on the locations they are built, so it makes sense to include them in the total water cost, and in most AI environmental reporting it’s become standard and expected. But Google is very clear throughout its paper that it’s only counting the water in the data centers itself. They define in their report:

Water consumption / prompt: the water consumed for cooling machines and associated infrastructure in data centers

It would seem deceptive if they were dodgy about what specific water they were measuring, but they give this very clear definition up front in their paper of exactly what they’re measuring. This does not seem deceptive.

Ren is correct to note that Google misleadingly compares their onsite water measurements to his total onsite + offsite measurements, and over-reports Ren’s original findings (they imply Ren’s maximum estimates were actually his averages). This is the place where Google did seem to behave somewhat deceptively, but to be clear, Ren’s original paper found that the US average data center’s water used per prompt was 16.9 mL. If we include both the offsite and onsite water, Gemini’s actual water cost is 1/10th of this value. While Google should have phrased this better, it’s still true that this is drastically lower than previous estimates.

Let’s reread one section of the article:

When it comes to calculating water consumption, Google says its finding of .26ml of water per text prompt is “orders of magnitude less than previous estimates” that reached as high as 50ml in Ren’s research. That’s a misleading comparison, Ren contends, again because the paper Ren co-authored takes into account a data center’s total direct and indirect water consumption.

The complaint here, that I think should have been spelled out in full, goes like this:

“Google incorrectly compares their new measurement of 0.24 mL of water per prompt to Ren’s original estimate of 16.9 mL, and they’re using that bad comparison to say it’s much smaller. In reality, each prompt uses 1.62 mL of water.”

To be clear, Ren has other complaints about misreadings of the data that are separate from the specific point I’m making here. I don’t want to misrepresent him, so you can read them in full here and here.

Why do this?

The title of this article could have been “Google says a typical AI text prompt only uses 5 drops of water — experts say that’s misleading… it’s 32 drops!” That’s obviously the fault of the person writing the headline, not the author.

It makes sense to write an article about how Google reported a number in a way that made it look lower than what industry standards require. It makes sense to focus on other specific issues with the paper.

It does not make sense to fill an article with talk about the “hidden offsite water cost” of the model without just telling your readers what that very publicly accessible water cost is somewhere in the article. It also does not make sense to imply that Google is “hiding” this when their report is very up front about only considering onsite water.

Someone could have bothered to just do the incredibly simple math to show what the numbers being talked about are. This information is easily available. Ren himself cites the study where I got the 4.52 L/kWh statistic. Everyone involved knows how to find this.

It definitely doesn’t make sense to leave readers hanging with the sense that their AI prompts use significant amounts of water. Everyone involved surely knows that a lot of people are under the mistaken belief that their personal AI use significantly adds to their daily water footprint.

I don’t want to single out the specific reporter, because this tendency to not give the simple small numbers involved in individual AI prompts is widespread across reporting on the topic. I do want to note that this reporter is not some rando who wouldn’t have known to include this number. Their bio says they are “a senior science reporter covering energy and the environment with more than a decade of experience.” I don’t understand why someone with their level of knowledge wouldn’t think “My whole article is about how Google is ominously leaving out the full water cost of AI. Let me just show my readers what that water cost is, using the incredibly easy method for calculating it.”

This leaves us with a misleading headline and article. Readers are left with the sense that Google’s actively lying to them about the true water costs of their models, and the unstated but seemingly obvious next step is to assume that the water costs are significant. In reality, Google’s paper makes it clear that they’re measuring only the water in data centers themselves. Including the offsite and onsite water means each Gemini prompt uses 1/38,000th of the water cost of a single beef patty. I think readers would come away from this article with a wildly different and wrong idea of how much water is involved.

Address misconceptions with me

We need a lot more people publicly calling out articles like this that leave readers with a sense that AI companies are using significant amounts of energy and water per prompt. Many many environmentalists have been tricked into thinking individual AI use is significantly bad for the environment, and are wasting a lot of time and energy encouraging others to boycott AI apps. This is a ridiculous waste of time for the climate movement, and it has been caused in part by reporters not sharing the simple numbers involved.

Calling this out is easy and just takes some simple math. If you ever see anyone make the complaint that reported water use isn’t using the offsite water, just use the fact that the average power plant providing energy for a date center uses 4.52 L of water per kWh (you can find it on page 57 here). Here’s a great summary of the Google report. You should consider writing about it publicly and posting it where other people can see it. I’ll have a much longer post soon on more ways to address this stuff.