Most of our media is not really science-literate (bad) and is therefore very wary of even performing basic math on the facts they're presented with (probably for the best). The problem is that they then rely on information from disparate scary quotes and sources with an axe to grind, or who themselves have also just collated factoids and soundbites.
So video essayist Bob reads a quote on X, which badly summarizes a preprint on the ArXiv; a big YouTube channel picks up the story; the BBC references the channel; the NYT references the BBC. Video essayist Bob says: "Don't believe me? Perhaps you believe the NYT and the BBC." The preprint goes to peer review and now includes a footnote quoting the NYT.
The past week, I have heard someone argue that every Sora 2 prompt uses 1 kWh of electricity and 1,700 liters of water which is then "forever" rendered undrinkable. This is so completely divorced from reality, you just want to despair.
This is the most grounded, data-driven take I’ve seen on AI and water use. The argument that AI is some outsized environmental villain when it comes to water just doesn’t hold up under scrutiny — especially when data centers are using a fraction of the water that golf courses or agriculture consume daily.
We should be critical of how tech impacts the environment, but let’s focus on the real issues (like electricity usage and e-waste), not the hyped-up distractions. Comparing AI water use to adding a few midsize towns across the U.S. really puts it in perspective.
> 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.
“You may love the AI, or you may hate it, but it consumes surprisingly few atoms which you can use for other purposes.”
Thanks for the thorough article. I agree there's a critical distinction between water withdrawal vs consumption. However, are there adverse impacts from excessive amounts of withdrawal? For instance, competition - albeit for a limited time - with other uses of water, particularly in contexts of scare water resources? Or the redistribution of water from one area to another via evaporation that negatively impacts the area that experiences reduced water resources?
Good point, I'm pretty busy but would like to circle back on this soon. I'm trying to push more people to explore the issue too, if you write something up I'd read it!
You know what would be great is an infographic that shows the different components of AI water use as a percentage of total use in 2030. Perhaps there’s already one out there?
You write that "consumptive use can harm total access to freshwater, but freshwater sources are also regularly being replenished", but this is certainly not true - there are myriad studies showing that most freshwater aquifers in the US are being depleted at rates much faster than they are being replenished, and that this imbalance will soon and literally destroy huge swaths of the nation's agricultural, commercial and residential enterprises. In addition, surface water supplies are being seriously impacted / reduced by climate change, so these cannot be assumed to replenish at historic rates either. Therefore the argument that consumptive water uses for AI will get replenished like any other industry use weakens your position, because in many areas, AI consumptive uses are coming on top of water systems that are already stressed.
Nobody in these comments seems to know that Andy is an "Effective Altruist" and doesn't really give a damn about the environment of today if caring about it gets in the way of visions for human evolution thousands of years in the future. His priorities and our basic needs as humans are misaligned.
I think quite a few people know this? I'm as public with it as I can be. This is a very goofy accusation in multiple directions. I wouldn't write a post full of lies to steamroll current humans. Instead of trying to psychologize me could you point me to anything in the article I'm getting wrong?
The writers and editors at rolling stone are apparently innumerate:
"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."
that's a 66% increase, not 46%. Even if it's a typo by the writer, an editor should catch a basic algebraic error like that.
Is "energy use multiplied by 10" a little low? Epoch is pointing to about a 100x multiplier per training run. Assuming other costs scale proportionally with training runs the 100x seems more appropriate. Still probably too little to worry about from a water perspective but a 10x difference is worth mentioning. https://epoch.ai/blog/power-demands-of-frontier-ai-training
I'm not sure how this squares in the total figure of how much energy all AI will use so I'm sticking to IEA and other sources for now. Like I'm not sure as many training runs will be happening if each one takes 100x as much energy as GPT-4, that'd be pretty nuts and only a few companies would be able to afford it right?
Considering how poorly other players deal with exponential growth, I'd rather take epoch's figure of the cost of a frontier model training run as a proxy for energy use as a whole, especially considering other work they have done suggesting training and inference costs tend to stay balanced https://epoch.ai/blog/optimally-allocating-compute-between-inference-and-training
IEA's generally one of the most trusted voices in energy world, and yeah they famously got solar panels wrong but I'm inclined to defer to them otherwise. I chat pretty regularly with a few people at Epoch who so far mostly seem to agree with IEA's overall energy estimates for all AI data centers, but I can circle back and double check
That is interesting. I have no idea how much they make similar mistakes. My perception is that large bureaucracies like the IEA make lots of status quo assumptions due to institutional caution/conservativism and consequently systematically fail to make good predictions of exponential trends. But the Epoch folks think carefully about forecasting and if they agree then I would have no further reason to object!
But the link to how much water AI consumes is just some guy on the internet guessing. The figures in the article range from 200-628M gallons a day depending on the source (a Berkeley Labs report or his guess). Seems to me that there is a gap in accurate knowledge on the crucial number.
I agree the water issue is completely fake. However, as an AI researcher, it's hard to take a strong "AI is not bad for the environment" stance because even if it is using small amounts of water and energy now, it's clear that it will grow exponentially over the next decades. Water may never be an issue, but soon AI will use tens of percents of the total US energy, if not more, and at that point there will be a real conflict between AI and the environment.
It's important for AI companies to see where this is headed, and make sure that they have impeccable environmental records. They should be paying double or triple to ensure that every data center is 100% renewable energy and has NO negative local impacts. The deck is stacked against them. PR is very important and right now AI companies are doing a terrible job of it (except maybe Anthropic)
Most of our media is not really science-literate (bad) and is therefore very wary of even performing basic math on the facts they're presented with (probably for the best). The problem is that they then rely on information from disparate scary quotes and sources with an axe to grind, or who themselves have also just collated factoids and soundbites.
So video essayist Bob reads a quote on X, which badly summarizes a preprint on the ArXiv; a big YouTube channel picks up the story; the BBC references the channel; the NYT references the BBC. Video essayist Bob says: "Don't believe me? Perhaps you believe the NYT and the BBC." The preprint goes to peer review and now includes a footnote quoting the NYT.
The past week, I have heard someone argue that every Sora 2 prompt uses 1 kWh of electricity and 1,700 liters of water which is then "forever" rendered undrinkable. This is so completely divorced from reality, you just want to despair.
This is the most grounded, data-driven take I’ve seen on AI and water use. The argument that AI is some outsized environmental villain when it comes to water just doesn’t hold up under scrutiny — especially when data centers are using a fraction of the water that golf courses or agriculture consume daily.
We should be critical of how tech impacts the environment, but let’s focus on the real issues (like electricity usage and e-waste), not the hyped-up distractions. Comparing AI water use to adding a few midsize towns across the U.S. really puts it in perspective.
I thank you for doing this, Andy, but IMO, few of these people are arguing in good faith.
I'm not saying they are "evil" or stooges. Just that Doom is their religion.
https://www.mattball.org/2023/05/doom-force-that-gives-life-meaning.html
Interesting blog post
> 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.
“You may love the AI, or you may hate it, but it consumes surprisingly few atoms which you can use for other purposes.”
Except for the CO2 production from energy generation. Big problem.
Me: [asking Claude for a recipe]
My conscientious 11-year-old: "Don't! AI is terrible for the environment!"
Me: "We've talked about this before, and it's actually not. Look, here's this piece by this person I know."
Her: "But Mom, people can be wrong."
Glad she at least knows not to trust what some person you know said on the internet, even if the fact-checking abilities aren't fully developed yet.
Thanks for the thorough article. I agree there's a critical distinction between water withdrawal vs consumption. However, are there adverse impacts from excessive amounts of withdrawal? For instance, competition - albeit for a limited time - with other uses of water, particularly in contexts of scare water resources? Or the redistribution of water from one area to another via evaporation that negatively impacts the area that experiences reduced water resources?
Sounds good, Andy. Thanks for the reply. If I write on this topic, I’ll send it your way.
Good point, I'm pretty busy but would like to circle back on this soon. I'm trying to push more people to explore the issue too, if you write something up I'd read it!
You know what would be great is an infographic that shows the different components of AI water use as a percentage of total use in 2030. Perhaps there’s already one out there?
You write that "consumptive use can harm total access to freshwater, but freshwater sources are also regularly being replenished", but this is certainly not true - there are myriad studies showing that most freshwater aquifers in the US are being depleted at rates much faster than they are being replenished, and that this imbalance will soon and literally destroy huge swaths of the nation's agricultural, commercial and residential enterprises. In addition, surface water supplies are being seriously impacted / reduced by climate change, so these cannot be assumed to replenish at historic rates either. Therefore the argument that consumptive water uses for AI will get replenished like any other industry use weakens your position, because in many areas, AI consumptive uses are coming on top of water systems that are already stressed.
Nobody in these comments seems to know that Andy is an "Effective Altruist" and doesn't really give a damn about the environment of today if caring about it gets in the way of visions for human evolution thousands of years in the future. His priorities and our basic needs as humans are misaligned.
I think quite a few people know this? I'm as public with it as I can be. This is a very goofy accusation in multiple directions. I wouldn't write a post full of lies to steamroll current humans. Instead of trying to psychologize me could you point me to anything in the article I'm getting wrong?
The writers and editors at rolling stone are apparently innumerate:
"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."
that's a 66% increase, not 46%. Even if it's a typo by the writer, an editor should catch a basic algebraic error like that.
Interesting analysis to circulate among the enviro community in Tucson, AZ (where I live) that defeated Amazon’s proposed data center 🌵.
In addition to misinformation about water use of data centers, there is also misinformation about battery storage projects, causing people to fight against those projects in local communities. Here’s a good article about it: https://www.wbur.org/news/2025/11/20/battery-energy-storage-system-bess-massachusetts-lithium-ion-safety
Sharing this, crazy
I have been searching for figures that substantiate AI’s rampant water guzzling and have been totally unable to, thanks for explaining why!
Thank you for this and your other articles on the topic - they have helped me refine and rethink my own views on the subject substantially.
Have you seen this paper? https://arxiv.org/pdf/2304.03271
It finds that aggregate global withdrawal could reach 4–6 billion m³ per year by 2027: comparable to the water use of a mid-size nation.
This is a much older paper where the authors had less to work with, I’d defer to the LBNL’s estimates instead
Is "energy use multiplied by 10" a little low? Epoch is pointing to about a 100x multiplier per training run. Assuming other costs scale proportionally with training runs the 100x seems more appropriate. Still probably too little to worry about from a water perspective but a 10x difference is worth mentioning. https://epoch.ai/blog/power-demands-of-frontier-ai-training
I'm not sure how this squares in the total figure of how much energy all AI will use so I'm sticking to IEA and other sources for now. Like I'm not sure as many training runs will be happening if each one takes 100x as much energy as GPT-4, that'd be pretty nuts and only a few companies would be able to afford it right?
They're talking about the energy cost of a frontier model training run: only a few companies doing those now and I suspect it'll be the same in 2030.
IEA have ridiculously underrated exponential growth in other areas like solar deployment and you must have seen the crazy graphs of their predictions being wrong in the same direction year after year. Do you really want to take them as a reliable source of future projections? (https://pv-magazine-usa.com/2020/07/12/has-the-international-energy-agency-finally-improved-at-forecasting-solar-growth/)
Considering how poorly other players deal with exponential growth, I'd rather take epoch's figure of the cost of a frontier model training run as a proxy for energy use as a whole, especially considering other work they have done suggesting training and inference costs tend to stay balanced https://epoch.ai/blog/optimally-allocating-compute-between-inference-and-training
IEA's generally one of the most trusted voices in energy world, and yeah they famously got solar panels wrong but I'm inclined to defer to them otherwise. I chat pretty regularly with a few people at Epoch who so far mostly seem to agree with IEA's overall energy estimates for all AI data centers, but I can circle back and double check
That is interesting. I have no idea how much they make similar mistakes. My perception is that large bureaucracies like the IEA make lots of status quo assumptions due to institutional caution/conservativism and consequently systematically fail to make good predictions of exponential trends. But the Epoch folks think carefully about forecasting and if they agree then I would have no further reason to object!
I can circle back with them so don't quote me here, will let you know!
But the link to how much water AI consumes is just some guy on the internet guessing. The figures in the article range from 200-628M gallons a day depending on the source (a Berkeley Labs report or his guess). Seems to me that there is a gap in accurate knowledge on the crucial number.
And the "some guy" is an Affective Altruist with his own agenda.
I agree the water issue is completely fake. However, as an AI researcher, it's hard to take a strong "AI is not bad for the environment" stance because even if it is using small amounts of water and energy now, it's clear that it will grow exponentially over the next decades. Water may never be an issue, but soon AI will use tens of percents of the total US energy, if not more, and at that point there will be a real conflict between AI and the environment.
It's important for AI companies to see where this is headed, and make sure that they have impeccable environmental records. They should be paying double or triple to ensure that every data center is 100% renewable energy and has NO negative local impacts. The deck is stacked against them. PR is very important and right now AI companies are doing a terrible job of it (except maybe Anthropic)