“GPT-4’s footprint is roughly equal to the annual emissions of 1,550 US citizens” Was this intended by the author to seem like a large amount of emission? Are people really seeing this and thinking "AI bad"? It's tiny for a global product. Like, *obviously* you divide it by the number of users. Maybe I'm too optimistic about how people view numbers. (It's actually so small that I'm left wondering whether it could possibly be right for what it's worth)
On another note, I want to float something I'm tentatively calling the Masley paradox - the idea that training and using LLMs has negligible electricity associated with them compared to other activities on a per person basis, while data centre growth has such strong local implications for electricity costs (via increased network and generation requirements).
New data centres are largely going in specific locations where a) the regulatory and investment frameworks are favourable, and b) the internet cable connectivity is favourable, such as greater Sydney in my backyard, leading to a lot of local efforts to see it goes right.
I actually refer to this in a few other places as the environmental paradox of data centers yeah! They’re so efficient that per capita and globally they’re small, but so concentrated that locally they’re huge
Yeah this one really struck me. US life expectancy is 79 years so this is like saying training GPT-4 is equivalent to an extra 20 people being born in the US. Not very scary!
This is very helpful, since I’m updating a lecture on the resource use of AI for an AI literacy class I’m teaching right now!
A few minor points - in my academic field (philosophy) there are some people who are complaining about the carbon emissions associated with conference travel, and arguing that we should have more online conferences and fewer in-person conferences. (See, eg, the comments here: https://dailynous.com/2026/01/15/apa-to-end-experiment-with-online-divisional-meetings/ ) I expect there are some people making similar arguments on computer science.
One thing that this really drives home for me is that there’s a difference between the general resource use of AI and specifically the electricity use. Unlike legos and CDs, which involve moving physical objects and making things out of refined petroleum, AI has its emissions almost entirely through electricity. It looks like there are individual data centers that use as much electricity as the entire city of San Diego! (For instance, the Amazon datacenter for Claude training in New Carlisle, IN: https://epoch.ai/data/data-centers/ .) But the carbon emissions and water use of this datacenter are more comparable to the emissions and water use of South Bend, or perhaps even New Carlisle. The water use change could be a significant effect locally, but it’s nowhere near the potential disruption of having to power a new San Diego in the middle of Indiana!
The issues of water and emissions will be handled by the same processes that are handling those issues for everything else in the world. But electricity use will be a significantly different type of issue to deal with, especially as we also electrify transportation, cooking, and climate control.
Arguing that we should have conferences online instead of in person seems like a tacit acknowledgement that data centers don’t require a lot energy relative to transportation!
Great research Andy! Thanks for sharing this. It has been really eye opening. I have a couple of doubts. The first one is why you didn't mention the paper "How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference" by Jegham, N., Abdelatti, M., Elmoubarki, L., & Hendawi, A. title (https://arxiv.org/abs/2505.09598).I personally useded it to create a calculator for the cost of personal queries with this models (spolier alert: it was a lot less than i thought it would be). And the second one is if you got the chance to look at this news report form Business Insider about how tricky it is to find public records about energy consumption of theses companies (https://www.youtube.com/watch?v=t-8TDOFqkQA&t=1548s). Would love your tohughts on that.
Hi Andy! Another excellent post, thank you. At least in my experience the level of impact you're having on people's thinking on this topic is very large, more than enough to justify the level of effort you're putting in! Obviously that's easy for me to say since *I'm* not the one putting in that effort, but...
I've been asked this by people with whom I've shared your posts, so just to get it on record: what if any financial incentives do you have to say that AI is good, and/or that AI is not causing environmental harm?
Happy to answer any questions. I’m not being paid by anyone to say that AI is good/is not causing environmental harm, I don’t take money from AI or data center companies etc. I’m also not invested in AI companies specifically which sometimes comes up. I guess I have some incentive because of audience capture etc. but I think I resist that. I sometimes take small speaking fees for small professional or educational gatherings.
The couple of times it's come up I haven't tried to drill deeper since I didn't have any answers for them other than to point to a history of involvement in EA, but I interpreted them as meaning 'has significant investments in AI companies' (anyone with an index fund is somewhat invested in AI at this point, but presumably that's excluded). Anyone who thinks you're literally a paid shill probably won't be much moved by you denying it 😂
This brings to mind the rant I always go off on when folks here in California get all outraged at how much water it requires to grow an almond tree. X # of people use the same amount of water etc. Maybe it is a poor use of the water, or maybe it isn't.
If we use that water instead for something else, how do you compare the costs and benefits? Gallons/calorie, or grams of fiber or fat, or some other nutrient? Gallons/$ of profit, or $ of income taxes to the state , etc etc. Drives me nuts.
Are we all better off if we stop growing almonds entirely (given that we produce pretty much the world's supply of a delicious and nutitious commodity)? I dunno, maybe, but I've sure never seen the recipts showing that to be the case.
Really solid analysis. Training is a fixation for journalists because the boundary between training and inference is blurry or nonexistent for most people. I'd be curious to see what the numbers are for daily electricity use by a widely-used frontier model. I imagine inference is where the bulk of the energy demand actually originates.
Great work as always. My gut reaction to hearing any of these comparisons for the first time has always been: "But... that's _tiny_!"
And speaking of GTA V, I also have to wonder about the emissions of hundreds of employees driving to work for six days a week, for five long years, all using high-end workstations and servers...
There's a talking point that surfaced over the past few months, where people claim that the rapid iteration of new models is making the training energy consumption that much worse (and the profitability of any models unlikely), e.g. "GPT-5.0 was followed by GPT-5.1 just six weeks later!" Of course, the various versions of GPT-5.x don't involve the same massive pretraining of the entire base model, as they're mostly fine-tunes, merges, RLHF/SFT and whatever other arcane arts the AI labs employ.
I’m on the author’s side overall, but I think it’s important to point out that we keep creating new energy- and resource-intensive products and services. Sure, data centers are no worse than producing Coke or developing Grand Theft Auto, but we aren’t swapping one for another, we’re adding more. So I think when people hear how demanding a data center is, they’re adding it to all the endeavors they know are already consuming vast amounts of energy and resources.
Fair enough, but what about the carbon footprint of inference? You’re right that the carbon footprint of training a single model is amortized over millions of users, but inference costs scale linearly with usage, and over time will be much larger than the training footprint.
Have you seen convincing estimates of the carbon footprint accounting for inference costs?
"where the regulatory and investment frameworks are favourable" = where the governing bodies don't mind exploiting and selling out their constituents who live nearby
“GPT-4’s footprint is roughly equal to the annual emissions of 1,550 US citizens” Was this intended by the author to seem like a large amount of emission? Are people really seeing this and thinking "AI bad"? It's tiny for a global product. Like, *obviously* you divide it by the number of users. Maybe I'm too optimistic about how people view numbers. (It's actually so small that I'm left wondering whether it could possibly be right for what it's worth)
On another note, I want to float something I'm tentatively calling the Masley paradox - the idea that training and using LLMs has negligible electricity associated with them compared to other activities on a per person basis, while data centre growth has such strong local implications for electricity costs (via increased network and generation requirements).
New data centres are largely going in specific locations where a) the regulatory and investment frameworks are favourable, and b) the internet cable connectivity is favourable, such as greater Sydney in my backyard, leading to a lot of local efforts to see it goes right.
I actually refer to this in a few other places as the environmental paradox of data centers yeah! They’re so efficient that per capita and globally they’re small, but so concentrated that locally they’re huge
Yeah this one really struck me. US life expectancy is 79 years so this is like saying training GPT-4 is equivalent to an extra 20 people being born in the US. Not very scary!
This is very helpful, since I’m updating a lecture on the resource use of AI for an AI literacy class I’m teaching right now!
A few minor points - in my academic field (philosophy) there are some people who are complaining about the carbon emissions associated with conference travel, and arguing that we should have more online conferences and fewer in-person conferences. (See, eg, the comments here: https://dailynous.com/2026/01/15/apa-to-end-experiment-with-online-divisional-meetings/ ) I expect there are some people making similar arguments on computer science.
One thing that this really drives home for me is that there’s a difference between the general resource use of AI and specifically the electricity use. Unlike legos and CDs, which involve moving physical objects and making things out of refined petroleum, AI has its emissions almost entirely through electricity. It looks like there are individual data centers that use as much electricity as the entire city of San Diego! (For instance, the Amazon datacenter for Claude training in New Carlisle, IN: https://epoch.ai/data/data-centers/ .) But the carbon emissions and water use of this datacenter are more comparable to the emissions and water use of South Bend, or perhaps even New Carlisle. The water use change could be a significant effect locally, but it’s nowhere near the potential disruption of having to power a new San Diego in the middle of Indiana!
The issues of water and emissions will be handled by the same processes that are handling those issues for everything else in the world. But electricity use will be a significantly different type of issue to deal with, especially as we also electrify transportation, cooking, and climate control.
Arguing that we should have conferences online instead of in person seems like a tacit acknowledgement that data centers don’t require a lot energy relative to transportation!
Great research Andy! Thanks for sharing this. It has been really eye opening. I have a couple of doubts. The first one is why you didn't mention the paper "How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference" by Jegham, N., Abdelatti, M., Elmoubarki, L., & Hendawi, A. title (https://arxiv.org/abs/2505.09598).I personally useded it to create a calculator for the cost of personal queries with this models (spolier alert: it was a lot less than i thought it would be). And the second one is if you got the chance to look at this news report form Business Insider about how tricky it is to find public records about energy consumption of theses companies (https://www.youtube.com/watch?v=t-8TDOFqkQA&t=1548s). Would love your tohughts on that.
Hi Andy! Another excellent post, thank you. At least in my experience the level of impact you're having on people's thinking on this topic is very large, more than enough to justify the level of effort you're putting in! Obviously that's easy for me to say since *I'm* not the one putting in that effort, but...
I've been asked this by people with whom I've shared your posts, so just to get it on record: what if any financial incentives do you have to say that AI is good, and/or that AI is not causing environmental harm?
Happy to answer any questions. I’m not being paid by anyone to say that AI is good/is not causing environmental harm, I don’t take money from AI or data center companies etc. I’m also not invested in AI companies specifically which sometimes comes up. I guess I have some incentive because of audience capture etc. but I think I resist that. I sometimes take small speaking fees for small professional or educational gatherings.
The couple of times it's come up I haven't tried to drill deeper since I didn't have any answers for them other than to point to a history of involvement in EA, but I interpreted them as meaning 'has significant investments in AI companies' (anyone with an index fund is somewhat invested in AI at this point, but presumably that's excluded). Anyone who thinks you're literally a paid shill probably won't be much moved by you denying it 😂
Thanks!
This brings to mind the rant I always go off on when folks here in California get all outraged at how much water it requires to grow an almond tree. X # of people use the same amount of water etc. Maybe it is a poor use of the water, or maybe it isn't.
If we use that water instead for something else, how do you compare the costs and benefits? Gallons/calorie, or grams of fiber or fat, or some other nutrient? Gallons/$ of profit, or $ of income taxes to the state , etc etc. Drives me nuts.
Are we all better off if we stop growing almonds entirely (given that we produce pretty much the world's supply of a delicious and nutitious commodity)? I dunno, maybe, but I've sure never seen the recipts showing that to be the case.
You are doing fantastic work, sir. Thank you.
Really solid analysis. Training is a fixation for journalists because the boundary between training and inference is blurry or nonexistent for most people. I'd be curious to see what the numbers are for daily electricity use by a widely-used frontier model. I imagine inference is where the bulk of the energy demand actually originates.
Great work as always. My gut reaction to hearing any of these comparisons for the first time has always been: "But... that's _tiny_!"
And speaking of GTA V, I also have to wonder about the emissions of hundreds of employees driving to work for six days a week, for five long years, all using high-end workstations and servers...
There's a talking point that surfaced over the past few months, where people claim that the rapid iteration of new models is making the training energy consumption that much worse (and the profitability of any models unlikely), e.g. "GPT-5.0 was followed by GPT-5.1 just six weeks later!" Of course, the various versions of GPT-5.x don't involve the same massive pretraining of the entire base model, as they're mostly fine-tunes, merges, RLHF/SFT and whatever other arcane arts the AI labs employ.
This is very interesting, thanks for sharing this! Do you have any data on running these models though? What is the operational carbon footprint?
I’m on the author’s side overall, but I think it’s important to point out that we keep creating new energy- and resource-intensive products and services. Sure, data centers are no worse than producing Coke or developing Grand Theft Auto, but we aren’t swapping one for another, we’re adding more. So I think when people hear how demanding a data center is, they’re adding it to all the endeavors they know are already consuming vast amounts of energy and resources.
Fair enough, but what about the carbon footprint of inference? You’re right that the carbon footprint of training a single model is amortized over millions of users, but inference costs scale linearly with usage, and over time will be much larger than the training footprint.
Have you seen convincing estimates of the carbon footprint accounting for inference costs?
"where the regulatory and investment frameworks are favourable" = where the governing bodies don't mind exploiting and selling out their constituents who live nearby
Sorry this quote doesn’t appear in my post?
Oh I see it's in the comments. How do you feel about pharmaceutical plants or car factories being allowed to be built?