Thanks for writing this. In addition check out this article.
“My overall takeaway is that you need to be careful when talking about water use: it’s very easy to take figures out of context, or make misleading comparisons. Very often I see alarmist discussions about water use that don’t take into account the distinction between consumptive and non-consumptive uses. Billions of gallons of water a day are “used” by thermal power plants, but the water is quickly returned to where it came from, so this use isn’t reducing the supply of available fresh water in any meaningful sense. Similarly, the millions of gallons of water a day a large data center can use sounds like a lot when compared to the hundreds of gallons a typical home uses, but compared to other large-scale industrial or agricultural uses of water, it's a mere drop in the bucket.”
The article also quotes Alex de Vries-Gao, who has apparently become the go-to source for these kinds of articles. He is not an expert on either AI or IT in general, and he was last in the news with this article that was repeated everywhere:
There is so much wrong with this study, I don't even know where to begin. It essentially assumes that every GPU manufactured is used exclusively for AI training (no consumer use, no non-AI data center use, no inference or inference-specific hardware) and runs at its theoretical max TDP forever. It conflates manufacturing energy with operational energy instead of amortizing it over the hardware's lifespan. It confuses instantaneous supply bottlenecks with long-term deployment.
I did a somewhat naive calculation to correct for this and found that the study overstates AI power consumption for 2025 by about 3x, and by at least 2x. Everyone in AI scoffed at it, but its narrative wasn't countered in major publications, and it passes for expert wisdom now.
Thanks for this and I greatly appreciate your call for clarification and the avoidance of hype.
Having studied water usage and data centers for around two years (and I recommend the surveys from The Uptime Institute, a global data center trade association) I believe there’s also the risk of not discussing or framing the entire system of water and energy in a particular region.
Meaning, on-site and offsite water use and availability is highly relevant for a hyperscale data center in terms of water treatment plants and where and how a specific center gets its energy.
It is laudable to try and maximize efficiencies of existing or future centers. It is questionable if not illogical and unsustainable to build hyperscale data centers in water scarce regions of these use (as reported for Bessemer, AL) upwards of 2M gallons of water a day. Due to their size, normal dewatering techniques (which happens for any large structures) can damage watersheds and local ecosystems for many years. And aquifer depletion and restoration don’t happen at even / regular intervals in water stressed areas even before any large / new drain.
Here I will say when comparing water use of GenAI at hyperscale data centers to other use of water there is a similar danger to misrepresent or distract as part of the larger paradigm of systemic issue of hyperscale datacenter proliferation.
First is discussing water use by agriculture or consumers for ongoing basic needs. Yes, water use by the beef industry is massive and another area of concern for aquifer depletion. So one recommendation among others is to consider eating less beef while working to determine how to make sure cattle farmers and the dairy industry could be supported in regions where their not having water becomes a business and societal issue, not just an environmental one.
My point here is that water is a core issue for life in relation to the need for food where GenAI is not. Or for a citizen having water for cooking and sanitation purposes along with needing it to drink.
So comparing water uses perhaps would be more compelling in terms of X tech uses Y water. I’ve seen this noted with blockchain, for instance.
The larger issue, however, is the vast proliferation of hyperscale data centers and the massive financial support for their growth, often times in water scarce regions. And oftentimes where local communities have voted against their approvals (in Detroit, in Bessemer, in multiple parts of Atlanta) and the projects move forward typically without any or all water and energy data revealed.
That is inclusive of local water sources and availability, issues related to crisis energy allocation (where reports have shown hyperscale data centers are prioritized to maintain energy usage over other community needs) etc.
What is also happening is that many communities are having their energy or utility bills increase due to hyperscale data centers that don’t release local water and energy data, claiming its under NDA.
So where the economic and larger holistic paradigms happening right now, to the point citizens protesting data center proliferation has become a bi-partisan issue in the US, the systemic aspects of any GenAI prompt have a lot more weight and framing to it than each individual prompt’s use.
The prompts don’t live in isolation.
I say all this because the often techno-solutionist answer to any of these issues is to use nuclear power, or quantum, etc. I look forward to the ongoing research in these domains.
But right now today, based on hyperscale data center increase in multiple global water scarce regions, any prompt exists in the larger socioeconomic paradigm mentioned here. And of course water scarcity exists in a larger set of systems in the biosphere. So there are escalating effects when all these structures drain water at such alarming and rapid rates.
Claiming nuclear might help or relating ways an org may be water secure by 2030 also ignores these larger systemic and economic issues.
Thank you again for your research and insights. I hope any of this is additive and helpful and I’m happy to send on research etc based on your thoughts.
And I really appreciate the call to avoid simplistic or non-holistic framings.
I saw a post on LinkedIn by Kate Brandt, Chief Sustainability Officer at Google, celebrating these key findings about water and energy use on per-prompt (text) basis. And I've been following you and Hannah Ritchie discussing this, reading MIT, Rocky Mountain Institute, etc. etc. I've been feeling better about per-prompt usage, especially in helping students in a climate studies M.S. program use AI enough to learn how to recognize it's flaws. - However, I'm evolving on this. This past weekend, I had breakfast with a Senator. They were alarmed by the data center expansion that is happening, and therefore pivoting toward supporting nuclear (SRMs). It dawned on me during that conversation, that I had fallen for the same framing trap that we all did back when BP proposed "carbon footprints" in the mid-2000's. (Probably before your time). The campaign encouraged people to measure their own emissions from daily activities (driving, heating, flying, etc.) and reduce them. This shifted attention toward individual responsibility for climate change rather than systemic responsibility by fossil fuel producers. It's corporate greenwashing. And I worry that those of us that are using our time on calculating per-prompt emissions and water drops, are not seeing the big picture.
It seems to me the better question isn’t “how many drops of water did my prompt take?” but “what choices are the system builders making?” Data centers aren’t multiplying because a few students run queries. They’re multiplying because developers decide model size, training frequency, and cooling methods. Because investors push for scale. Because the system is set up to reward bigger, faster, more—without asking if it’s wiser.
When we focus on per-prompt numbers, we’re repeating the BP trick. They told us to change our lightbulbs while they kept drilling. Now tech companies are telling us to count our prompts while they choose how to power and cool their server farms.
The real accountability belongs with them:
– Are they designing models for efficiency, or just dominance?
– Are they sourcing energy from renewables, or from coal-heavy grids?
– Are they using water wisely, or draining rivers in drought zones?
– Are they being transparent, or pushing the math onto end-users?
That’s the frame I want to see. Not guilt over one more question to an AI, but pressure on the people making the structural choices.
Thanks for helping me think through this issue. I'm ready to address it with my students today.
I assume this is rhetorical. It is to get clicks from the Doomers. You and Hannah don't feed their dogma, so you'll get squat for traffic. (Except for the fanatics yelling at Hannah.)
A lot of people have preexisting objections to AI use (copyright, diminishing human thought, association with tech bros), and because of those biases they easily fall for bogus environmental claims about AI. I think we should just discuss those objections instead of this dumb water argument.
Are we talking about a full life cycle view of water use by an exceptionally large AI data centre (DC) or just the instantaneous use by one consumer type search? Anyone with fab experience knows how much DI water is used to clean the chips between machine cycles and for humidity control- I assume all this consumption is being ignored.
My experience with certain mid-size older Google type DCs was cooling of server floors with water chillers, cooling towers with evaporative losses and blow down water discharges plus humidification for the server floors.
The new DCs are significantly greater in scope as they are using server water cooling of much higher operating temp video chips. The cooling load is significantly higher in total and you just can't play around with the denominator of a much higher capacity chip to get a lower water use per search.
You need to do a DC operating cycle view of water use and not play around with the numerator/denominator numbers.
It's better to ask about how the DC equates to municipal water use per person for the total water consumption of the site.
You will then see how a new AI DC compares to what you and I would be as water intensive users for the DC hosting location being impacted.
I doubt if any local impact assessments have been done for these new mega-DCs.
I agree that this is an important thing to clarify; But is the median prompt a relevant measure ? Is AI use similar to social media that has a steady increase or is it more like alcohol where a few whales who are around 10% of users account for the most spend (Video generation, coding loops & deep research agents as opposed to pasta recipes); if it's the latter then the calculations by Google are helpful, but not relevant.
Thanks for writing this. In addition check out this article.
“My overall takeaway is that you need to be careful when talking about water use: it’s very easy to take figures out of context, or make misleading comparisons. Very often I see alarmist discussions about water use that don’t take into account the distinction between consumptive and non-consumptive uses. Billions of gallons of water a day are “used” by thermal power plants, but the water is quickly returned to where it came from, so this use isn’t reducing the supply of available fresh water in any meaningful sense. Similarly, the millions of gallons of water a day a large data center can use sounds like a lot when compared to the hundreds of gallons a typical home uses, but compared to other large-scale industrial or agricultural uses of water, it's a mere drop in the bucket.”
https://www.construction-physics.com/p/how-does-the-us-use-water
The article also quotes Alex de Vries-Gao, who has apparently become the go-to source for these kinds of articles. He is not an expert on either AI or IT in general, and he was last in the news with this article that was repeated everywhere:
https://vu.nl/en/news/2025/ai-rapidly-on-its-way-to-becoming-the-largest-energy-consumer
There is so much wrong with this study, I don't even know where to begin. It essentially assumes that every GPU manufactured is used exclusively for AI training (no consumer use, no non-AI data center use, no inference or inference-specific hardware) and runs at its theoretical max TDP forever. It conflates manufacturing energy with operational energy instead of amortizing it over the hardware's lifespan. It confuses instantaneous supply bottlenecks with long-term deployment.
I did a somewhat naive calculation to correct for this and found that the study overstates AI power consumption for 2025 by about 3x, and by at least 2x. Everyone in AI scoffed at it, but its narrative wasn't countered in major publications, and it passes for expert wisdom now.
You should write about it if you haven’t!
I know, I keep meaning to get round to it - maybe I'll use this latest article as an excuse!
Please do! I'll share it once you write it
Thanks for this and I greatly appreciate your call for clarification and the avoidance of hype.
Having studied water usage and data centers for around two years (and I recommend the surveys from The Uptime Institute, a global data center trade association) I believe there’s also the risk of not discussing or framing the entire system of water and energy in a particular region.
Meaning, on-site and offsite water use and availability is highly relevant for a hyperscale data center in terms of water treatment plants and where and how a specific center gets its energy.
It is laudable to try and maximize efficiencies of existing or future centers. It is questionable if not illogical and unsustainable to build hyperscale data centers in water scarce regions of these use (as reported for Bessemer, AL) upwards of 2M gallons of water a day. Due to their size, normal dewatering techniques (which happens for any large structures) can damage watersheds and local ecosystems for many years. And aquifer depletion and restoration don’t happen at even / regular intervals in water stressed areas even before any large / new drain.
Here I will say when comparing water use of GenAI at hyperscale data centers to other use of water there is a similar danger to misrepresent or distract as part of the larger paradigm of systemic issue of hyperscale datacenter proliferation.
First is discussing water use by agriculture or consumers for ongoing basic needs. Yes, water use by the beef industry is massive and another area of concern for aquifer depletion. So one recommendation among others is to consider eating less beef while working to determine how to make sure cattle farmers and the dairy industry could be supported in regions where their not having water becomes a business and societal issue, not just an environmental one.
My point here is that water is a core issue for life in relation to the need for food where GenAI is not. Or for a citizen having water for cooking and sanitation purposes along with needing it to drink.
So comparing water uses perhaps would be more compelling in terms of X tech uses Y water. I’ve seen this noted with blockchain, for instance.
The larger issue, however, is the vast proliferation of hyperscale data centers and the massive financial support for their growth, often times in water scarce regions. And oftentimes where local communities have voted against their approvals (in Detroit, in Bessemer, in multiple parts of Atlanta) and the projects move forward typically without any or all water and energy data revealed.
That is inclusive of local water sources and availability, issues related to crisis energy allocation (where reports have shown hyperscale data centers are prioritized to maintain energy usage over other community needs) etc.
What is also happening is that many communities are having their energy or utility bills increase due to hyperscale data centers that don’t release local water and energy data, claiming its under NDA.
So where the economic and larger holistic paradigms happening right now, to the point citizens protesting data center proliferation has become a bi-partisan issue in the US, the systemic aspects of any GenAI prompt have a lot more weight and framing to it than each individual prompt’s use.
The prompts don’t live in isolation.
I say all this because the often techno-solutionist answer to any of these issues is to use nuclear power, or quantum, etc. I look forward to the ongoing research in these domains.
But right now today, based on hyperscale data center increase in multiple global water scarce regions, any prompt exists in the larger socioeconomic paradigm mentioned here. And of course water scarcity exists in a larger set of systems in the biosphere. So there are escalating effects when all these structures drain water at such alarming and rapid rates.
Claiming nuclear might help or relating ways an org may be water secure by 2030 also ignores these larger systemic and economic issues.
Thank you again for your research and insights. I hope any of this is additive and helpful and I’m happy to send on research etc based on your thoughts.
And I really appreciate the call to avoid simplistic or non-holistic framings.
I saw a post on LinkedIn by Kate Brandt, Chief Sustainability Officer at Google, celebrating these key findings about water and energy use on per-prompt (text) basis. And I've been following you and Hannah Ritchie discussing this, reading MIT, Rocky Mountain Institute, etc. etc. I've been feeling better about per-prompt usage, especially in helping students in a climate studies M.S. program use AI enough to learn how to recognize it's flaws. - However, I'm evolving on this. This past weekend, I had breakfast with a Senator. They were alarmed by the data center expansion that is happening, and therefore pivoting toward supporting nuclear (SRMs). It dawned on me during that conversation, that I had fallen for the same framing trap that we all did back when BP proposed "carbon footprints" in the mid-2000's. (Probably before your time). The campaign encouraged people to measure their own emissions from daily activities (driving, heating, flying, etc.) and reduce them. This shifted attention toward individual responsibility for climate change rather than systemic responsibility by fossil fuel producers. It's corporate greenwashing. And I worry that those of us that are using our time on calculating per-prompt emissions and water drops, are not seeing the big picture.
It seems to me the better question isn’t “how many drops of water did my prompt take?” but “what choices are the system builders making?” Data centers aren’t multiplying because a few students run queries. They’re multiplying because developers decide model size, training frequency, and cooling methods. Because investors push for scale. Because the system is set up to reward bigger, faster, more—without asking if it’s wiser.
When we focus on per-prompt numbers, we’re repeating the BP trick. They told us to change our lightbulbs while they kept drilling. Now tech companies are telling us to count our prompts while they choose how to power and cool their server farms.
The real accountability belongs with them:
– Are they designing models for efficiency, or just dominance?
– Are they sourcing energy from renewables, or from coal-heavy grids?
– Are they using water wisely, or draining rivers in drought zones?
– Are they being transparent, or pushing the math onto end-users?
That’s the frame I want to see. Not guilt over one more question to an AI, but pressure on the people making the structural choices.
Thanks for helping me think through this issue. I'm ready to address it with my students today.
Thanks, Andy.
>Why do this?
I assume this is rhetorical. It is to get clicks from the Doomers. You and Hannah don't feed their dogma, so you'll get squat for traffic. (Except for the fanatics yelling at Hannah.)
A lot of people have preexisting objections to AI use (copyright, diminishing human thought, association with tech bros), and because of those biases they easily fall for bogus environmental claims about AI. I think we should just discuss those objections instead of this dumb water argument.
Are we talking about a full life cycle view of water use by an exceptionally large AI data centre (DC) or just the instantaneous use by one consumer type search? Anyone with fab experience knows how much DI water is used to clean the chips between machine cycles and for humidity control- I assume all this consumption is being ignored.
My experience with certain mid-size older Google type DCs was cooling of server floors with water chillers, cooling towers with evaporative losses and blow down water discharges plus humidification for the server floors.
The new DCs are significantly greater in scope as they are using server water cooling of much higher operating temp video chips. The cooling load is significantly higher in total and you just can't play around with the denominator of a much higher capacity chip to get a lower water use per search.
You need to do a DC operating cycle view of water use and not play around with the numerator/denominator numbers.
It's better to ask about how the DC equates to municipal water use per person for the total water consumption of the site.
You will then see how a new AI DC compares to what you and I would be as water intensive users for the DC hosting location being impacted.
I doubt if any local impact assessments have been done for these new mega-DCs.
I agree that this is an important thing to clarify; But is the median prompt a relevant measure ? Is AI use similar to social media that has a steady increase or is it more like alcohol where a few whales who are around 10% of users account for the most spend (Video generation, coding loops & deep research agents as opposed to pasta recipes); if it's the latter then the calculations by Google are helpful, but not relevant.