The Cloud Is Drinking the River
The Hidden Water and Power Bill of Always-On AI, and the Towns Paying It So You Can Have an Answer
Every prompt has a body. It runs on electricity, and the machine that answers it has to be kept cool, which in much of the world means water. At the scale of billions of queries a day, that physical cost stops being a rounding error and turns into a line item on a utility bill somewhere. The catch is that the bill rarely arrives at the address of the person who asked the question.
The cloud was sold to us as weightless, and for most of the internet era the abstraction held. It is in fact one of the heaviest, thirstiest machines ever built, and its weight is being set down quietly on a handful of specific places: their grids, their aquifers, their power rates. The people paying for the answer are mostly not the people asking the question.
Global data-centre electricity is set to more than double
Terawatt-hours per year (IEA)
Published by Kymata Labs · Independent Research Institution.
You asked the cloud a question this morning. Where did the water go?
You typed something into a chatbot before coffee, and the answer came back in a second from nowhere you could point to. It came from a building. Somewhere there is a windowless hall of servers running hot, kept from cooking itself by air conditioning and, very often, by evaporating fresh water into the sky. That building sits in a real county, draws from a real grid, and in many cases pulls from a municipal water supply that somebody else also drinks.
If you have never once wondered whereyour prompts are answered, that is rather the point of this paper. The geography is hidden by design, and the word “cloud” does a great deal of work to keep it that way. The cost underneath the word is neither abstract nor evenly shared. A typical AI data centre, the IEA notes, draws as much power as 100,000 households,1 and that load has to land somewhere. It is already landing on towns most of the people using AI have never heard of.
Ask where the water went, and the cloud turns out to have a postcode.
“Somewhere a town is handing over its river so the rest of us can have an answer in a second.”
Kymata LabsThe numbers are large, and the honest ones are more interesting than the scary ones.
It would be easy to write this paper as a horror story of apocalyptic grids and runaway aggregate figures. We have chosen not to, because the real argument is sharper than the frightening one and it survives the caveats. The authoritative numbers come from the International Energy Agency, and where a figure is contested we will say so on the page. What both kinds of number describe is not AI swallowing the planet’s power but a small number of places quietly absorbing a load the rest of us generated.
The authoritative block: 415 terawatt-hours, and set to double
In its 2025 “Energy and AI” report, the IEA put global data-centre electricity at roughly 415 TWh in 2024, about 1.5% of the world’s electricity, and projected it to more than double to around 945 TWh by 2030, a figure comparable to the entire electricity consumption of Japan today. The United States alone accounted for 45% of global data-centre electricity in 2024. In advanced economies, data centres are set to drive more than 20% of electricity-demand growth to 2030, nearly half of it in the US.
IEA, “Energy and AI,” 2025.The honesty anchor: this is concentration, not domination
Here is the figure that makes the paper credible rather than alarmist. The same IEA report notes that data centres account for only about 10% of global electricity-demand growth to 2030. At the planetary level the larger pressures on the grid are electrification, industry, and cooling, not AI. The thesis of this paper is therefore not aggregate dominance but local concentration plus socialised cost, and the proof of it is geographic: nearly half of US data-centre capacity sits in just five regional clusters. A gentle global average can hide a brutal local one, and in those clusters it does.
IEA, “Energy and AI,” 2025. The figure that strengthens the case.What the companies say: one fifteenth of a teaspoon
In June 2025, OpenAI’s Sam Altman offered his own numbers. An average ChatGPT query, he wrote, uses about 0.34 watt-hours of electricity and roughly 0.000085 gallons of water, “about one fifteenth of a teaspoon.” The figure is tiny and reassuring, and the part that matters most is that it was published with no methodology. This is the company’s own framing of its own footprint, with no way for anyone outside to check it, and we cite it for exactly that reason: to hold it up against the independent work, which is messier and far less flattering.
Altman, “The Gentle Singularity,” June 10, 2025. OpenAI’s own figure, unmethodologised.The water figure: real, but handle with care
The most-cited independent estimate comes from UC Riverside’s 2023 study, “Making AI Less Thirsty.” It found roughly 500 mL, about 17 ounces, of fresh water per 20 to 50 ChatGPT queries, counting both the water evaporated to cool the data centre and the water consumed at the power plants feeding it, and it estimated that training GPT-3 alone consumed on the order of 700,000 litres of freshwater. The qualifiers deserve as much attention as the headline number. That is a GPT-3-era estimate, per 20–50 queries, including indirect power-plant water, not 500 mL for a single query. Newer and more efficient models almost certainly draw less, and some critics put the figure nearer 5 mL per query. What is in dispute is the magnitude; what is not in dispute is that the water cost is real.
Li, Ren et al., “Making AI Less Thirsty,” 2023. A contested, GPT-3-era estimate.The energy per query: a moving target
How much electricity does one query take? The answer keeps shrinking. EPRI’s 2024 analysis put a Google search at about 0.3 Wh and a ChatGPT query at about 2.9 Wh, roughly ten times more. Epoch AI, working a year later, estimates a typical modern GPT-4o query at around 0.3 Wh, about ten times lower than the older figure, as the models grew more efficient. The sensible way to hold per-query energy is as a range that is still in motion, because anyone who quotes a single fixed figure has frozen a number that is still falling.
EPRI, 2024; Epoch AI, 2025. Present as a range, not a point.
A moving target: the per-query cost keeps shrinking
Electricity per query, watt-hours
Two towns. Two rivers. The same machine.
Aggregate statistics are easy to ignore in a way that towns are not. The clearest way to understand what “socialised cost” means is to visit the places where it has already happened, where a named utility logged a real draw and a named company went to some length to keep the number quiet.
What one data centre takes from one town’s water
Share of a community’s water drawn by a single facility
West Des Moines, Iowa.In July 2022, the month before OpenAI finished training GPT-4 in Microsoft’s Iowa data-centre cluster, Microsoft pumped roughly 11.5 million gallons of water into that cluster. According to the local water utility and the Associated Press, that came to about 6% of the district’s water for the month. One company drew six percent of a community’s monthly water to cool the machine that trained the model now answering your prompts.6
The Dalles, Oregon.Here the story is as much about secrecy as it is about water. Google fought in court to keep its data centres’ water use classified as a trade secret, and when the figure was finally disclosed it showed the company had used about 274.5 million gallons in 2021, roughly a quarter of all the water used by the town. The community hosting the facility had to sue to learn how much of its own river the cloud was drinking.7
The draw is also rising. In their own disclosures, Google reported water use up about 20% to 5.6 billion gallons in 2022, and Microsoft reported a 34% rise from 2021 to 2022.7 Those are the years AI scaled, and the water curve climbed alongside it.
A quarter of a town’s water, and the town had to go to court to find out.
We called it “the cloud” so we’d stop asking where it was.
The metaphor did its job. “The cloud” took a continent of concrete, copper, and coolant and turned it into something that sounds like weather, at once everywhere and nowhere and weightless and free. For most of the internet era that fiction was harmless, because the footprint was small enough that the abstraction cost nothing.
AI changed the arithmetic. Training a frontier model and then serving it to hundreds of millions of people is a far more physical act than serving a web page. The IEA’s projected doubling of data-centre electricity by 2030 is the moment the abstraction becomes visible, the weather turning out to have a water bill. Because the load concentrates, it becomes visible unevenly, surfacing first in the five US clusters that hold nearly half the capacity and in the counties where the next gigawatt of demand is being sited right now.
The deeper trap is that efficiency does not rescue us; it accelerates us. Per-query energy keeps falling while total demand keeps climbing anyway, because each cheaper query invites a great many more. Microsoft’s own CEO, Satya Nadella, named the mechanism in January 2025 by invoking the Jevons paradox: “As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.”8 Every efficiency gain is reinvested as more usage, and the footprint grows on the back of the very improvements meant to shrink it.
Each gain in efficiency is spent on more demand, not less.
The same machine, read by three different readers.
This is not an argument against AI, nor a guilt trip about asking a chatbot a question. It is an argument about who pays and whether they got any say in the matter, and what to do about it depends a good deal on who you are.
For individuals
Stop treating the cloud as weightless, and then put your attention where it actually counts.
- Recognise that the honest leverage here is collective rather than personal, since skipping a prompt does nothing to a county’s water table.
- When a data centre is proposed near you, ask the questions that decide the deal: whose grid, whose water, whose rates.
- Reward disclosure. A company that publishes its real consumption figures has not earned the same treatment as one that litigates to hide them.
For employers
If you build on AI, part of the footprint is yours, so account for it honestly.
- Choose providers and regions on water stress and grid mix, not on price and latency alone, because the same query costs very different things in different places.
- Measure and report your AI’s real consumption rather than repeating a vendor’s un-methodologised teaspoon figure.
- Plan for the rebound. As your AI gets cheaper per query your total usage will climb, so budget the footprint for where demand is heading rather than where it sits today.
For policymakers
This is a problem of siting, disclosure, and cost allocation, and those levers are yours.
- Mandate water and energy disclosure before a facility is approved. The Dalles had to sue for its number, and no community should have to.
- Decide explicitly who pays for the grid upgrades new data-centre load requires, the operator or, by default, every other ratepayer.
- Govern siting by local water stress rather than by available land and tax incentives alone. The cost concentrates in five clusters, and the scrutiny should concentrate there too.
FAQ
Both claims are wrong, and the honest answer sits between them. At the global level the IEA is blunt: data centres were only about 1.5% of the world's electricity in 2024, and they account for roughly 10% of the growth in global electricity demand to 2030, which is not the profile of a dominant driver. The aggregate average is where the real story hides, because the load is so concentrated. The IEA notes that nearly half of US data-centre capacity sits in just five regional clusters, and that in advanced economies data centres drive more than 20% of electricity-demand growth to 2030, nearly half of it in the US. The grid is not melting everywhere. It is straining in particular places, and those places carry the cost for everyone else.
Nobody can give you a clean number, and you should distrust anyone who claims to. OpenAI's own figure is roughly 0.000085 gallons, about one fifteenth of a teaspoon, but it arrives with no published methodology. The most-cited independent estimate, from UC Riverside in 2023, works out to roughly 500 mL of freshwater per 20–50 queries, and even that figure carries heavy qualifiers: it is GPT-3-era, it counts the water evaporated at the power plants supplying the electricity, and newer, more efficient models almost certainly use less, with some critics putting it closer to 5 mL per query. The exact number is less important than what every version of it confirms, which is that a single query has a real physical water cost, and at billions of queries a day that cost lands somewhere specific.
It is working now, but it had to be dragged into the light first. In The Dalles, Oregon, Google argued in court that its data centres' water consumption was a trade secret, and only after a legal fight did the number come out: about 274.5 million gallons in 2021, roughly a quarter of the whole town's water. The disclosures that followed are genuine progress, yet the starting posture was secrecy, and the communities hosting these facilities often learned the scale of the draw only after the contracts had been signed.
That is the comforting assumption, and the evidence cuts against it. Per-query energy has indeed fallen, with Epoch AI estimating a typical GPT-4o query at around 0.3 Wh, roughly ten times lower than older numbers. The trouble is that efficiency has not historically shrunk total demand; it has grown it. Microsoft's own CEO invoked the Jevons paradox in January 2025, saying that as AI gets cheaper and more accessible its use 'skyrockets' into a commodity 'we just can't get enough of.' When each query gets cheaper, we simply run far more of them. The IEA's projected doubling of data-centre electricity by 2030 already bakes in the efficiency gains, and the footprint grows regardless.
Honestly, very little moves the needle at the individual level, and pretending otherwise would be its own kind of greenwashing. This is not a recycling problem you solve at the household sink. The leverage is collective, and it sits with how utilities approve new data-centre load, who pays for the grid upgrades that load requires, what water a facility is permitted to draw in a drought-prone county, and whether companies must disclose consumption before the deal is done rather than after a lawsuit. The one genuinely useful individual act is to stop treating the cloud as weightless, so that when a data centre comes to your county you think to ask exactly whose power and whose water it plans to use.
The answer came to you free. Somewhere, it was paid for.
None of this is a case for switching off the machine. AI is useful, the efficiency gains are real, and at the global scale it is not the thing breaking the grid. The global scale is simply the wrong place to look. The story lives in the five clusters, and in the county that gave up a quarter of its water and then had to sue to learn it had. The cost of the cloud is real and physical and it has a street address, and the people at that address mostly never got to vote on the question you just asked.
Next time the cloud answers, ask whose river it drank.
Sources
Every figure in this paper is drawn from the primary sources below. Where a figure is contested, estimated, or company-supplied without methodology, we have said so in the text, because on this subject the authority lives in the caveats.
- International Energy Agency (2025). "Energy and AI": Executive Summary. (Data centres ~415 TWh / ~1.5% of global electricity in 2024; projected to more than double to ~945 TWh by 2030; US = 45% of global data-centre electricity in 2024; data centres ~10% of global electricity-demand growth.) https://www.iea.org/reports/energy-and-ai/executive-summary
- Altman, S. (2025). "The Gentle Singularity." (OpenAI's own framing: average ChatGPT query ~0.34 Wh and ~0.000085 gallons of water, no published methodology.) https://blog.samaltman.com/the-gentle-singularity
- Li, P., Yang, J., Islam, M. A. & Ren, S. (2023). "Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models." arXiv:2304.03271. (GPT-3-era estimate: ~500 mL per 20–50 queries, including indirect power-plant water; training GPT-3 ~700,000 litres of freshwater.) https://arxiv.org/abs/2304.03271
- Electric Power Research Institute (2024). "Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption." (~0.3 Wh per Google search vs ~2.9 Wh per ChatGPT query.) https://www.epri.com/research/products/3002028905
- Epoch AI (2025). "How much energy does ChatGPT use?" (A typical GPT-4o query likely ~0.3 Wh, roughly 10× lower than older estimates, due to efficiency gains.) https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
- O'Brien, M. & Fingerhut, H. (2023). "Artificial intelligence technology behind ChatGPT was built in Iowa, with a lot of water." Associated Press. (West Des Moines: Microsoft pumped ~11.5M gallons into its Iowa cluster in July 2022, ~6% of district water that month.) https://apnews.com/article/chatgpt-gpt4-iowa-ai-water-consumption-microsoft-f551fde98083d17a7e8d904f8be822c4
- Solomon, C. (2023). "Data centers and their impact on Oregon's water and power." Reynolds Center / The Oregonian. (The Dalles: Google fought to keep its water use secret; disclosed ~274.5M gallons in 2021, ~a quarter of the town's water. Corporate disclosures: Google's water use rose ~20% to 5.6B gallons in 2022; Microsoft's rose 34% from 2021 to 2022.) https://businessjournalism.org/2023/11/oregonian-data-centers/
- Kurtz, A. & Selyukh, A. (2025). "DeepSeek, AI and the Jevons paradox, explained." NPR Planet Money. (Microsoft CEO Satya Nadella, Jan 27, 2025, on AI and the Jevons paradox.) https://www.npr.org/sections/planet-money/2025/02/04/g-s1-46018/ai-deepseek-economics-jevons-paradox
Read it as a PDF
The complete argument, the cross-domain evidence, and all primary sources, typeset as a white paper. Free, no signup.
↓ Download the PDF