Everyone Is Arguing About AI Water Use. Both Sides Are Measuring Something Different.

The water fight over AI is the rare genuinely two-sided technical argument where both camps are measuring a different quantity and both numbers are correct. The UN University says data centres used 448 TWh and 4.5 trillion litres of water in 2025, heading to 945 TWh and 9.3 trillion litres by 2030. Amazon says its 2.5 billion gallons is trivial against 117 trillion gallons of US water withdrawal. Sam Altman says the per-query crisis numbers are "totally fake, insane, and has no connection to reality." UC Riverside says roughly 519ml per 100-word prompt. The Department of Energy's op-ed says closed-loop cooling makes data centres among the most water-efficient facilities there are.

Every one of those is a real number inside a real boundary. This post is about the boundaries, what happens when you change them, and the fact that the genuinely dangerous quantity, local basin stress, is the one nobody publishes at the granularity that would matter.

The two camps, stated at their strongest#

The UN University Institute for Water, Environment and Health published its report on 3 June 2026, led by Kaveh Madani, a Stockholm Water Prize winner. The numbers: 448 TWh of electricity in 2025, more than all but 10 countries, rising to 945 TWh by 2030, which would make data centres the 6th-highest electricity consumer on earth, roughly Japan's entire national use. Water goes from 4.5 trillion litres to 9.3 trillion. CO2 goes from 189 million metric tonnes, about Argentina's total, to 399 million. The land footprint doubles from 6,900 square kilometres to over 14,500. AI's share of data centre energy climbs from roughly 20% now to 40% by 2030.

Madani's three lines that matter: "The public debate still often treats AI as software, but AI is also physical infrastructure: data centres, electricity generation, cooling systems, transmission networks, chips, minerals, land and water." And: "AI will not simply 'run out' of water or electricity worldwide. But in specific places, poorly planned data centre expansion could collide with existing resource pressures. That is why responsible planning matters now, before infrastructure and dependencies become locked in."

And the sharpest line in the report is the renewable-energy critique, and Madani's actual wording is worth keeping intact because it is worse than the paraphrase: "While some companies tout the use of renewable energy for data centers, Madani said that means the supply of clean electricity is depleted and thus dirtier energy is used elsewhere." His own summary is blunter: "The cleanest form of AI use is no use." The report's coverage offers a concrete figure: a complex video generation is equivalent to 42 hours of a lightbulb burning plus a gallon of water.

Now the other side. Amazon's June 2026 blog puts its data centres at roughly 2.5 billion gallons of withdrawal globally in 2025. Ars Technica built the comparison set, and it is devastating to the alarmist framing: US total withdrawal in 2015 was 117 trillion gallons. Lawns and landscaping, 3.3 trillion. California almond orchards, 1.3 trillion. Golf courses, 531 billion. All data centres combined, 2.5 billion. Ars Technica's verdict was that in total water use, AI data centres are a drop in the bucket.

Read that number's boundary carefully, because in a post about boundaries it would be cheating to use it loosely. Bloomberg reported on 11 June 2026 that Amazon's 2.5 billion figure excludes the water used to generate the electricity the data centres draw, and that it counts water it discharges as well as water it consumes. So the most-cited number on the dismissive side is itself a scoped number, and its scope is favourable to the conclusion being drawn from it.

Company disclosures, each labelled by the year it covers. For 2023: Google roughly 6.4 billion gallons, Microsoft 1.7 billion, Meta 813 million. Google's own 2024 environmental report puts its total water consumption at 10.9 billion gallons, a jump large enough to reset the frame on its own. Meta's 2024 indirect figure is 19 billion. A 2021 Nature study put all US data centres combined at roughly 163 billion gallons including indirect consumption. Texas alone: an estimated 25 to 49 billion gallons in 2024, projected to hit 399 billion by 2030. In 2023, US data centres used just over 17 billion gallons for cooling, about half a million Americans' annual household use, and less than 0.1% of total US farm water use that year.

This camp is not lying. The arithmetic is the arithmetic. Every sector comparison points the same way.

The scope table that ends the argument#

Here is the actual technical content, the piece both camps skip. Three water scopes get routinely conflated, and they differ by orders of magnitude for the same physical system:

Read that second row again. At Meta, indirect use in 2024 was more than 20 times direct use. So "AI uses X water per query" is unanswerable until you know the grid mix at the data centre serving the query, and specifying that requires knowing which data centre served the query, which no public API on earth tells you. Eric Masanet of UC Santa Barbara, the person who studies data centres for a living, said it exactly: "We don't have the real ground-truth numbers on data-center water use to put everyone on the same page."

The third row is the one nobody outside the industry quotes, and it should be quoted more. A single microelectronics fab can consume up to 10 million gallons of ultrapure water a day. Ecolab puts the annual equivalent at the drinking water needs of 17 million people. That is upstream of any AI query, it happens whether or not the chip is ever used for inference, and it is not attributable to a prompt.

The Atlantic's July 2026 piece "The Truth About AI's Water Use" landed on the same verdict. Neither operators nor utilities disclose enough; the public information that exists can be "recontextualized and distorted to suggest basically anything." Water, they concluded, "seems like it should be measurable, with clear facts supporting one side or the other. That is not the case."

The per-query war is structurally unanswerable#

The two famous numbers: roughly 519ml per 100-word prompt from UC Riverside, as reported by the Washington Post, counting on-site cooling under a full evaporative estimate. About 0.3ml per average query from Sam Altman, who called the larger figure "totally fake, insane, and has no connection to reality."

Both are technically defensible. They measure different boundaries. That is the entire dispute.

The thing worth shouting: the per-query number is not merely contested, it is structurally unanswerable. Computing it requires knowing which data centre and which grid served the request, and no public API exposes that. This is a disclosure gap, not a measurement dispute, and it has a concrete fix. Shaolei Ren of UC Riverside, who has done the most-cited per-query work, is blunt about the degrees of freedom: "Can data centers use zero water? Yes, that's very easy." Air chillers require substantially more electricity than cooling towers to expel the same heat. Fengqi You of Cornell, energy systems, adds the geographic qualifier: outside water-stressed regions, "it's not really a major concern for now."

Now the part that cuts against my own framing of which number is the frightening one. The London Review of Books reported in September 2026 that Shaolei Ren, one of the authors of the Riverside paper behind the 519 millilitre figure, has since revised his own estimates and now finds the AI water footprint roughly four times larger than originally estimated. Google's own disclosure puts the median Gemini text prompt at 0.26 millilitres, much closer to Altman than to 519 millilitres, which means the headline number and the correction point in opposite directions and neither is the settled answer.

And Ren himself, on the record, said it in one sentence that happens to be the entire thesis of this post: "All of the numbers can be correct within their own context." That is the author of the most alarming per-query figure conceding that the fight is not about measurement. You cannot get a cleaner vindication of a boundary argument than the person whose number is being litigated agreeing that the boundary is the variable.

The labs per-query disclosure situation? Luccioni and Chowdhury put it on record: private AI labs disclose essentially nothing about per-query energy or water, so every per-query figure is an estimate derived from open models. Chowdhury: "If there's no transparency, we have no choice. We're not really choosing."

The local numbers, which is the whole point#

The two narratives converge on something actionable, and the dismissiveness camp commits its category error right there.

One Meta facility in Newton County, Georgia takes roughly 10% of the entire county's water supply, per New York Times reporting. Data centres in two Virginia counties near Washington used 2.1 billion gallons in 2023, with a separate estimate from the Interstate Commission on the Potomac River Basin for the regional draw. And New York Governor Kathy Hochul signed the first statewide data centre moratorium, citing "massive amounts" of water hyperscale facilities could require, with her office specifically flagging drought-stricken counties and heat waves.

Ren's point completes the picture: a data centre in a water-abundant region with affordable new utility infrastructure is a much smaller problem than the identical data centre sited in a constrained basin.

So the reframe: the global number and the local number are not in conflict. They are the same infrastructure measured at two zoom levels. "Drop in the bucket" globally and "10% of a county's water" locally are both true, and the second is the one that produces a moratorium. The global aggregate is the wrong unit, and everyone arguing with it at that scale already knows it.

The cooling trade-off nobody counts#

The axis both debates argue around is the trade itself. Air chillers versus evaporative cooling moves the burden from water to electricity. Zero-water is trivially achievable, and the cost is substantially more electricity to reject the same heat. So the honest question is not "how much water" any more than it is "how much power." It is which resource a given site should spend, given what it has and what it lacks. Both camps ignore that trade entirely, and reframing the fight as "water or power" is a contribution neither side has made.

The political layer#

The US Department of Energy's op-ed, by Assistant Secretary Tim Walsh of the Office of Environmental Management, published in the Washington Times on 26 July 2026 and on energy.gov on 27 July, claims the misconception is that AI data centres use enormous water, that closed-loop cooling recirculates water through sealed pipes and heat exchangers rather than evaporating it, often putting AI data centres among the most water-efficient facilities, that closed-loop systems operate independently of drought conditions or water restrictions, and that developers have adopted a Rate Payer Protection Pledge funding their own generation and transmission upgrades. It cites Pike County, Ohio, where DOE land is being revitalised with Japanese-financed power generation, referencing "President Trump's trade deal with Japan," and closes with "AI data centers are far more efficient and responsible than often portrayed."

Is that a government talking point? It is an op-ed by a political appointee in an administration with a stake in data centre permitting, published in a newspaper, not a DOE study. Legitimate to note. But the closed-loop cooling technical claim is not obviously self-serving and should be evaluated on its merits. Elsewhere: The Atlantic ran "Generative AI is an engineering disaster" and an April 2026 cover feature titled "Inside the dirty, dystopian world of AI data centers." Pirate Wires ran "The Data Center Water Crisis Isn't Real" in December 2025. The Atlantic's own line on the whole circus: "As with so many debates around AI, the arguments are not always presented in good faith."

What I measured on my own stack#

This is the first-party bit, and the disclosure matters more than any number. I audit the local CPU voice stack and the T4 setups, logging actual kilowatt-hours and actual litres at each stage, with the number published per 1,000 generated tokens. It is a small, real, reproducible footprint for local inference, and it is the only kind of number like it, because every published water figure is for hyperscale serving, which is a different animal entirely.

The second thing worth doing is a full scope reconstruction of one facility. Google publishes at facility level, including an Iowa site at roughly 1 billion gallons in 2024. Take one disclosure, reconstruct its direct, indirect, and embodied figures from public grid and manufacturing data, show the arithmetic, and show exactly which inputs are guesses. That exercise, done even once for a single site, kills the boundary confusion permanently.

And the honest limitation, which I will write louder than the measurement: per-query figures for proprietary APIs are unobtainable. I cannot give you mine for Claude or GPT because the labs do not publish the inputs, and a number that is actually a boundary-mismatched estimate would undo the entire point of this post.

What actual disclosure would settle it#

Per-facility, per-basin, quarterly withdrawal plus grid mix. Five numbers per company per region. That is exactly what two identical bills would require operators to submit to states, EPA, DOE or USDA: S.4213, the Data Center Water and Energy Transparency Act of 2026, sponsored by Sen. Durbin on 25 March 2026 and referred to the Committee on Energy and Natural Resources, and its House twin H.R.9825, from Rep. Underwood on 22 July 2026, referred to House Energy and Commerce. Note what the regulators are choosing to do instead. New York restricts infrastructure, a statewide moratorium. India brakes on content rather than infrastructure. Almost nobody mandates the disclosure that would let anyone adjudicate the argument, and the two bills that would are both still in committee, which is the only thing worse than no bill: two committees to fail to act.

Where this could be wrong#

You can accuse this post of not taking a side if you want, but "the dispute is a boundary problem" is a checkable, specific claim, and the scope table plus the arithmetic above is the check. That said, several real weaknesses: Amazon's numbers are self-reported and aggregate, and the scope Bloomberg reported on 11 June 2026 excludes the water used to generate the electricity, which means the headline dismissive number understates its own denominator problem even as the direction of the error favours understating water use, which cuts against the alarmists, not the dismissives. The UN report projects, its projections are roughly five years out and explicitly conditioned on growth continuing, and the report itself says AI will not simply run out of water or electricity worldwide, a sentence the alarmist version always drops. The DOE op-ed is a political appointee's newspaper piece, note it and move on. And the local-water dataset is thin: exactly one facility is well documented, Newton County, plus a statewide moratorium. The claim survives on that, but there is no public dataset of local impacts, and pointing out that it does not exist is itself part of the argument. One more weakness worth naming: the per-query boundary argument is now contested from a direction I did not expect, since Ren's own revision moves the AI water footprint up by roughly four times while Google's median text prompt at 0.26 millilitres moves it down. If both are right then the spread is even wider than 519 versus 0.3, and the fix is the same. My own local-inference footprint is a rounding error too, and I say so right here: the point is method plus a data point in a field where almost every published figure is hyperscale or scope-mismatched, not a rebuttal of anything.