AI Infrastructure

Vol. 1·Thursday, June 25, 2026

AI's Water Problem Is Real. It's Just Not the One You've Been Told About.

Tips, corrections, or questions? support@omniscient.media

TopicsIndustry Strategy Bio & Science Research

Two numbers about artificial intelligence and water are circulating, and they sit about a thousand times apart. One holds that generating a single 100-word email with a chatbot can consume a 500-milliliter bottle of water.^[1] The other, offered by OpenAI's chief executive, puts an average query at about 0.000085 gallons, "roughly one fifteenth of a teaspoon."^[2] Both come from credible-sounding places. Both, in the way they are usually quoted, miss what actually matters.

The honest account of AI's water use is not in the per-query figure the argument keeps circling. It is in three facts that rarely lead the story. Most of the water is consumed where you cannot see it, evaporated at the power plant rather than inside the data center. The national total is genuinely small, but the local burden is not. And the companies doing the drawing have, until very recently, disclosed remarkably little about it. Put those together and the picture is neither the catastrophe of the viral posts nor the rounding error of the corporate blog. It is a real and fixable problem that almost everyone is describing with the wrong number.

Where the bottle-of-water number comes from

The figure that launched the panic traces to a 2023 paper by Pengfei Li, Shaolei Ren, and colleagues at UC Riverside and UT Arlington, titled "Making AI Less Thirsty."^[3] Its headline estimate: training GPT-3 in Microsoft's U.S. data centers could directly evaporate about 700,000 liters of clean freshwater, and roughly 5.4 million liters once the water used to generate the electricity is counted.^[3] For everyday use, the authors wrote that GPT-3 "needs to 'drink' (i.e., consume) a 500ml bottle of water for roughly 10 to 50 medium-length responses."^[3]

Read that sentence carefully, because almost everyone drops half of it. The bottle covers ten to fifty responses, not one, and the range depends on where and when the model runs. In September 2024 a Washington Post investigation, done with Ren's group, restated the idea for GPT-4 as roughly one 500-milliliter bottle to write a single 100-word email, a number that bundles both on-site cooling and off-site electricity water.^[1] That reframing, one email equals one bottle, is the version that went everywhere.

It did not go unchallenged. The engineer Sean Goedecke recomputed the chain and argued that a realistic conversation, on far more efficient modern models, lands closer to 5 milliliters than 500.^[4] The writer Andy Masley pointed out that the original figure covered 10 to 50 prompts, not one, and that roughly 85 percent of the water in that accounting comes from electricity generation rather than the data center itself.^[5] Then came Sam Altman's teaspoon. In a June 2025 blog post he wrote that an average ChatGPT query "uses about 0.34 watt-hours" of energy and "about 0.000085 gallons of water; roughly one fifteenth of a teaspoon."^[2] He offered no methodology, no model, and no hint of whether the figure includes the electricity or only the cooling. Google, publishing measured figures for its Gemini models the same year, put a median text prompt at about 0.26 milliliters, "the equivalent of five drops of water."^[6]

Line the estimates up and the spread is almost comic: from a quarter of a milliliter to more than half a liter for "one query." The gap is not fraud. It is the absence of a shared definition. The numbers differ on which model, on whether a "query" is one response or a whole conversation, and above all on whether the count includes the water burned to make the electricity. The per-query figure, in other words, is the least useful number in the entire debate, and it is the one the public fights about.

Table comparing per-query AI water estimates from Ren et al., the Washington Post, Sean Goedecke, Andy Masley, Sam Altman, and Google, ranging from about 0.26 milliliters to about 519 milliliters — Estimates for the water behind a single AI query span roughly two-thousandfold. The disagreement is mostly about what gets counted: one response or a whole conversation, and whether the electricity behind it is included.

The water you cannot see

Strip away the per-query noise and the first real finding appears: the water that matters most is not the water that cools the servers. It is the water evaporated at the power plants that feed them. Most electricity in the United States still comes from thermoelectric generation, from coal, gas, and nuclear plants that boil and evaporate water to spin turbines and shed heat. Every kilowatt-hour a data center draws carries that hidden water with it.

The numbers are lopsided. The International Energy Agency estimated that data centers worldwide consumed about 560 billion liters of water in 2023, of which roughly two-thirds was indirect water tied to electricity and only about a quarter was direct cooling water.^[7] Lawrence Berkeley National Laboratory put U.S. direct cooling water at about 17.4 billion gallons in 2023, against an indirect figure several times larger.^[9] However it is sliced, the indirect water exceeds the direct, often by a wide margin. That ratio also appears to be understated: where the IEA implies roughly 1 liter of water per kilowatt-hour of electricity, Meta's own disclosures imply closer to 3.6, and Berkeley Lab's U.S. figure is about 4.5.^[8]

Table showing direct cooling water versus indirect electricity-related water for data centers, from the IEA, Berkeley Lab, and Ren et al., with indirect water exceeding direct in every case — Across global and U.S. estimates, the water consumed generating a data center's electricity outweighs the water used to cool it. The ratio varies with assumptions about how much water each kilowatt-hour carries.

This is why the cleanest-sounding fix can backfire. Cooling a data center with air instead of evaporating water saves on-site water but raises electricity use, and that extra electricity carries its own water at the power plant. Microsoft, announcing a "zero-water" cooling design in December 2024, was candid about the trade. The closed-loop system, it said, is "filled during construction" and then "continually circulate[s] water between the servers and chillers... without requiring a fresh water supply," avoiding "more than 125 million liters of water per year per datacenter."^[10] But the company also conceded that replacing evaporative cooling with mechanical cooling would raise its power use, a "nominal increase in our annual energy usage."^[10] On a fossil-heavy grid, water saved at the rack can reappear at the generator. There is no free coolant.

Small nationally, heavy locally

At a national scale, data centers are a minor water user. The U.S. Geological Survey put total American water withdrawals at about 322 billion gallons per day in 2015, led by thermoelectric power at 133 billion and irrigation at 118 billion.^[11] Against those rivers of water, even the high-end data-center totals are a rounding error, and the direct cooling portion is smaller than what the country's golf courses consume.

The aggregate is the wrong lens, though. As Shaolei Ren and Microsoft's Amy Luers wrote in IEEE Spectrum in 2025, "on the national level, data centers' water use is relatively modest," but "the biggest stress may not be total use, but timing. On hot days when residents and businesses need water most, data-center water demand spikes too."^[12] The siting compounds it: by their account, "about two-thirds of U.S. data centers built since 2022 are in high water-stress areas."^[12] Much of that water does not come back. Companies estimate that between 45 and 60 percent of the water they withdraw is consumed, evaporated rather than returned to the local source.^[12] In Newton County, Georgia, the authors note, "some proposed data centers have reportedly requested more water per day than the entire county uses daily."^[12] A use that is trivial as a national average can still drain a single watershed at the worst possible moment.

Four places where it got real

The abstractions turn concrete in the towns that host the machines. In The Dalles, Oregon, Google fought a legal battle, bankrolling the city's effort to keep its data centers' water use secret, until a 2022 settlement forced disclosure. The records showed the data centers used 355 million gallons in 2021, about 29 percent of the city's entire water supply, after which Google agreed to stop treating site-level water as a trade secret.^[13]

In West Des Moines, Iowa, the Associated Press found that in July 2022, the month before OpenAI finished training GPT-4 in Microsoft's cluster there, the site drew about 11.5 million gallons of water, roughly 6 percent of the district's total. The same reporting documented Microsoft's global water use jumping 34 percent in a single year, to nearly 1.7 billion gallons, a rise researchers tied to AI.^[14]

In Uruguay, during the worst drought in seven decades, a proposed Google data center planned to draw about 7.6 million liters of potable water a day, the daily supply of roughly 55,000 people, while Montevideo's tap water turned salty enough to trigger government warnings. Protesters adopted the slogan "no es sequia, es saqueo" - it is not drought, it is pillage - and Google later redesigned the project to use air cooling.^[15]

In Memphis, Tennessee, xAI's "Colossus" supercomputer draws cooling water from the Memphis Sand Aquifer, the region's sole source of drinking water, even as the company fought separate battles over unpermitted gas turbines. xAI broke ground in October 2025 on an 80-million-dollar recycled-water plant meant to supply the data center with treated wastewater, then paused the project in April 2026 to prioritize a second data center, drawing rebukes from the city and its utility.^[16]

The pattern across all four is consistent, and it is not about volume. It is about drawing potable water, in stressed places, at stressed times, with as little public disclosure as the operator can manage.

What the hyperscalers use, and what "water positive" hides

The disclosure picture has improved, unevenly. Google reported consuming roughly 7 billion gallons of freshwater across its operations in 2024, replenishing about 4.5 billion gallons of that total, representing roughly 64 percent of its freshwater consumption.^[17] Microsoft reported consuming about 5,800 megaliters, around 1.5 billion gallons, in fiscal 2024, though that figure dipped partly because the company changed how it estimates sites without meters, not only because it used less.^[18] Meta reported about 3,100 megaliters consumed in 2024.^[19] Amazon, the laggard, disclosed an absolute figure for the first time only in June 2026, saying it withdrew about 2.5 billion gallons across its global data centers in 2025.^[20] The gap between "withdrawal" and "consumption," routinely blurred in coverage, matters: the first is what you take, the second is what you do not give back, and they can differ by roughly half.^[21]

All four have pledged to become "water positive," to replenish more water than they consume by 2030, largely by funding restoration projects. The promise is softer than it sounds. The pledges generally exclude the larger indirect water embedded in the electricity these companies buy. And the geographic mismatch is fundamental: a company restoring a wetland in one river basin provides no relief to the aquifer its data center is drawing down in another state. One widely cited survey found that fewer than a third of data-center operators even tracked their water consumption, and while that 2016 finding has improved, site-level disclosure remains the exception rather than the rule.^[21]

The fixes, and the catch

The engineering answers are real, and each carries a cost. Closed-loop and air-cooled designs, like Microsoft's zero-water sites whose pilot locations began construction in 2026 and are slated to come online in late 2027, can cut on-site cooling water to almost nothing, at the price of more electricity and therefore more indirect water.^[10] Recycled and non-potable water can replace freshwater for cooling, as xAI plans in Memphis and as Google and Meta already do at several sites, where supply and permits allow. Smarter siting, putting thirsty workloads in cool, water-rich regions rather than the deserts of the U.S. Southwest, helps most of all. And sheer efficiency moves the needle: Google credits hardware and cooling gains for getting a Gemini prompt down to five drops.^[6]

But the central constraint does not disappear. Water and energy are a seesaw. Spend less of one and you usually spend more of the other, and on most grids more energy means more water somewhere upstream. The only real escape is a cleaner grid, since solar and wind consume almost no water per kilowatt-hour, which is why the water question and the carbon question are, in the end, the same question.^[8]

The honest bottom line

So how thirsty is AI, really? Per query, trivially so, somewhere between a few drops and a few milliliters, and the viral "bottle per email" is an artifact of counting a whole conversation plus the power plant behind it. In aggregate, AI's water use is small against agriculture and conventional power. The IEA projects global data-center water consumption will roughly double, to about 1.2 trillion liters, by 2030, which means the window for better siting and disclosure rules is shorter than it looks.^[7] The problem that survives scrutiny is none of those headline framings. It is that this water is increasingly drawn from potable supplies, in drought-prone counties that did not choose the burden, at the hottest hours of the year, to power the least visible part of the system, and that the firms doing it have been slow to say how much.

The volume is manageable. The siting, the timing, the source, and the secrecy are not.

Sources

Consequential AI, explained and evaluated, every weekday.

The Omniscient Bulletin: 5 to 7 items a day with the take, not the recap.

Discussion