How Much Water Does AI Actually Use? The Real Numbers, With Error Bars

How Does Artificial Intelligence Use Water?

AI uses water in two places: directly, evaporated to cool the servers running the model inside a data center, and indirectly, at the power plants generating the electricity those servers draw. Most public debate only counts the first and misses the second, which is often the bigger share.

Data centers running AI workloads generate enormous amounts of heat, and a lot of them still shed that heat with evaporative cooling towers. Outside air is pulled across water-wetted media, some of the water evaporates, and the resulting cooler air (or a chilled-water loop) keeps the racks of GPUs from overheating. It is a cheap, effective way to cool very dense compute, and it is also a direct, physical loss of fresh water to the atmosphere.

The second, less visible piece is the electricity itself. Most of the US grid still runs on thermoelectric power, natural gas, coal, and nuclear plants that use water for their own cooling. Every kilowatt-hour an AI model draws from that grid carries a water cost before it ever reaches a server rack. Researchers call this the offsite footprint, and the 2023 UC Riverside paper 'Making AI Less Thirsty' was the first widely cited study to add onsite and offsite water together into one number per query.

How Much Water Does One AI Query Actually Use?

Published estimates for a single text prompt range from about 0.26 milliliters, five drops, to roughly 500 milliliters, a full bottle. That is close to a 2,000-fold gap between the smallest and largest numbers currently in circulation, and which one is 'right' depends on the model, the hardware, and who is doing the estimating.

None of these four numbers is measuring the same thing. Google's and OpenAI's figures are self-reported medians across their real production fleets, not independently audited but grounded in actual hardware and grid data. The UC Riverside numbers are third-party estimates built on public assumptions about model size, GPU count, and cooling system, assumptions the vendors don't confirm or deny. Model size, region, cooling technology, and whether a study bundles in a share of training cost all move the number by orders of magnitude before anyone even disagrees about methodology.

Why Do These Estimates Vary By Almost 2,000x?

The largest gap traces back to one number getting misapplied. A 2023 study estimated GPT-3's water cost per prompt. A 2024 report plugged in a far larger, unverified energy estimate for GPT-4 into the same formula, and the water figure ballooned toward a full bottle. The study's own author now says the real number is closer to 15 milliliters.

The 500 mL 'AI uses a bottle of water per prompt' claim came from a 2024 Washington Post story, which credited the number to Shaolei Ren, the UC Riverside researcher behind the original 2023 paper. Blogger Andy Masley traced the math: Ren's 2023 study estimated GPT-3 used about 10 to 50 prompts per bottle of water, using an energy estimate of roughly 4 watt-hours per prompt. The 2024 figure instead assumed 140 watt-hours per GPT-4 prompt, 35 times higher, and ran it through the same onsite-plus-offsite formula. That single assumption, never independently confirmed, is what pushed the water number up by roughly 30x. Ren has since told Masley the real per-prompt figure is closer to 15 mL total, and about 5 mL if you count only the water evaporated onsite at the data center.

This is why the honest answer to 'how much water does AI use' is a range with a citation on every number, not a single figure. Self-reported vendor numbers like Google's and OpenAI's are the closest thing to ground truth we have right now, because they come from real fleets, but they are not independently audited, and a company has an obvious incentive to publish its best-looking number. Independent academic estimates are transparent about their assumptions but depend on public specs that change fast and that vendors rarely confirm.

What Do Google, Microsoft, And OpenAI Actually Report?

Google says a typical Gemini text prompt uses 0.26 mL of water and 0.24 watt-hours of energy, about what nine seconds of TV costs. OpenAI's Sam Altman has cited a similar 0.32 mL figure for ChatGPT. Microsoft doesn't publish a per-query number but says it has cut data center water intensity by nearly 90% since the early 2000s.

Google published its methodology in August 2025, the first time a major AI lab detailed how it measures energy, carbon, and water per prompt rather than just per data center. The company reported the energy footprint of a median prompt fell roughly 33x over a recent 12-month stretch, driven by model and hardware efficiency work rather than by asking people to prompt less. Sam Altman's figure for ChatGPT, posted the same month, was not peer reviewed and did not specify what counts as an 'average' query, so treat it as a directional company disclosure, not an audited result.

Microsoft takes a different approach: instead of a per-query number, it reports fleet-wide water use effectiveness (WUE), liters of water per kilowatt-hour delivered. That figure dropped from 2.3 L/kWh in the early 2000s to 0.27 L/kWh in 2025. About 90% of Microsoft's 2025 fleet runs on direct air cooling with evaporative assist, a system that uses up to 90% less water than traditional cooling towers and needs no water at all for large stretches of the year in cooler climates. A newer, AI-optimized data center design introduced in 2024 uses a closed chip-level loop that consumes zero water for cooling during normal operation. Microsoft has committed to being water positive, replenishing more than it withdraws, by 2030.

How Does AI's Water Use Compare To Everyday Things?

One AI prompt is a rounding error next to ordinary daily water use. A single hamburger's water footprint is roughly ten million times larger than one AI text prompt, and a single average shower uses more water than hundreds of thousands of prompts combined.

Using Google's own median figure of 0.26 mL, one AI text prompt sits far below almost anything else people do with water in a day. A quarter-pound hamburger, mostly the water it takes to grow feed for cattle, carries a footprint of about 660 gallons, roughly 2,498 liters. An average American shower uses about 59.8 liters. Even an hour of HD video streaming, estimated by the UK's Jisc National Centre for AI at 2 to 12 liters, dwarfs a single prompt by four to five orders of magnitude.

That comparison is honest but it is not the whole story. Nobody drinks one hamburger's worth of AI in a lifetime, they send billions of prompts across a population, and the data centers built to serve that volume are real, physical facilities that draw from real, local water sources. The per-query number and the aggregate, sited impact are two different questions, and both deserve an honest answer.

How Does AI Affect The Environment Where Data Centers Are Built?

The real fight over AI's water use is not about per-query math. It is about specific communities competing with newly built data center campuses for water during active droughts, as happened in Georgia and Arizona in 2025 and 2026.

In Fayette County, Georgia, residents of the Annelise Park subdivision spent a year watching their water pressure drop before a county investigation found two industrial-scale water hookups feeding Project Excalibur, a 615-acre data center campus owned by Blackstone and developed by Quality Technology Services. One hookup wasn't authorized by the county and the other wasn't billed; by the time the error was caught, QTS had drawn more than 29 million gallons, the equivalent of about 44 Olympic swimming pools, while the entire state sat under severe to exceptional drought conditions.

In Tucson, Arizona, which averages just seven to ten inches of rain a year, the city council rejected involvement with the proposed Project Blue data center complex in 2025 over water and power concerns. Amazon withdrew, but the developer kept building. A resident's question about dust-control water later triggered an investigation that found a contractor had taken water from Tucson's own service area and trucked it outside city limits to the site without authorization. Nationally, US data centers directly consumed about 17 billion gallons of water in 2023 according to a Department of Energy-commissioned Berkeley Lab report, with hyperscale and colocation facilities using 84% of that. Hyperscale water use alone is projected to reach 16 to 33 billion gallons a year by 2028.

What Are Companies Doing To Cut AI's Water Use?

The two biggest levers are cooling technology and site selection. Closed-loop, chip-level cooling can bring a data center's water use during operation close to zero, and building in cooler, wetter regions instead of deserts avoids much of the tradeoff before construction even starts.

Microsoft's move to zero-water chip-level cooling for new AI-optimized designs is the clearest example: instead of evaporating water to reject heat, coolant runs in a sealed loop straight past the chip, similar in principle to liquid cooling in a gaming PC, just built at data center scale. Google's public per-prompt methodology serves a related purpose. By publishing the number and the method behind it, the company opens itself up to being checked, and other labs now have a public bar to match or beat.

The efficiency gains that already happened are real: Google's roughly 33x drop in per-prompt energy use over a year, Microsoft's nearly 90% cut in water use intensity since the early 2000s. Neither company reports these figures as reasons to stop scrutinizing them, and neither should you. The honest framing is that per-query efficiency is genuinely improving while total, aggregate demand is growing even faster, which is why local, sited impacts like Fayette County and Tucson are the part of this story that per-query charts can't show.

So How Much Water Does AI Actually Use?

Per query, AI's water use is small and shrinking, a fraction of a milliliter to a few tens of milliliters by the best current estimates, smaller than a shower, a hamburger, or an hour of video streaming. At the scale of billions of daily queries and the data centers built to serve them, that adds up to a measurable draw on local water supplies, concentrated in specific places.

Both things are true at once, and neither cancels the other out. If someone tells you AI uses a whole bottle of water per prompt, that specific number has been walked back by the researcher who published it. If someone tells you AI's water use is nothing to worry about, point them to Fayette County. The number that matters depends on the question you're actually asking: what does my prompt cost, or what is this data center doing to the town next to it. We'd rather give you both answers with their sources attached than pick the one that makes a better headline.