How much water does AI use?
Google measured its median Gemini text prompt at 0.26 milliliters of water — about five drops. Meanwhile all U.S. data centers combined consumed 66 billion liters on-site in 2023, plus roughly 800 billion liters at power plants. Both numbers are real. They differ because they measure different things.
Why — the first-principles explanation
AI uses water in two separate places, and almost every argument about this topic comes from mixing them up. The first is the data center itself. Chips turn nearly all the electricity they draw into heat, and that heat has to go somewhere. Many facilities dump it by evaporating water in a cooling tower, because evaporation is extraordinarily good at absorbing heat: turning one liter of water into vapor soaks up roughly 2,260 kilojoules, about 0.6 kWh. That water leaves as vapor and does not come back — it is consumed, not just borrowed. The second place is the power plant. Coal, gas, and nuclear plants boil water and cool their own condensers, so every kilowatt-hour a data center buys carries hidden water use somewhere else. Berkeley Lab puts that at about 4.52 liters per kWh for the U.S. grid mix serving data centers.
Once you see the two locations, the wild spread in published numbers stops being mysterious. Berkeley Lab's 2024 report to Congress found U.S. data centers directly consumed 66 billion liters on-site in 2023 (up from 21.2 billion in 2014), while their indirect footprint at power plants was nearly 800 billion liters — about twelve times larger. Per day, those 2023 totals average out to roughly 180 million liters on-site and about 2.2 billion liters at power plants. Same industry, same year, two numbers that differ by more than 10x purely from where you draw the boundary. A second boundary trick is withdrawal versus consumption: withdrawn water is taken from a river and mostly returned; consumed water is evaporated and gone. Scary headlines often quote withdrawal and call it consumption.
Per-prompt figures diverge for the same reason, plus two more. Google's 2025 measurement of a median Gemini text prompt gives 0.24 Wh and 0.26 mL. The widely-cited 2023 estimate from Li and colleagues at UC Riverside put GPT-3-class inference at roughly a 500 mL bottle per 10–50 responses — that is 10–50 mL each, up to about 40x higher. Why? Google's number covers on-site water only (add power-plant water at Berkeley Lab's intensity and 0.24 Wh implies roughly another 1 mL); Google reports a 33x energy improvement in twelve months, so hardware from 2023 is simply not hardware from 2025; and outside researchers must model from public specs while companies measure their own fleets and pick their own boundaries. Google's paper is also not peer-reviewed.
The honest synthesis: a single prompt's water use is trivial — you cannot meaningfully conserve water by not chatting with a bot. The aggregate is not trivial, and it is concentrated. Data center water lands in specific counties and specific aquifers, not spread evenly across the country, which is why a number that looks tiny nationally can still matter enormously in one drought-stressed town.
An example that makes it click
Think about how your body dumps heat. On a hot day you sweat, and the sweat evaporates off your skin — that evaporation is what actually cools you, and the water is gone for good. A data center with evaporative cooling is a building that sweats. A closed-loop, air-cooled data center is more like a car radiator: it moves heat around with fans and never loses the coolant, but it burns more energy to do it.
Now the boundary trick. Say you drink a glass of water: that's your "on-site" use — one glass. But growing the coffee beans, cooling the power plant that ran your kettle, and manufacturing the mug consumed hundreds of glasses somewhere else. If one report counts your glass and another counts the whole chain, they'll disagree by 100x and both will be telling the truth. That is exactly the gap between Google's 0.26 mL per prompt and the scarier figures you see quoted.
Key facts
- Google measured the median Gemini Apps text prompt at 0.24 Wh of energy, 0.26 mL of water (about five drops), and 0.03 gCO2e, published August 2025 (arXiv:2508.15734, not peer-reviewed).
- Using a narrow chip-only boundary instead, the same prompt scores 0.10 Wh and 0.12 mL — roughly half, from boundary choice alone.
- U.S. data centers consumed 66 billion liters (about 17.4 billion gallons) of water on-site in 2023, up from 21.2 billion liters in 2014 (Berkeley Lab, December 2024).
- Indirect water consumed at power plants for those same data centers was nearly 800 billion liters in 2023 — about 12x the on-site figure (Berkeley Lab).
- Training GPT-3 in Microsoft's U.S. data centers directly evaporated about 700,000 liters on-site, and roughly 5.4 million liters including power generation (Li et al., arXiv:2304.03271).
- Berkeley Lab estimates the U.S. grid mix serving data centers consumes 4.52 liters of water per kWh generated, versus a 4.35 L/kWh national average.
▶ The 60-second explainer (script)
How much water does AI use? One prompt: about a quarter of a milliliter — five drops. That's Google's own 2025 measurement of a median Gemini text prompt. So why do you keep seeing headlines about bottles of water per question? Because of boundaries. AI touches water in two places. First, the data center evaporates water in cooling towers to dump the heat coming off the chips. Second — and this is the bigger one — the power plants making the electricity cool themselves with water too. Berkeley Lab found U.S. data centers consumed 66 billion liters on-site in 2023, but nearly 800 billion liters at the power plants. That's twelve times more, same industry, same year. The difference is just where you draw the line. Google's 0.26 milliliters counts only on-site water. Add the power plant and you're at roughly a milliliter. And a 2023 estimate of GPT-3 put it near 10 to 50 milliliters per answer — but that was older hardware, and Google says efficiency improved 33-fold in a single year. So the honest answer: per prompt, water use is genuinely negligible. In aggregate, it's large and growing fast — and it lands in specific towns and specific aquifers, which is where it actually bites.
What authoritative sources say
People also ask
Does one ChatGPT question really use a bottle of water?
No. That claim comes from a 2023 estimate that GPT-3 used a 500 mL bottle per 10–50 responses — so 10–50 mL per answer, not 500 mL. Google's 2025 measurement of a comparable text prompt is 0.26 mL.
Why do the numbers disagree so much?
Mostly boundaries: on-site cooling water versus power-plant water, withdrawal versus consumption, and training versus per-query use. Hardware generation matters too — Google reports a 33x energy improvement in twelve months.
Is the water gone forever?
Evaporated water returns to the atmosphere and eventually rains somewhere, but it leaves the local watershed. That is why "consumption" is tracked separately from "withdrawal," which is mostly returned to the source.
Can data centers run without water?
Yes. Closed-loop and air-cooled designs use almost no on-site water, but they draw more electricity — which means more water at the power plant. It's a trade, not a free win.
The same question, asked other ways
- Does AI use water?14,800/mo
- How much water does AI use per day?4,400/mo
- Does AI use a lot of water?3,600/mo
- How much water does AI use per prompt?1,000/mo
- How many gallons of water does AI use?1,000/mo
- Does AI use water to run?720/mo
- How much water is AI using?590/mo