TL;DR
One AI prompt uses somewhere between a quarter of a millilitre of water (Google's figure, cooling only) and about 50 millilitres (the highest independent estimate, power plants included). One beef burger takes about 2,350 litres. Even counting only river and groundwater, a burger equals thousands of prompts. So asking a chatbot a question isn't where your water footprint lives.
But the worry isn't wrong either. Most of them cool with evaporation, so that water is gone from the local supply. It doesn't come back. A single large site can pull over a million gallons on a hot day from one town's system, and many new ones are planned in drought areas like Arizona, Texas and Georgia. The power plants that feed them use even more water than the cooling does.
The fair summary: AI's water use is small per person and large per location. The fix is about siting and rules (don't build evaporative sites in drought basins, require public water reporting, use dry or closed loop cooling where water is short), not about feeling guilty for each prompt.
Two arguments that talk past each other
When people argue about AI and water, one side says "a prompt uses a few drops, a burger uses a bathtub, relax." The other side points at a town in Arizona that just lost a fight with a data center developer. Both sides have real numbers. They're measuring different things.
The per prompt argument is about scale across a whole country. The local argument is about concentration in one place. A city doesn't plan its water around national averages. It plans around how much it can pull from its own river or aquifer in August.
First, a correction: the water often does disappear
A common defence of data centers is that cooling water "just gets warm and goes back." That is true for once through cooling, where water passes a heat exchanger and returns to the river. Most big sites don't work that way.
Most use evaporative cooling towers. Heat leaves the building by turning water into vapour. That vapour goes into the air and comes down as rain somewhere else, usually far from the watershed it came from. Hydrologists count this as consumed water, the same category as irrigation water that a crop pulls up and transpires.
The water left in the tower gets saltier as the rest evaporates. Operators flush it out ("blowdown"), and it carries concentrated minerals and treatment chemicals to the sewer. Not returned, and not clean.
There are designs that use almost no water on site: closed loop systems, dry coolers, and direct liquid cooling of chips. The catch is that they usually need more electricity, and electricity has its own water cost at the power plant.
The per prompt numbers
| What | Water | Source |
|---|---|---|
| Median Gemini text prompt, on site cooling only | 0.26 mL | Google, August 2025 |
| Same prompt, chips only (Google's "optimistic" case) | 0.12 mL | Google, August 2025 |
| Independent estimate including power plant water | about 50 mL | Shaolei Ren, UC Riverside |
| One 150 g beef burger | about 2,350 L | Water Footprint Network |
Google's figure has two known gaps. It counts only the water used on site for cooling, and no independent auditor has checked it. Meta, the only hyperscaler that publishes its indirect water, reports that the indirect share is more than 20 times the direct share. So the high estimate is the safer one to use.
Even with the high estimate, one burger equals about 47,000 prompts.
The fair objection to the burger comparison
The burger number has a weakness, and anyone who knows water accounting will raise it. About 87 percent of livestock water is "green" water: rain that falls on pasture and feed crops and would have fallen there anyway. Only about 6 percent is "blue" water pumped from rivers and aquifers, which is the kind data centers compete for.
Count blue water only, and a burger drops to something over 100 litres. That's still around 2,000 prompts at the highest per prompt estimate. The comparison holds. It's just less dramatic than "a burger equals millions of prompts."
The national totals
Lawrence Berkeley National Laboratory estimates that US data centers directly consumed about 17.4 billion gallons of water in 2023, or about 48 million gallons a day. The water used to generate their electricity was about 12 times larger, around 211 billion gallons.
That indirect figure is contested. It includes hydroelectric dams, where reservoir evaporation is large and hard to assign to any one user. But the direction is clear, and NPR reported this month that experts now agree powering data centers takes more water than cooling them.
Where the critics are right
Concentration. A large evaporative data center can need more than a million gallons a day in summer, from one municipal system sized for homes and shops. In a heat wave.
Location. One national tally found that 517 of 809 planned US data center projects are in areas officially classified as drought stricken. Two Arizona communities rejected proposals after public pushback over water. Georgia asked data centers to cut use during drought, and residents near one Georgia site reported trouble getting water from their wells. A University of Texas report projects Texas data center water demand rising from 120 million gallons a day in 2025 to 640 million by 2030.
Power. A Ceres analysis estimates that if all planned data centers in the Phoenix area come online, water demand for electricity there rises by nearly 400 percent.
Disclosure. When the Texas Water Development Board asked 341 data centers to report their 2025 water use, about 80 percent had not responded a week before the deadline. You can't plan around a user that won't say how much it takes.
What would actually help
None of this means you should use AI less. The per person footprint is small, and the same server farms run search, video and email. It means building them in the right places, the right way:
- Do not approve evaporative cooling in overallocated basins. Require dry or closed loop cooling there.
- Make water reporting a condition of the permit, including peak daily draw, not just the annual total.
- Count the power plant water when a region plans for new load.
- Prefer sites with wind, solar or reclaimed water supplies, which use far less water per kilowatt hour than thermal plants.
The short version: stop measuring AI's water by the prompt, and start measuring it by the town.
Sources
- LBNL, 2024 United States Data Center Energy Usage Report
- ITIF, The Data Center Water Problem Is Soluble
- Small Bottle, Big Pipe: data centers and public water systems (arXiv)
- Google Cloud, Measuring the environmental impact of AI inference
- TechRadar, what Google's five drops figure leaves out
- Mekonnen and Hoekstra, water footprint of farm animals
- Ercin, Aldaya and Hoekstra, water footprint of a beef burger
- NPR, Experts say powering data centers uses more water than cooling them (2026-11-20)
- SACE, data centers use 80 to 215 percent more water when power generation is included
- Fortune, data center water fights in Georgia and Arizona
- Al Jazeera, Arizona locals fight data centres
- Route Fifty, Texas emergency water supply delays
