The "Bottle of Water per AI Prompt" Claim Never Was True

The claim that every AI prompt drinks a bottle of water is wrong by roughly 1,500 times on the best-measured figures. Google and OpenAI each report about a quarter to a third of a milliliter per prompt. That works out to roughly 1,500 to 1,900 prompts per 500 mL bottle.

Where the bottle came from

The figure traces to a paper by researchers Li, Yang, Islam, and Shaolei Ren, "Making AI Less 'Thirsty,'" first posted in April 2023. The key line reads: "GPT-3 needs to 'drink' (i.e., consume) a 500ml bottle of water for roughly 10 – 50 medium-length responses."

The paper's medium request was about 800 words or fewer in and 150 to 300 words out. Its estimate counted two kinds of water, and they're different. Onsite water cools the servers. Offsite water evaporates at the power plants that supply the electricity.

The onsite share was small: near zero in places like India, Denmark, and Finland, 0.24 mL per response in Georgia, and 7.6 mL at the high end in Indonesia. The offsite share ran from about 6.6 mL in Texas to 44 mL in Washington. Location and time of day moved the numbers a lot.

Add the two and the US totals in the paper's table run from about 7.6 mL in Texas to 47.5 mL in Washington, which the paper rounds to 10 to 50 responses per bottle. That's one tenth to one fiftieth of a bottle per response. It also described GPT-3, a model several generations old. The same paper estimated that training GPT-3 evaporated about 700,000 liters of freshwater in Microsoft's US data centers.

What the companies now report

In a blog post published in August 2025, Google said a median text prompt to its Gemini assistant uses 0.26 mL of water, about five drops, and 0.24 watt-hours of energy. Google's figures rest on May 2025 data. It got there by multiplying energy per prompt by its fleet's water usage effectiveness (WUE), a measure of water used per unit of energy.

Ren told The Register that Google's figure counts onsite cooling only and leaves out power-plant water. Add the power plant and the per-prompt number does go up.

OpenAI CEO Sam Altman wrote that an average ChatGPT query uses "about 0.000085 gallons of water; roughly one fifteenth of a teaspoon," along with 0.34 watt-hours. He didn't give a method. Converted, that's about 0.32 mL.

Per 500 mL bottle:

Google: 500 mL ÷ 0.26 mL = about 1,900 prompts per bottle
OpenAI: 500 mL ÷ 0.32 mL = about 1,550 prompts per bottle

No company figure, and no figure in the original paper, reaches one bottle per prompt.

The national scale, in context

The Lawrence Berkeley National Laboratory (LBNL) published a report in December 2024 on 2023 US data center use. It put direct cooling water at about 17 billion gallons a year. Indirect water, used at the power plants that supply data centers, came to about 211 billion gallons. That's roughly 12 times the direct figure, and it depends heavily on the kind of generation.

Coal and nuclear plants need the most cooling water, gas needs less, and wind and solar need almost none. Hydropower can inflate the number: studies that count the evaporation from a dam's reservoir as consumption put hydropower at a national average of 16.8 liters per kilowatt-hour, against 1.25 for thermoelectric plants, though how much of that evaporation belongs to power generation is contested, since a reservoir loses water whether or not a data center exists.

Seventeen billion gallons a year is about 47 million gallons a day. The US Geological Survey counted 118 billion gallons a day in irrigation withdrawals in 2015. Direct data center use comes to about 0.04% of that. The Golf Course Superintendents Association of America (GCSAA) survey put US golf course water at about 1.68 million acre-feet in 2020, roughly 1.5 billion gallons a day. Direct data center use is about 3% of that.

Some caveats: The USGS counts water withdrawn but the LBNL report counts water consumed. The years differ too: 2015 for irrigation, 2020 for golf, 2023 for data centers. Use them for perspective, not as a precise ratio.

The bigger lesson from the report is the 12-to-1 split. Most data center water flows through power plants, and it doesn't show up on the campus meter. Electricity demand drives the water story more than cooling towers do.

Where the real water fight is

National averages hide local strain, and they don't help a town with one big campus. Civil Beat reported that Google's campus in Council Bluffs, Iowa, which opened in 2009 and cools with evaporation, consumed about 1 billion gallons of water in 2024, roughly 2.7 million gallons a day. Microsoft doesn't break out its data center water use.

A Houston Advanced Research Center (HARC) report estimated that Texas data centers used about 25 billion gallons in 2025, roughly 68 million gallons a day. The state has 464 data centers, with more than 70 more under construction as of September 2025. HARC's 2030 projection runs from 29 to 161 billion gallons a year, a range that spans more than a fivefold gap.

Advances in technology matter. Haynes Strader, chief development officer of Skybox Datacenters, told Rep. Senfronia Thompson at the August 19, 2026 House State Affairs hearing, "Every building we've built since 2014 uses closed-loop technology, so it uses very minimal water."

A reduction in electricity usage per query cuts water usage at the building and at the plant. HARC finds that a one percent gain in energy efficiency cuts data center water use more than a one percent gain in cooling efficiency, all else equal.

And Google measured a 33-fold drop in energy for its median prompt from May 2024 to May 2025.

So even with the power plant included, a bottle of water per prompt is still off by orders of magnitude, and should continue to shrink as technology continues to improve.

The "Bottle of Water per AI Prompt" Claim Never Was True

Where the bottle came from

What the companies now report

The national scale, in context

Where the real water fight is

Sources

Get daily intelligence like this delivered to your inbox