Myth or fact: how much water does one AI prompt use?
The claim
“Every AI prompt uses a bottle of water.”
Verdict: Unknown
No one knows precisely how much water a single AI prompt uses, and the published figures differ by more than 100 times. Google measured 0.26 milliliters for a median Gemini text prompt, about five drops, counting only cooling water in its data centers, in an August 2025 paper. Mistral AI put a 400-token answer from its Le Chat assistant at 45 milliliters, the marginal impact of inference not counting users' devices, in a July 22, 2025, analysis.
The claim
"Every AI prompt uses a bottle of water."
Verdict: unknown
The honest answer depends on what you count, and the companies that could measure it precisely publish little. The best-documented figures range from a fraction of a milliliter to tens of milliliters per prompt.
Where did the idea come from?
The bottle comes from a research paper. In Making AI Less "Thirsty", researchers at the University of California, Riverside, and the University of Texas at Arlington estimated that GPT-3 "needs to 'drink' (i.e., consume) a 500ml bottle of water for roughly 10 – 50 medium-length responses, depending on when and where it is deployed." That works out to 10 to 50 milliliters per response, not a bottle per prompt.
The paper also projected that global AI demand could account for "4.2 – 6.6 billion cubic meters of water withdrawal in 2027." Withdrawal is water taken from a source; consumption is the part that does not return, mostly evaporated.
What does the evidence say?
The three best-documented estimates measure different things.
Three published estimates of water per AI prompt, and what each counts
| Source | Water per prompt | What it counts |
|---|---|---|
| Google, Gemini Apps (2025) | 0.26 mL (median text prompt) | On-site data center cooling |
| Li et al., GPT-3 (2023-2025) | 10 to 50 mL (500 mL per 10 to 50 responses) | Cooling, electricity generation |
| Mistral AI, Le Chat (2025) | 45 mL (400-token response) | Marginal inference, excluding users' devices; training reported separately |
What gets counted matters most. The Riverside paper counts water evaporated to cool data centers (scope 1), water used by power plants to make the electricity (scope 2), and water used to manufacture servers (scope 3). Google's 0.26 milliliters uses its data centers' Water Usage Effectiveness, a measure of on-site cooling water, and does not include water used to generate the electricity, according to the paper.
Mistral's life-cycle figure is a separate number. Its 45 milliliters covers only the marginal impact of inference, excluding users' devices, though Mistral says its figures include upstream impacts such as manufacturing servers. Its life-cycle analysis of Mistral Large 2, done with the consultancy Carbone 4 and France's ecological transition agency, ADEME, put training plus 18 months of use, as of January 2025, at 281,000 cubic meters of water.
Location and time change the answer. The Riverside researchers stress that the figure depends "on when and where" a model runs, because climate and the local power grid change how much water cooling and electricity require.
Efficiency is improving. Google says software and clean-energy changes cut the energy use of a median Gemini prompt 33 times and its carbon footprint 44 times over 12 months. In its discussion section, the paper says its figures "end up being one or two orders of magnitude smaller than many existing estimates or measurements of AI inference benchmarks."
What we don't know
OpenAI, Anthropic and most other developers have not published a per-prompt water figure with a method that can be checked. Prompts also vary: a long answer, an image or a reasoning model uses more computing than a short text reply.
FAQ
How much water does a ChatGPT prompt use?
OpenAI has not published a per-prompt figure with a checkable method. A 2023 study estimated GPT-3 used about 10 to 50 mL per medium-length response.
How much water does a Gemini prompt use?
Google measured 0.26 mL for a median Gemini Apps text prompt in 2025, about five drops, counting on-site cooling water only.
Does one AI prompt use a bottle of water?
No. The bottle figure comes from a study estimating that GPT-3 consumed 500 mL for roughly 10 to 50 responses, not one.
Why do AI water estimates differ so much?
They count different things: on-site cooling only, water used to generate electricity, or the full life cycle including training and hardware.
Sources
- Google, Measuring the environmental impact of delivering AI at Google Scale (arXiv 2508.15734, Aug. 21, 2025) · archived copy
- Mistral AI, Our contribution to a global environmental standard for AI (July 22, 2025) · archived copy
- Li, Yang, Islam and Ren, Making AI Less "Thirsty" (arXiv 2304.03271v5, March 26, 2025) · archived copy
Cite this
Myth or fact: how much water does one AI prompt use?, New Mind Museum, November 6, 2026. https://newmindmuseum.com/culture/2026/11/myth-how-much-water-does-one-ai-prompt-use/
New Mind Museum. (2026, November 6). Myth or fact: how much water does one AI prompt use?. https://newmindmuseum.com/culture/2026/11/myth-how-much-water-does-one-ai-prompt-use/
“Myth or fact: how much water does one AI prompt use?.” New Mind Museum, 6 Nov. 2026, https://newmindmuseum.com/culture/2026/11/myth-how-much-water-does-one-ai-prompt-use/.
@misc{new-mind-museum-2026-myth-or-fact-how-much-water-does-one-ai-,
author = {{New Mind Museum}},
title = {Myth or fact: how much water does one AI prompt use?},
year = {2026},
url = {https://newmindmuseum.com/culture/2026/11/myth-how-much-water-does-one-ai-prompt-use/}
}Free to quote and cite with a link to this page.