How much water does AI use?
By the figures Google and OpenAI publish, a text prompt uses about a quarter to a third of a millilitre — a few drops. Mistral's life-cycle study puts a longer response at 45 millilitres, because it counts far more than cooling. For images and video, nobody publishes a water figure at all.
Copilot Microsoft
Disclosed by the maker
0 – 0.067 mL
0 – 0.49 L a year
Gemini Google
Disclosed by the maker
0.26 mL
1.9 L a year
ChatGPT OpenAI
Disclosed by the maker
0.32 mL
2.3 L a year
Mistral Vibe (Mistral Large 2) Mistral AI
Disclosed by the maker
45 mL
330 L a year
Yearly figures assume 20prompts a day. Microsoft's range is for typical queries to its largest models, not Copilot specifically.
Claude; Grok, Meta AI, DeepSeek & others.
Not disclosed
Why the figures differ a hundredfold
Cooling only, or the whole life cycle. Google, OpenAI and Microsoft describe water used around the data centre. Mistral's life-cycle analysis also counts the water behind making the hardware — a much wider boundary, and a much bigger number.
How a data centre is cooled.Some facilities evaporate water to shed heat; others don't. Microsoft's range starts at zero, which is what a design without evaporative cooling looks like on paper.
Where it is. The same prompt costs a region with plenty of water very little, and a dry one more — a difference no per-prompt average can show.
Common questions
- How much water does ChatGPT use?
- OpenAI states an average query uses about 0.000085 gallons, roughly 0.32 millilitres — a few drops. It has not published how that was measured.
- How much water does Gemini use?
- Google reports 0.26 millilitres for a median Gemini Apps text prompt, measured across its serving infrastructure and published with a methodology.
- Why does Mistral report so much more water?
- Mistral's 45 millilitres per 400-token response comes from a life-cycle analysis. It counts water across the whole life cycle — including making the hardware — rather than only what the data centre uses, so it answers a broader question.
- Does generating AI images or video use water?
- Almost certainly, since it uses far more energy per item than text. But no maker publishes a water figure for image or video generation, so we don't show one.
- Is AI's water use a problem?
- Per prompt, the published figures are tiny. The concern is local: data centres draw water in specific places, and where water is already scarce that matters regardless of how small each prompt's share is.
Sources
- Google — technical paper on the energy, emissions and water of Gemini prompts (arXiv 2508.15734) (2025-08-21)
- Sam Altman (OpenAI) — The Gentle Singularity (2025-06-10)
- Mistral AI — Our contribution to a global environmental standard for AI (2025-07-22)
- Microsoft — Scaling AI with 8 to 20x energy efficiency (2026-06-15)
- Microsoft researchers — per-query energy of frontier AI serving (Joule, 2026) (2026)
Last reviewed 2026-09-15.