In short
Water use follows energy use: multiply the energy of your AI work by the data centre's water usage effectiveness (WUE) for on-site cooling, and add the water used to generate the electricity if it is in your boundary. Published per-prompt figures range from a fraction of a millilitre to tens of millilitres, mostly because they count different things. Where the water is used matters as much as how much.
Steps
- Estimate the energy of the AI work first, in kilowatt-hours.
- Multiply by the data centre's WUE, in litres per kilowatt-hour, for on-site water.
- Add off-site water used in electricity generation if your boundary includes it.
- Note the location, and whether the region is short of water.
- State the boundary and give a range rather than a single figure.
Estimate water from energy, then say where it was used. Data centres use water mainly for cooling, and power stations use water to generate the electricity, so both sit behind an AI answer.
How big are the published figures?
| Source | What it measured | Water |
|---|---|---|
| Google (2025) | Median Gemini text prompt | 0.26 mL |
| Mistral AI life-cycle assessment (2025) | A 400-token response, counting more stages | 45 mL |
| Li, Yang, Islam and Ren | Training GPT-3 in Microsoft's US data centres | about 700,000 litres directly evaporated |
Li and colleagues also estimated that GPT-3 needs to "drink" a 500 mL bottle of water for roughly 10 to 50 medium-length responses, depending on when and where it runs, and projected global AI demand could account for 4.2 to 6.6 billion cubic metres of water withdrawal in 2027. These are estimates for one older model and set of data centres. Newer company figures, such as Google's, are much lower; compare what each counts before drawing conclusions.
What goes into an estimate?
- On-site water: energy × WUE. See making sense of PUE and WUE.
- Off-site water: the water used per kilowatt-hour by the local power stations.
- Manufacturing: chip making uses water too, if your boundary is wide.
Common mistakes
- Mixing withdrawal and consumption. Water taken and returned is not the same as water evaporated.
- Ignoring location. A litre in a drought-hit region matters more than one where water is plentiful.
- Comparing figures with different boundaries, as in the table above.
To cut water use, see reducing the water footprint of AI and hosting and data centres.
Sources
- Making AI less 'thirsty': uncovering and addressing the secret water footprint of AI models, Li, Yang, Islam and Ren (2023)
- Measuring the environmental impact of delivering AI at Google scale, Elsworth et al., Google (2025)
- Our contribution to a global environmental standard for AI, Mistral AI (2025)
Related
How do I reduce the water footprint of AI?
Cut AI's water use by using less compute, choosing hosts and regions with low water stress and good WUE, and timing heavy work where it helps.
How do I make sense of PUE and WUE figures?
PUE shows a data centre's energy overhead; WUE shows its on-site water per kWh. What the numbers mean, how to compare them and what they leave out.
How do I check a claim about AI's energy use?
Six questions to test any claim about AI energy, water or carbon: per what, who says so, what's included, how old, compared with what, total or per use.
Green AI hosting and data centres
How to choose where AI runs: grid carbon intensity, PUE and WUE, renewable energy claims, water use and siting, with questions to ask hosts.