Why two honest water numbers can differ 170 times

ecogpt.coPublished Oct 20, 2026 · 2 min read

Google says a prompt uses 0.26 mL of water. Mistral says a response uses 45 mL. Neither is wrong. They draw the boundary in different places, and the boundary is the whole story.

Google's 0.26 mL counts water evaporated cooling its own data centres, derived from fleet water-use effectiveness.
Mistral's 45 mL also counts the water consumed at power plants generating the electricity, following a full life-cycle method.
Generating a kilowatt-hour at a US thermoelectric plant evaporates about 1.8 litres, more than most data centres use to cool with it.
The index records the boundary next to every water figure and never compares across boundaries without saying so.

Two figures, two questions

Google's inference study reports 0.26 mL of water per median Gemini prompt. Mistral's life-cycle assessment reports 45 mL per 400-token Le Chat response. The gap is about 170×, and it has nothing to do with one company being careless.

Google answers the question "how much water evaporates at my data centre to cool this prompt?" Mistral answers "how much water does this response consume anywhere in the world?" The second question includes the first and adds the biggest term: the water used to make the electricity.

Where the water goes

A data centre turns almost all its electricity into heat. Carrying that heat away with evaporative cooling consumes water on site; Google's fleet runs at about 1.15 litres per kilowatt-hour, AWS at 0.15, Microsoft at 0.30. That is the on-site term, and it is what water-use effectiveness (WUE) measures.

The off-site term is larger and invisible. Thermoelectric power plants, which still supply most grids, evaporate water in their own cooling towers. NREL's long-standing estimate for the United States is about 0.47 US gallons, 1.8 litres, per kilowatt-hour delivered. So a prompt that uses 0.3 Wh of grid electricity has roughly 0.5 mL of water behind it before any data-centre cooling is counted.

Why Mistral's number is bigger still

Mistral's 45 mL is not just on-site plus generation. The life-cycle method amortises training over the model's useful life and includes the water embedded in manufacturing the hardware. For a model served from France, where nuclear plants have large cooling-water demands, the generation term dominates. It is a different, broader accounting, reviewed by two independent firms under a French standard. It is also the only such figure any frontier lab has published.

OpenAI's figure has no boundary

Sam Altman's blog gives 0.000085 gallons, about 0.32 mL, per average query. No scope is stated, so we cannot tell which question it answers. It sits close to Google's on-site figure, which suggests it is the same kind of number, but that is an inference, not a disclosure. The index records it as disclosed and flags the missing scope.

How the index keeps them apart

Every water figure on a company profile carries a note stating its boundary. Where we estimate water for a company that publishes nothing, we use the on-site boundary, energy × the fleet's WUE, so the estimate is comparable with Google and OpenAI, and we print the full-scope figure in the method note. Sorting the rankings by water per prompt sorts on that on-site boundary.

The practical reading

Five drops per prompt is a fair description of on-site cooling at an efficient hyperscaler.
Two to three tablespoons per response is a fair description of the whole supply chain.
Both can be true at the same time, for the same prompt.

If a headline gives one number without saying which it is, it has told you less than it seems.