How much water does AI use per prompt?
Between 0.26 millilitres and 45 millilitres, depending on where you draw the line. The two ends of that range are not competing estimates; they are different questions.
| Who published it | mL per prompt | What it counts |
|---|---|---|
| OpenAI | 0.320 | On-site cooling only |
| 0.260 | On-site cooling only | |
| Microsoft | 0.067 | On-site cooling only |
| Independent research | 10.000 | Including power-plant water |
| Mistral | 45.000 | Full life cycle, 400-token answer |
Two boundaries, a thousand-fold difference
On-site cooling is measurable and disclosed. Google and OpenAI publish figures close to a third of a millilitre. Independent research that follows the electricity back to the power station arrives at around ten millilitres, because generating electricity consumes water too, often more than the data centre itself.
Neither number is dishonest. A disclosure answers the question a company can measure directly. A life-cycle study answers the question a policymaker needs. Headlines compare them because the gap is spectacular, not because one of them is wrong.
“Google Gemini, disclosed - direct cooling only (0.26 mL/query).”
“About 500 mL for 20 to 50 exchanges, roughly one bottle of water, once power-plant water use is included.”
Why per-prompt water is a hard quantity in the first place
A data centre does not switch its cooling on for your question. Water evaporates continuously, and dividing a facility's total by its query count produces an average that no single conversation actually consumed. Every per-prompt figure, including the ones on this site, is a division of totals, not a measurement of an event.
That does not make them useless. It makes them comparable only when the boundary is stated, which is why the calculator above makes you choose one.
“A typical AI query needs 0 to 0.067 mL of cooling water, with a median below a single drop.”