How much water do data centres use? The answer depends entirely on whether you count the power plants

Every figure published for data centre water use answers one of two different questions, and they are almost never labelled.

**On-site cooling water** is what evaporates in the cooling towers of the building itself. **All-in water** adds the water consumed generating the electricity that building draws — thermal power plants evaporate substantial volumes, and that consumption is real even though it happens somewhere else.

The gap between those two accountings is where nearly all of the apparent disagreement lives.

Google's published measurement is a useful anchor because its scope is stated: a median Gemini text prompt consumes **0.26 millilitres** of water. That figure counts water used in Google's data centres for cooling. It excludes water consumed by the power plants supplying them.

For contrast, take the most-cited academic estimate. In 2024 The Washington Post worked with researchers at the University of California, Riverside to estimate that GPT-4 producing a 100-word email consumed **519 millilitres** — cooling water plus the water behind the electricity. That figure became "a bottle of water per email" and travelled everywhere.

Those two numbers are roughly two thousand times apart and both were produced carefully. The distance comes from three things: a 100-word email is many prompts' worth of output where 0.26 ml is a single one; one estimate is all-in and the other is cooling-only; and a 2024 model of an average US data centre is not a 2025 measurement of Google's own fleet.

There is a further wrinkle that gets left out. **Shaolei Ren, the UC Riverside researcher whose team produced the 519 ml figure, has since revised it** to roughly 15 millilitres per GPT-4 prompt, of which about 5 millilitres is on-site cooling. He describes the original as outdated for current systems. That revision is roughly thirty-five fold and has travelled nowhere near as far as the number it corrects.

Line those up with their scopes attached and they stop contradicting each other. Cooling-only, per prompt: about 0.26 ml measured by Google, about 5 ml in Ren's revised estimate for a different model on different hardware. All-in, per prompt: about 15 ml. The remaining spread is infrastructure and model, which is what you would expect.

So the per-prompt water cost of a text query is millilitres under every current estimate, including the corrected version of the one that started the argument.

That does not settle the thing actually being argued about. Per-prompt figures say nothing about aggregate withdrawal, and nothing at all about **where** the water is drawn from. A facility with a modest per-query footprint sited in a water-stressed county is a local problem regardless of how the arithmetic divides out. Whether a given datacentre strains its watershed depends on its cooling design, its location, the season and the local supply — none of which any per-prompt number addresses.

Facility-level withdrawal figures are rarely disclosed, which is why that local question usually cannot be answered from public data even when it is the only question that matters to the people asking it.