How much water does ChatGPT use? The researcher behind the famous number has revised it

If you have seen a number for how much water ChatGPT uses, it was probably around half a litre — a bottle of water for a single email. That figure is real, it came from serious researchers, and the person behind it no longer stands by it.

Here is where it came from. In 2024, The Washington Post worked with researchers at the University of California, Riverside to estimate the water cost of generative AI. Their model put GPT-4 producing a 100-word email at an average US data centre at **519 millilitres**. That number counted two things: the water evaporated to cool the servers, and the water consumed at power plants generating the electricity those servers used.

It travelled. It became the standard citation, usually stripped of its conditions and restated as a bottle of water per prompt.

Shaolei Ren, the UC Riverside researcher whose team produced it, has since revised the estimate to roughly **15 millilitres** for a GPT-4 prompt, of which about 5 millilitres is on-site cooling. He describes the original figure as outdated for today's systems. That is a reduction of roughly thirty-five fold, and it has not travelled anywhere near as far as the original did.

Then there is Google's own number. In August 2025 Google published a technical paper measuring its production systems and reported that the median Gemini text prompt consumes **0.26 millilitres** of water — about five drops — alongside 0.24 watt-hours of energy.

So the published figures span from 519 millilitres to 0.26. That looks like a scientific dispute. It mostly is not.

The numbers differ on three separate axes, and comparing them directly gets all three wrong.

**They measure different things.** 519 ml was for a 100-word email — many prompts' worth of output. Google's 0.26 ml is a single median text prompt. These are not the same unit.

**They count different water.** The UC Riverside estimate is all-in: cooling water plus the water consumed generating the electricity. Google's figure counts only water used in its data centres for cooling, and excludes the water consumed by the power plants supplying them. Ren's own revised split makes the size of that gap visible — about 5 ml of his 15 ml is cooling; the rest is upstream.

**They describe different infrastructure.** A 2024 estimate of an average US data centre and a 2025 measurement of Google's own fleet are not describing the same buildings. Efficiency has moved, and Google is measuring hardware it designed.

One further caveat belongs on Google's figure specifically: it is self-reported by the company whose environmental footprint is in question, drawn from systems no outside party can independently measure. That does not make it wrong. It does make it a company disclosure rather than an independent finding, and it should be read as one.

What can be said with confidence is narrower than either camp's headline. The per-prompt water cost of a text query is small — millilitres, not litres — under every current published estimate including the revised one from the researcher who produced the original. The widely circulated bottle-of-water figure overstates a single prompt by a wide margin, and its own author says so.

None of which settles the question people are actually arguing about. Per-prompt figures say nothing about aggregate demand, and nothing at all about where the water is drawn from. A data centre with a modest per-query footprint sited in a water-stressed county is a local problem regardless of how the arithmetic divides out per prompt. That is a question about siting, permitting and regional supply, and it is not answered by any of the three numbers above.