Why AI Water Estimates Differ by 2000x

Published estimates of how much water an AI prompt uses range from about 0.26 millilitres to over 500 millilitres. That is a factor of roughly two thousand, between figures that are both defensible and both produced by competent people.

The disagreement is not really about water. It is about where the boundary of the system is drawn, and almost nobody states which boundary they used.

The two headline numbers

Google has reported a median of around 0.26 millilitres of water per text prompt. Researchers at UC Riverside estimated roughly 519 millilitres for a 100-word response, about a standard bottle.

Both can be true at once, because they are measuring different systems:

	Google's ~0.26 mL	UC Riverside's ~519 mL
Counts on-site evaporation	Yes	Yes
Counts water behind the electricity	No	Yes
Unit	One median prompt	A 100-word response
Facility assumption	Google's own fleet	Modelled US average

Three multipliers, stacked. A longer response is more tokens. Including the power sector's water adds a large indirect term. And an efficient, cool-climate, closed-loop facility differs enormously from an evaporatively cooled one in a hot region.

None of those steps is a trick. Multiply them and you get three orders of magnitude.

Withdrawal is not consumption

The second confusion is more basic and does more damage.

Withdrawal is water taken from a source. Consumption is the part that does not go back, normally because it evaporated. A facility can withdraw a great deal and return most of it slightly warmer, or withdraw less and consume nearly all of it.

Headline comparisons routinely put a withdrawal figure from one source next to a consumption figure from another. The resulting contrast is meaningless, and the distinction is usually the first casualty when a study is summarised.

When you encounter any AI water number, the first question is not how big but which of these two, and measured where.

What the boundary choice does, arithmetically

The following is illustrative arithmetic, not a measurement, and it exists only to show how quickly the boundary dominates the result.

Take a facility-level consumption figure and divide by a query count. If you assume the facility serves ten million queries a day, you get one number. Assume a hundred million and you get a tenth of it. The physical facility has not changed at all; only the denominator has.

That denominator is almost never published. Providers report infrastructure totals, not queries per facility, so every per-prompt figure in circulation contains an outside estimate of a number only the provider knows.

This is why per-prompt figures should be treated as rhetoric rather than measurement. They are useful for establishing that the cost is not zero. They cannot support a comparison between providers, and they should not carry policy.

What is actually solid

Some numbers in this area are much firmer than others, and the ranking is not the one the coverage implies.

Claim type	How verifiable	Why
Facility withdrawal and consumption	Good	Reported by operators and often recorded by local water authorities
Regional data centre water use	Good	Aggregated from utility data
Training run totals for named models	Moderate	Traceable to a specific study, but modelled
Per-prompt figures	Weak	Depend on an unpublished denominator

Facility-level reporting is the level at which the question is answerable, and it is also the level at which it matters. A community sharing an aquifer with a data centre cares about the volume that facility consumes, not about a bottle-of-water metaphor.

Where the number genuinely is large

Being sceptical about per-prompt figures is not the same as saying the consumption is small.

Individual large facilities can consume on the order of millions of gallons a day, comparable to a small town. Training runs for large models have been estimated in the hundreds of thousands of litres. Those are facility-scale and campus-scale numbers, and they are the ones with the firmest provenance.

The physical constraint is real, and it belongs in the same category as the power and interconnect limits that decide what a large training cluster can actually do. Water is one of the resources that determines where compute can be built at all, which is a more consequential fact than any per-query figure.

What we do not know

We do not have a defensible way to convert facility consumption into a per-query figure, because the query denominator is not public and we do not expect it to become public.

We also do not know how quickly cooling technology is shifting the picture. Closed-loop and air-cooled designs change the direct term substantially, and the deployment mix across the industry is not published in enough detail to say what the current average actually is. Estimates modelled on a few years ago may already be wrong in either direction.

Finally, the indirect term inherits every uncertainty in grid water intensity, which varies by region and by hour. Any lifecycle figure is a estimate resting on another estimate, and should be read that way.

How to read the next figure you see

Ask three questions, in order:

1. Withdrawal or consumption? If unstated, the number cannot be compared to anything.
2. Does it include the water behind the electricity? This alone accounts for most of the spread.
3. What is the denominator? If it is per-prompt, someone has guessed a query count.

A figure that survives all three is worth taking seriously. Most do not, and the ones that circulate most widely tend to be the ones that survive fewest — the same pattern as confident output that no component in the pipeline is positioned to question.

Sources: EESI, Data Centers and Water Consumption · Forbes, How Much Water Does AI Use