The Invisible Trillion · The Water Question, Honestly

The “Water Per Query” Thing

There was a study in 2023 — a good one, by researchers at UC Riverside — that estimated the water cost of an AI query. The number they came up with was that a ChatGPT conversation of 20–50 responses used about 500 milliliters of water, or roughly one plastic water bottle.

This number went everywhere. It is the single most quoted statistic about AI and the environment. It is also more misleading than it appears, for reasons worth understanding.

First: that 500 ml includes both the data center’s direct cooling water and the water used to generate the electricity the data center consumes, upstream at the power plant. Power plants — especially coal and gas plants — use a lot of water for their own cooling. If the data center is running on coal, the water footprint is big. If it’s running on solar or wind, the water footprint is almost zero. The 500 ml figure assumes an average American grid mix circa 2022, which is mostly coal and gas. A data center running on solar could have a water footprint 50 times smaller.

Second: the 500 ml is for a 20–50 response conversation, not a single query. A single query is more like 10–25 ml — one or two tablespoons. The number got repeated as “a bottle of water per query,” which is, at a minimum, off by a factor of 20.

Third: even 500 ml of water per 50-response conversation is a small number in context. One hamburger contains about 1,700 gallons of embedded water — enough water for 13,000 ChatGPT conversations at the high-end estimate. The water you used to shower this morning was enough for several thousand of those conversations.

Fourth — and this is the part that matters most — the water cost of a query depends almost entirely on where the data center is and how its electricity is made. A query answered by a data center in Iowa running on wind-and-solar with closed-loop cooling uses essentially zero fresh water. The same query, answered by a data center in a desert running on gas with evaporative cooling, uses quite a lot. Averaging these together produces a number that describes no actual data center anywhere.

None of this means the 500 ml number is “wrong.” It’s a real estimate of a real, if complicated, quantity. It means that using that number as a cudgel to argue that AI is destroying water supplies is a misuse of the science. The researchers themselves have said as much, in interviews that got much less traction than the original number.

All 19 chapters of The Invisible Trillion · Library · All courses