700,000 litres to train, a bottle to talk
On 6 April 2023 researchers at Riverside and Arlington proposed a way to count the water AI spends and applied it to GPT-3. Training in Microsoft's US data centres evaporates 700,000 litres of fresh water on site and a further 2.8 million at the power stations, which is 3.5 million together. A conversation of 20 to 50 exchanges costs about half a litre.
Why it matters
Carbon was being counted by then, water was not: it was treated as an incidental detail of cooling. The work separated direct withdrawal in the cooling towers from indirect withdrawal in generating the electricity, and showed the second to be four times the first. And the bottle-per-conversation figure made visible, for the first time, the price of an answer rather than of a training run.
In Microsoft's Asian data centres the same training would evaporate three times as much on site, and 4.9 million litres once the indirect footprint is added. The indirect footprint is computed from the United States average water efficiency of electricity generation, 1.8 litres per kilowatt-hour, and a data centre efficiency coefficient of 1.2. Half a litre per conversation is not a constant: a data centre's water efficiency changes with the season and the region, and the authors stress exactly this dependence on when and where the model is running. What the record does not claim. The estimate of 4.2 to 6.6 billion cubic metres of global withdrawal by 2027 is not in this version: the strings 4.2, 2027 and Denmark occur in it zero times. It appeared in a later revision in October 2023 and cannot stand under the April date. The GPT-3 figures here are computed from published characteristics, not read off a Microsoft meter.