The carbon footprint of large language models gets a lot of attention. Their water footprint gets much less. A paper by Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren, first posted on arXiv in April 2023 and published in Communications of the ACM in 2025, tries to put a number on it.
Where the water goes
Data centers produce a lot of waste heat. Most of them get rid of it with evaporative cooling towers, which consume water by turning it into vapor. That is the on-site water. The power plants that generate the electricity for those servers also consume water, which the authors count as off-site water. A third scope, water used to manufacture the servers and chips, sits in the supply chain and is the hardest to measure.
Freshwater is not an unlimited resource. The paper points out that severe water scarcity already affects about 4 billion people, and that the gap between global supply and demand could reach 40 percent by 2030.
How they estimated it
Because companies publish little about their water use, the authors built a bottom-up model from the thermodynamics of data center cooling. Total water is the sum of on-site evaporation and off-site water for electricity. The model depends on where and when the computation runs, since cooling efficiency changes with the weather and the local energy mix changes by the hour.
Applied to Microsoft’s data centers in the United States, the model estimates that training GPT-3 consumed about 700,000 liters of clean freshwater. That is roughly the water needed to build 370 BMW cars or 320 Tesla vehicles. Training the same model in a hotter Asian climate would have used about three times as much, around 2.1 million liters.
For everyday use, the authors estimate that a conversation of roughly 20 to 50 questions with ChatGPT uses about a 500 mL bottle of water. They also note a tension between carbon and water: scheduling work for when solar power is abundant lowers emissions but raises evaporation, because it moves computation into the hottest part of the day.
What they propose
The paper recommends what it calls an “unfollow the sun” approach, scheduling large training runs during cooler hours and using battery storage to balance carbon and water goals. It also asks for liquid cooling and alternative water sources such as greywater or seawater, and for transparency: water usage effectiveness reported in model cards, and supply-chain water disclosed alongside carbon.