AI Data Centers Can Cut Water Use — Just Not Where You'd Guess

Every time you ask an AI chatbot a question, it runs on rows of computers in a warehouse-sized building that gets very hot. There are two main ways to cool those buildings. Evaporative cooling uses water that literally evaporates away, a bit like how sweating cools you down. Air cooling uses fans and chillers instead, more like a giant version of the fan in your laptop. But there's a hidden third water cost: the power plants generating all that electricity need water too, mostly for steam and cooling.

Researchers at the University of Michigan built a detailed computer model to track both kinds of water use at the same time. They tested it on proposed data centers in Michigan. The result was a trade-off. Air-cooled data centers cut their own on-site water use roughly in half compared with evaporative cooling. But they need more electricity, and that extra demand pushed indirect water use up by about one-third, because the power plants supplying the electricity consumed more water.

The really important finding is about location. Older methods of counting water use blame the data center's own neighborhood. But in reality, most of that water is consumed wherever the power plants sit — which can be in completely different watersheds, the areas that drain into a shared river or lake. When the team mapped this out, they found water stress rising in places far from the data center sites, in communities that typical corporate reporting would never mention.

The takeaway: you can't judge a data center's environmental footprint by looking at the building alone. The authors argue that companies, planners, and regulators need to model water and electricity together before choosing where to build and what cooling system to install. As AI keeps expanding, this quiet water cost is likely to become a bigger fight in towns hosting — or powering — the next wave of data centers.