How do AI data centers stay cool, and how much water does that really use?

Nearly all the electricity a chip uses turns into heat, and a modern AI rack can release about 120 kilowatts of it. The cooling system has one job, which is to carry that heat outside. How it does the job decides whether the site uses a little water or a lot.

Our article on the energy and water behind each AI request covers the cost per prompt. This one looks at the building, because for the people living nearby, the local total matters more than the figure per prompt.

Three ways to get rid of heat

Inside the building, modern AI racks use direct-to-chip liquid cooling, with coolant flowing through plates on the processors. The main difference is what happens to the heat next.

Evaporation is efficient because turning water into vapor absorbs a lot of heat. That water is consumed: it leaves as vapor and does not go back to the local source. A site with a closed loop fills its system during construction and then recirculates it.

What the numbers say

The widely cited totals come from a 2024 Lawrence Berkeley National Laboratory report, as summarized by researchers writing for The Conversation:

So most of the water footprint is usually at the power plant, not the data center. This is why indirect water matters when you compare options, and why company reports that leave it out can look better than they are. The same researchers note that these reports are voluntary and hard to compare.

Single sites vary enormously. Google reported that its data center in Council Bluffs, Iowa, used 3.8 billion liters in 2024, the most of any of its sites, while an air-cooled site in Pflugerville, Texas, used 38,000 liters. The first figure is much bigger than the second, and neither one is the "average" data center.

The trade-off nobody can avoid

Saving water usually costs electricity. Microsoft says so plainly: it began using closed-loop designs for all new data centers in August 2024, saying each would avoid more than 125 million liters a year, and it acknowledges that replacing evaporation with mechanical cooling raises its power usage effectiveness (PUE), the ratio of total facility power to computing power. Its chips can run with warmer coolant, which offsets part of the penalty.

Epoch AI, a research group that tracks AI sites, sees the same pattern in the Stargate project. It reports that at least six of the seven US sites will use closed-loop liquid cooling, chosen in response to public concern about water, and it says this raises costs. It estimates that one Texas site's total power is about 2 gigawatts for 1.4 gigawatts of computing, partly because of the hot climate and closed-loop cooling. Its records for Colossus 2 in Memphis show air-cooled chillers and condensers instead of cooling towers.

More electricity can mean more indirect water and more emissions, depending on what generates it. There is no universally correct choice: a closed loop makes sense where water is scarce and the grid is clean, while the balance may differ in a wet climate with a gas-heavy grid.

A worked example: 100 megawatts

Water efficiency is measured by water usage effectiveness (WUE), in liters per kilowatt-hour of computing energy. In a December 2024 post, Microsoft reported a fleet average of 0.30 L/kWh for its previous fiscal year, and said new designs aim for near zero.

Illustrative. A real site's WUE depends on climate and design, and this leaves out indirect water. At 2.5 people per household and about 300 liters per person a day, an assumption we chose for illustration, 263 million liters would be about the yearly use of 960 households.

What should you ask about water when a data center is proposed near you?

Start with what will be evaporated, not what will be withdrawn. Ask which cooling method the site will use, the expected peak-day water use as well as the annual average, and whether the source is drinking water, recycled water or a private well. Peak days matter, since they fall in the hot, dry periods when a local system is already stretched.

Then ask how much extra electricity the design uses and where it comes from, since a dry design can shift the burden to a power plant. Finally, ask whether the figures will be reported publicly once the site is running. Promises made at the permit stage only help if someone can later check them. New York's executive order, for example, asks state regulators to review whether their water withdrawal program fully captures large users, including data centers.