The Atlantic Examines AI Data Centers' Real Water Use

AI's water story has spent the past year swinging between two extremes, and The Atlantic's latest look at how much water data centers actually use lands in the middle of a debate that recently got a lot more complicated.

The turning point was a very public correction. Karen Hao, whose 2025 book Empire of AI reported that a proposed Google data center in Cerrillos, Chile could require 'more than one thousand times the amount of water consumed by the entire population,' acknowledged the figure was off by a factor of 1,000, tracing it to a mix-up between liters per second and cubic meters per hour in a Chilean government document. She thanked researcher Andy Masley, who identified the error. The corrected number is still substantial, roughly 104.5% of the residential water used by the nearby city of Cerrillos in 2019, but nowhere near the viral 1,000x claim that had circulated across coverage of AI's environmental cost.

Where does that leave the bigger picture? By one widely cited estimate, the total water use of all U.S. data centers is less than 0.5% of American freshwater use, and Cornell researchers project global AI-related water withdrawals could exceed 6 billion cubic meters per year by 2027, a lot in absolute terms, still small next to the roughly 3 trillion cubic meters agriculture consumes annually. Per-query estimates remain contested: one figure has a 100-word AI response consuming roughly 519ml, while OpenAI's Sam Altman has said an average query uses about 0.3ml. Take the specifics as reported, not settled.

The honest caveat is that national averages hide local pain. In Chile, a court partially revoked the environmental permit for the Google site citing aquifer risk and climate concerns, and Google later redesigned the project with less water-intensive cooling. The story didn't go away after the number shrank. Similar dynamics are showing up in the U.S. Southwest, where a Microsoft data center campus in the Arizona desert was the subject of Hao's original Atlantic investigation.

What the reporting doesn't give you is a clean single number both sides trust, or a map of which U.S. jurisdictions will follow Chile's lead. The upside for readers is that the loudest numbers are finally getting audited. For utilities, hyperscalers, and communities negotiating new campuses, that means the fight will increasingly turn on measured local draw and closed-loop cooling redesigns rather than viral comparisons, which is where the policy conversation belongs anyway.