The Atlantic Digs Into the AI Data Center Water Numbers

The debate over how much water AI actually consumes has become one of the most contested arguments in tech reporting, and The Atlantic is the latest outlet to work through the numbers. The gap between the loudest estimates is enormous. A widely cited UC Riverside estimate puts a 100-word ChatGPT response at about 519 ml of water, roughly a bottle. OpenAI's public counter is that a typical query uses far less, in the fraction-of-a-milliliter range. Both cannot be right at the same time, and the gap comes down to what you count.

Researchers now describe AI's water use in three scopes. Scope 1 is the on-site water for cooling the servers. Scope 2 is off-site water used to generate the electricity the data center consumes. Scope 3 is embodied water in the supply chain, most visibly the roughly 2,200 gallons of ultra-pure water used to make a single microchip, as Undark laid out in its own recent breakdown of the same debate. Companies tend to speak in Scope 1 terms, which produces the tiny per-query figures. Independent researchers include Scope 2, and that is where the numbers balloon.

The more useful frame, and the one the reporting keeps returning to, is that national totals are misleading. Total US data center water use is a small share of overall American freshwater withdrawals, but the local picture is uneven. Roughly one in five US data centers draws from watersheds that are already stressed, and a Meta facility in Newton County, Georgia has been reported to use about 500,000 gallons per day, roughly ten percent of the entire county's water. That is the kind of concentration where a national-average defense stops being reassuring.

The honest caveat is that the viral end of this coverage has a credibility problem. Karen Hao, whose earlier Atlantic reporting helped popularize the "AI is guzzling water" framing, publicly acknowledged that a figure in her book about a proposed Google data center near Santiago, Chile was off by roughly a factor of a thousand due to a unit misunderstanding. Take the specifics as reported, not settled, and note what the coverage still does not give you: consistent, audited, site-level disclosure from the hyperscalers themselves.

The forward-looking piece is the cooling technology. Liquid cooling in sealed loops can operate with near-zero fresh water consumption, and Nvidia has been pitching its newer data center designs on exactly that basis. If those designs deploy at scale, the argument shifts from asking how bad AI water use is in the aggregate to asking which specific sites are still evaporating drinking water in the wrong watersheds, which is a much more tractable question for regulators and local officials than the current all-or-nothing shouting match.