How much water does a ChatGPT query really use? Sam Altman's almond answer, examined

Sam Altman, on the Sources Podcast with Alex Heath this week, gave a specific number for the water cost of a ChatGPT query: 38,000 queries use about as much water as a single California almond. He followed it with a second comparison: a modern large data center uses about as much water as an office building (Tom's Hardware). The second claim doubles as a rebuttal to a viral line Altman called a "robust meme." Neither number is settled, and the more useful question is which "water" Altman is actually counting.

Altman said, "for every 38,000 ChatGPT queries, that is the same amount of water that is used in the production of a single almond in California," then added "I don't have the exact calculation in front of me" and "doing this from memory, it might be wrong, but it's close" (Tom's Hardware). The arithmetic is a comparative argument, not a measured number, and it is also a specific pushback. The viral claim Altman is responding to: "every time you run a single ChatGPT query, it's like, you know, you run your shower for like six hours. And the water never comes back." Altman called that line "difficult to disprove" but said it "does not hold up to any scrutiny."

The 38,000-per-almond math can be cross-checked against a number Altman himself published earlier. In his blog post "The Gentle Singularity," Altman put the per-query water cost at roughly 0.32 milliliters. That puts 38,000 queries at about 12 liters, which sits near the working figure for a California almond, which a USGS Landsat study places in the same rough order of magnitude. The comparison lands inside a defensible band, but the underlying intensity figure is still Altman's own, and the band is wide enough that an order-of-magnitude claim like "38,000 queries = one almond" tells the reader the figure is small, not what it is.

The office-building comparison is partly a real cooling-technology shift and partly a narrow slice of the footprint. Altman said, "if you look at a modern very large data center, it uses the equivalent amount of water as an office building in terms of, you know, people like running the sinks and the toilets and whatever." That is partly a real shift: evaporative cooling was once common at hyperscale sites, and Altman acknowledged "at one point" it was used but said "they have not done that in a long time." The office-building comparison captures only on-site operational water (drinking, sanitation, the small leak from a closed-loop system) and skips two other layers that any "total true water accounting" has to include.

Those layers are the water embedded in the electricity the data center draws, the water embedded in construction and lifetime maintenance of the facility, and the regional water context of where the data center sits. A Business Insider piece on the same exchange noted that the broader AI-vs-water debate is "hard to dispel" because the embedded-energy and regional layers are not on Altman's chart. A modern data center can use office-building amounts of tap water and still sit in a watershed where any new large draw changes the local balance, and the electricity that powers it is water-cooled at the plant that generated it. Both of those layers are real, and neither is visible in the office-building analogy.

The almond unit is the right shape for a clip: small enough to be graspable, large enough to make the point that the figure is not zero. The office-building unit is the right shape for a CEO defending a buildout, folding a real cooling-technology shift into a comparison that lets the listener file the rest of the debate under "overblown." Both moves are defensible from inside a single layer of the water footprint. Neither survives the "total true water accounting" Altman invoked to introduce them.

The 38,000-per-almond figure is a clean number in a way the comparison is not. The reader's job is to know which layer a future claim is sitting in, and to ask what was left out, not to settle the meme, but to keep the perimeter in view.