Sam Altman Says One Almond Costs More Water Than 38,000 ChatGPT Queries. The Math Doesn't Hold Up.

PolitiFact rated the claim Mostly False, finding that Altman's comparison "skews the estimates for both almonds' and AI's water demands" ^[1]. On the almond side, Altman's framing draws on an inflated older estimate; a more direct measure of the water actually poured on an almond crop is closer to 6 liters per almond ^[1][2]. On the ChatGPT side, Altman's own figure of roughly 0.32 milliliters per query is well below independent estimates, which range from 0.6 to 17 milliliters per prompt depending on the model and method ^[1].

Run the real numbers and the picture flips: independent analysis suggests it takes something like 1,000 to 10,000 ChatGPT-style prompts to match the roughly 6 liters of water used to grow a single almond — not 38,000 ^[1][2]. That's a comparison that still favors AI over almonds, but by a much smaller and much less certain margin than Altman claimed.

The larger problem, according to researchers, is that nobody outside the AI companies can actually check any of this. UC Riverside's Shaolei Ren, who studies data center water use, put it plainly: "There is very little public information about how today's ChatGPT systems operate and what resources they use" ^[1][3]. Water draw varies by a data center's location, cooling system, and even how long or complex a given prompt is, which makes any single flat number — Altman's included — essentially unverifiable ^[2][3].

Data center water consumption in the US climbed from about 5.6 billion gallons in 2014 to roughly 17 billion gallons in 2023, with projections putting cumulative use as high as 150 billion gallons between 2025 and 2030 — comparable to the annual water use of millions of US households ^[2][3]. Unlike farm irrigation, that water is largely treated, potable municipal supply, which is part of why California lawmakers are pushing for disclosure in the first place ^[2].

Altman acknowledged his own number might be shaky. He told Heath the figure "might be wrong, but it's close" ^[1] — a hedge that undercuts the confidence with which the comparison was delivered and then repeated across social media as settled fact.

MISLEADING. Altman's 38,000-queries-per-almond comparison overstates almond water use, understates ChatGPT's water use relative to independent estimates, and rests on data center figures that outside researchers say can't currently be verified at all ^[1][2][3]. AI likely does use less water per interaction than growing a single almond — but by a margin far smaller, and far less certain, than the number Altman gave a national podcast audience.