OpenAI CEO Sam Altman stated on the Alex Heath podcast that the water used for approximately 38,000 ChatGPT queries is equivalent to the amount needed to produce one almond in California, and mentioned that data centers face a reputation issue rather than a water usage problem.

He noted that the water consumption of modern data centers is roughly comparable to that of large office buildings, citing a Virginia 2024 committee report indicating that most facilities use no more water than large office buildings. He previously wrote that an average query consumes about 0.34 watt-hours and approximately 0.000085 gallons of water (about 0.32 milliliters, roughly one-fifteenth of a teaspoon). Google claims that Gemini text prompts use about 0.26 milliliters. A 2019 study by the U.S. Geological Survey estimated that the average water footprint for one almond in California is about 3.56 liters. Business Insider stated it could not independently verify the precise conversion of 38,000 queries.

The team led by Shaolei Ren at the University of California, Riverside, previously collaborated with media to provide a higher water usage per prompt, later revising the GPT-4 level to about 15 milliliters (about 5 milliliters in the data center), stating that the earlier “one bottle of water” metric is outdated. The discrepancy depends on whether cooling from power plants, training amortization, and ultra-pure water for chips are included. Large-scale data centers can still use millions of gallons daily, with peaks occurring on hot days, and community water networks are designed for peak rather than average queries.

Altman acknowledged that water usage was higher during the era of evaporative cooling, now emphasizing closed-loop or low-water cooling. He argued that the public feels no guilt about consuming a handful of almonds but is sensitive about queries, indicating a cognitive asymmetry. Community opposition to new data centers revolves around water rights and electricity prices, not teaspoon arithmetic.

Single milliliters and one nut can be juxtaposed, but overall installations are still priced according to power stations and cooling towers. When metrics are not standardized, the juxtaposition favors dissemination.

In market mechanisms, buyers are cloud and model companies that need to build data centers, while sellers are water and electricity from drought-stricken areas. The driving force is that site selection conflicts have become memes. Funding comes from electricity costs and capital expenditures for recirculating water systems. Beneficiaries are platforms that can reduce disputes to almond metaphors in public relations, while the pressured parties are municipalities that need to expand pipelines for peak days.

Source: Public Information