UC Riverside walks back AI's viral water-per-prompt figure

The eye-catching number that has powered two years of AI-water reporting is the one where a 100-word ChatGPT reply costs about a bottle of water, and The Atlantic is the latest outlet to revisit whether that framing survives contact with what the researcher who coined it now says.

That researcher, Shaolei Ren at UC Riverside, is reportedly walking the figure back for current systems. His group's revised estimate for a GPT-4 prompt is closer to 15ml, or roughly 5ml if you count only water used inside the data center rather than water at the power plants generating its electricity. OpenAI's Sam Altman has said an average query uses about 0.3ml, and Google has put a typical query at around five drops. Take the specific numbers as reported, not settled: the gap between bottle and drops is mostly an argument about which scope you count. Researchers describe on-site cooling as Scope 1, off-site power generation as Scope 2, and the supply chain, roughly 2,200 gallons of ultra-pure water for a single microchip, as Scope 3.

Where the story does not shrink is at the facility level. A single large hyperscale data center can use up to five million gallons of water a day, comparable to a town of tens of thousands of people. Roughly one in five US data centers draws from watersheds that are already stressed. A Meta facility in Newton County, Georgia has been reported to use about 500,000 gallons a day, roughly ten percent of the entire county's water. Globally, Ren's group projects AI-driven withdrawals of 4.2 to 6.6 billion cubic metres a year by 2027, on the order of half the United Kingdom's annual freshwater withdrawal.

There is a credibility overhang shaping the newer, more hedged reporting. Karen Hao's book Empire of AI cited a Chilean government document that mixed up liters per second with cubic meters per hour, an error that inflated a Google site's projected use by roughly a factor of a thousand and has since been corrected. Independent critic Andy Masley walked through the miscalculation in a widely cited blog post. That does not vindicate the industry; it explains why per-prompt shock numbers are getting quieter.

The honest caveat is that most operators still do not publish Scope 1 and Scope 2 figures at the facility level, which is exactly what a county water board would want. If per-prompt guilt is fading as a lever, siting policy is where the fight moves next: which watershed the next hyperscale gets built above, and who gets to see the number.