Why AI's Water-Use Numbers Are All Over the Map

In 2023, a widely cited study estimated that having ChatGPT draft a 100-word email cost roughly 519 milliliters of water — most of a standard water bottle. In 2025, Google measured its own Gemini chatbot's typical text prompt at five drops: 0.26 milliliters. Same basic action, both numbers from credible researchers, a gap of nearly 2,000 times between them.

The confusion isn't a conspiracy in either direction. It's a measurement problem, and understanding it explains more about how AI actually uses water than either headline number does on its own.

The researcher behind the number has already revised it

The 519-milliliter figure traces back to Shaolei Ren, an electrical and computer engineering professor at UC Riverside, whose 2023 paper on AI's "secret water footprint" put the issue on the map. Ren has since said that estimate, based on an earlier version of GPT-4, is outdated: newer models are far more efficient, and his current per-prompt figure, water from cooling plus the water used to generate the electricity behind it, lands closer to 15 milliliters. He's also the one still sounding the alarm. "Every time you ask an AI chatbot a question, you are also consuming water — without realizing it," Ren told a TEDAI Vienna audience in October 2024.

Three different things get counted as "water use"

Every credible estimate is measuring one or more of three separate pools, and most of the public confusion comes from headlines that name only one:

Onsite cooling water — what actually evaporates at the data center to keep servers from overheating. Electricity-generation water — the water power plants use to make the steam that spins turbines, which typically dwarfs onsite use. Embodied water — the water baked into building the hardware in the first place, including the roughly 2,200 gallons of ultra-pure water it can take to fabricate a single high-end chip.

Layer on top of that: which model, how efficient its chips are, how hot the local climate is, whether the plant nearby burns gas or spins wind turbines, and what counts as "consumed" versus merely "withdrawn and returned." Change any one variable and the per-prompt number moves by an order of magnitude, which is exactly why a single data center's water and electricity bill can look nothing like another's a few states away.

Why the aggregate number matters more than the per-prompt one

Zooming out past any single query, the totals are more solid. U.S. data centers directly consumed an estimated 17.5 billion gallons of water in 2023, according to Lawrence Berkeley National Laboratory, well under 1% of the country's total water use. Cornell energy systems researcher Fengqi You's team projects that could roughly double to quadruple by 2030. The number that should worry a given town isn't the national average, though. Bloomberg has found that roughly two-thirds of U.S. data centers built since 2022 sit in already water-stressed regions, and demand tends to spike hardest on the same hot summer days when everyone else's tap is running too — which is also when the local power grid strains hardest.

Ren's proposed fix is almost embarrassingly simple: shift AI training to cooler hours. "We don't water our lawns at noon because it's inefficient," he told the TEDAI crowd. "Similarly, we shouldn't train AI models when it's hottest outside." Cooling-technology upgrades already underway, closed-loop liquid systems that lose no water to evaporation, chips engineered to tolerate hotter coolant, are pushing in the same direction. None of it will settle the per-prompt argument, because there will never be one true number for a footprint that changes by the hour and the zip code. But it explains why two honest experts, using the same industry data, keep landing worlds apart.