Google, Microsoft, and Amazon report only on site cooling water. The larger upstream volume, evaporated at power plants, goes uncounted, and Berkeley Lab says it averages 12 times the direct number.
Most of the water that the AI buildout consumes does not appear in the sustainability reports. Among the major U.S. tech companies building AI data centers, only Meta discloses the water its suppliers use to generate the electricity that runs its facilities. Microsoft, Google, and Amazon report only what evaporates on-site. That single accounting choice determines whether a company's published footprint is a fraction of actual consumption or the real thing.
The gap is not theoretical. In its 2025 sustainability report, Google said it consumed 10.9 billion gallons of water, almost all of it for cooling, a 34 percent jump from 2024. The figure does not include the water consumed at the coal, gas, nuclear, and hydro plants that feed those data centers electricity. According to a 2025 paper by Alex de Vries-Gao at VU Amsterdam, summarized in the Hacker News discussion of the WSJ report, Google's indirect water use is roughly three times its direct number. Apply that ratio and the company's actual 2025 water cost is closer to 30 billion gallons, well beyond the headline number. The Hacker News thread on the WSJ article noted that even the larger figure, expressed in agricultural terms, would be enough to irrigate tens of thousands of acres of alfalfa for a season.
The asymmetry is structural, not accidental. No U.S. law requires companies to report the full scope of their water use, both direct and indirect. The disclosures that do exist vary by methodology, which is why direct comparisons of "total water" between Microsoft, Google, and Amazon are not apples-to-apples. The most quoted benchmark is from Lawrence Berkeley National Laboratory, whose 2024 analysis found that, historically, indirect water consumption for U.S. data centers runs about 12 times the direct figure. That average has been corroborated across multiple outlets, including Honolulu Civil Beat, Truthout, and Water Security Newswire, though none ran an independent primary study, and the 12x number is a historical average, not a per-operator current-year ratio.
Why the indirect number swings so widely comes down to the local power grid. Coal and nuclear plants consume large volumes of water for cooling, natural gas uses less, and solar and wind use almost none. A data center in West Virginia drawing from a coal-heavy grid leaves a far larger upstream water footprint than an equivalent facility in a wind-rich region. That grid dependence is the second reason the disclosure gap matters: a company that reports only direct water can still draw most of its real water cost from a region it does not name.
The scale of what is being built makes the loophole newly consequential. Microsoft, Google, and Amazon are among the tech companies spending an estimated $1 trillion on AI infrastructure in 2024 and 2025, what experts called the largest U.S. infrastructure expansion in history. Consumer Reports has connected that buildout to rising electric bills and water stress, and NPR has documented how hard the true water footprint is to verify, because every layer of the supply chain is allowed to keep its own books. The Lincoln Institute has framed the boom as a land-and-water impact story, and researchers at the United Nations University have warned that AI's combined carbon, water, and land footprint is already pressuring natural resources for billions of people.
For communities hosting new data centers, the math is what matters. The Most Policy Initiative's science note on data-center water use puts the regional pressure in plain terms: every gigawatt of computing capacity brings its own upstream water demand, and the regions absorbing that demand are often already short. That is the level where the disclosure gap becomes a siting decision. Counties approve projects on environmental impact statements that count only the on-site number, while the larger volume is paid for elsewhere, in watersheds that may not know they are part of the deal.
What would change the picture is straightforward: a reporting standard that asks companies to disclose indirect water alongside direct, a way to attribute that water to the grid region it actually came from, and disclosure at the level of individual sites rather than only the company-wide aggregate. Meta's water report already includes the upstream number. The others could, too. Until they do, the headline figures, from Google's 10.9 billion gallons to whatever Microsoft and Amazon publish next year, will continue to describe the smaller half of the water the AI buildout actually uses.