AINEWS
An analysis from sustainability nonprofit Ceres says 78% of that electricity comes from thermoelectric plants — coal, gas, and nuclear — that withdraw water for cooling, and two thirds of those plants sit in water stressed basins.
Data centers in seven U.S. states depend on roughly 3.4 trillion gallons of freshwater a year for the electricity that runs them, and most of that water sits upstream of the campus. A fresh Ceres synthesis puts the dominant share in the thermoelectric power plants that feed the campus its electricity, and argues operators are not yet pricing it into the multi-year power contracts the AI buildout is signing now.
The seven states on Ceres' list host roughly half of all U.S. data center capacity: Virginia, Texas, California, Illinois, Georgia, Ohio, and Arizona. To put a human-scale number on the 3.4 trillion figure, the report frames it as roughly twelve times the combined annual water use of Los Angeles, Phoenix, and Washington, D.C., three cities with their own famously strained supplies. The visible cooling tower is a small piece of the picture. The bigger piece is the steam cycle at the natural-gas, nuclear, and coal plants that still dominate dispatch in most of these states.
Those plants withdraw water to operate. They pull it from a river, a reservoir, or an aquifer, run it through the condenser to turn low-pressure steam back into liquid water, and return most of it to the source. Across the seven analyzed states, 78% of the electricity serving data centers came from plants that withdraw water this way. The cooling tower on the data center campus is, in that sense, downstream of a much larger water tap.
What makes the picture uncomfortable is geography. Two-thirds of the water-using power plants feeding these data centers sit in basins the report classifies as medium-high to extremely high water stress, or in regions already coping with active drought. The states on the list are not arbitrary. Loudoun County in northern Virginia is the densest data center market on the planet. Texas, Arizona, and California are the most drought-tracked states in the lower forty-eight. Georgia and Ohio are running their own basin-allocation fights. Illinois holds critical Great Lakes cooling capacity with its own political constraints.
The report's blunt claim is a governance gap rather than a technology gap. Most power producers already name data centers as the primary driver of rising electricity demand. Few of them, Ceres finds, have folded water risk into the procurement and siting decisions that turn that demand into a power purchase contract. On the other side of the meter, most data center operators do not yet account for the water risk embedded in the electricity they buy. The number has been in the engineering literature for years. It has not been priced into most of the contracts being signed now.
Loudoun's data center moratorium debates, the Arizona groundwater fights around Phoenix exurbs, and the Texas Panhandle siting fights around new natural-gas plants all run on the same water question. They are downstream of the same procurement decision that did not price the water.
Ceres lays out four leverage points where the gap could close. Power producers can build water-risk disclosure into generation planning and into the terms they offer data center buyers. Data center operators can map their contracted electricity against basin-level water stress, and treat water the way they already treat power cost and carbon. State-level planners can require disclosure of the water footprint embedded in any new large load, not just the on-site use. Investors can price water risk into the cost of capital for the buildout.
Discussion on Hacker News around the report's release leaned into the same picture. The comments that took the numbers seriously treated the upstream load as the under-counted one, and treated the on-site cooling debate as a distraction from the bigger planning question. The planning question is not new, but the AI load is. The contracts being signed this year for new data center campuses in these seven states will lock in power purchases that run for ten to fifteen years. The water risk inside those contracts has not yet been written down on the same page as the megawatt.