Data centers use less than 1% of America's water. The towns fighting them aren't wrong.

A hand-painted sign in a small New Mexico town reads "Water for people not AI." The slogan shows up at water-board hearings and county commission meetings from west Texas to southern Arizona, wherever a new data center project is moving through permitting in a basin that has spent the last decade fighting drought.

Researchers estimated that US data center cooling consumed about 66 billion liters of water in 2023, well under 1% of total national demand, according to an Ars Technica analysis of public numbers. The 500-milliliter-per-query figure that circulates online is a contested example rather than a measured claim, and it should not be the headline of the conversation. The 66-billion-liter number, in turn, has been used in the same public hearings to argue that the protests are overblown, which is a use the underlying data does not actually support.

Most data centers cool by evaporating water, the same mechanism as a backyard swamp cooler, because the processors that train and serve AI models run hot enough that internal rack temperatures can reach 176°F. The water that disappears into the air has to come from somewhere, and in counties that have already cut residential use, watched reservoirs drop, and let fields go fallow, a single hyperscale project's withdrawal rate lands in the same room as the people.

What would settle the local-versus-national argument is per-site, per-month water withdrawal, but most hyperscale operators don't publish that figure. The ones that do use different scopes, different counterfactuals, and different definitions of "water-stressed," and the resulting numbers are what Fengqi You, an energy systems expert at Cornell University, calls "incomplete and inconsistent." In a basin that already runs dry, an undisclosed draw is a draw that nobody can plan around. The national 66-billion-liter aggregate, useful as it is, sits on top of that same fragmented reporting layer, which means "less than 1%" is more a measurement artifact than a settled fact, and the local-versus-national argument is being run, in public, on numbers that don't quite exist.

The engineering menu to change the picture is already on the table, and most of it isn't speculative. Air-cooled and dry-cooled architectures can replace most or all of the evaporative draw, with a heat-rejection penalty paid in electricity rather than water. Closed-loop liquid cooling keeps the same water recirculating inside the data center rather than bleeding it off to the atmosphere. Water replenishment projects, where an operator funds basin-level restoration that returns more water than the site withdraws, are being contracted. Siting is a lever in its own right: a project placed near a wetter region, near a coastline, or near a heat sink that can absorb the load without evaporation has a fundamentally different water profile than one sited in a closed basin. Each option has a cost, mostly in capex or electricity, but the price difference is small compared with the cost of fighting a project in court for a decade.

The on-site target You names is net-zero water for the cooling system, meaning a site that withdraws no net freshwater from the local basin for cooling. That isn't a research milestone; it's a build decision. The path runs through architecture (less evaporation, more closed loop), through siting (avoiding basins that are already over-drawn), and through replenishment contracts that turn the remaining draw into a net positive. None of this is hypothetical. All of it is being done at individual sites, and the open question for the next hundred projects is whether the same standard gets applied in the counties where the protests are loudest.

The hand-painted signs are not a fringe position. They are a demand that the engineering that can deliver net-zero on-site cooling also show up at the sites where the buildout is landing now.