Operators publish only the water they use on the hottest summer day. The average draw and the off site water tied to a campus's power plants, cooling, and chip supply never enter the math that decides where new campuses can be built.
For a permitting meeting, a data center can show exactly how many gallons of water it will pull on a 100-degree afternoon. That number is real, but it is also the most flattering number in the file. The figure that actually determines whether a region's rivers, aquifers, or treatment plants can absorb a new campus is the average and embedded water footprint. By current convention, that figure does not have to appear in the application.
Site selection for hyperscale data centers has become a board-level question for utilities, municipalities, and AI builders, and water is the constraint that is suddenly binding. The math that decides who gets built, however, is built around peak cooling demand on the hottest day of the year, the moment a campus is designed to handle. Shannon Markham, a vice president and regional manager for water at AECOM, told Area Development that the projections utilities use to plan capacity are typically peak scenarios: hottest day, full compute load, and maximum cooling stress. Average daily use is rarely the planning number.
Peak framing flatters two places at once. It flatters the operator, whose worst-case number looks modest compared with legacy industrial users. It also flatters the siting region, because the same peak number is small relative to a reservoir or river system's design allocation. John Newsome, the water administrator for Columbus, Ohio, explained to Area Development why his system could absorb more load: Midwestern utilities like his were built to serve heavy industry that has since shrunk, so the same pipes now sit below the capacity they were sized for. For cities chasing tax base and power capacity, that under-used capacity reads as a competitive edge.
The framing collapses, though, when you look at what is not in the permit. Independent modeling by Xylem and Global Water Intelligence, summarized by TechTimes, estimated that on-site evaporation from cooling will account for only about 4 percent of the additional water AI will pull globally through 2050. The remaining 96 percent sits outside the data center walls: in the power plants feeding the campus, in semiconductor fabrication for the chips, in the local distribution losses that move water from treatment to tap, and in the construction supply chain. None of that has to appear in the disclosure.
The cumulative footprint shows up in places nobody requested a permit for. The Ecologist, reporting on a July 2026 Environment Agency analysis, warned that seasonal spikes from data centers already pose a risk to water-stressed regions even where industry claims supply is sufficient. Google's own June 2026 disclosure, reported by Axios, shows roughly two-thirds of its data centers use evaporative cooling. The remaining third is split across air-cooled designs and recycled or non-conventional sources. That mix is the actual portfolio, and it is what determines whether a regional cluster of AI campuses drains an aquifer through a multi-year drought.
Disclosure standards are starting to bend. In June 2026, Google released detailed facility-level water data for its data centers, accounting separately for recycled and non-conventional water sources, per Axios. Western utilities have begun tightening allocation rules for any large new user, including data centers. Recycled-water systems long used in parts of the West are spreading eastward, per the Area Development interviews with Markham and Newsome.
The harder question is whether the average and embedded number ever joins the disclosure. The University of Georgia's CAES Field Report TP121 on data centers and surface and ground water, drawing on the EPA's 2026 framework for water-impact analysis, treats the cumulative footprint as the unit of analysis. Most hyperscale siting regimes do not. Until those regimes treat the cumulative number as the planning number too, the gap between what gets measured and what gets built will keep driving the next round of siting fights, not because data centers use more water than they say, but because regulators and communities are still signing off on less than they will actually draw.