The water problem: what cooling AI actually costs the ground it sits on

Power gets the headlines and the lawsuits. But the resource that quietly decides where a data center can actually be built — and where the buildout runs into a wall it can't buy its way past — is water.

A Gridlas analysis · built from public data · figures 2021–2026, primary sources cited throughout

Our first two analyses were about electricity — the interconnection queue and the workarounds operators use to get power fast. Water is the other meter, and it behaves differently. A cluster that can't get grid power can burn gas on-site or sign a nuclear PPA. A cluster that runs a region's aquifer or river dry has no equivalent escape hatch. Cooling is where AI touches the ground most directly, it is increasingly landing in the driest places in the country, and it is the constraint least amenable to a clever deal. Here is what the public record shows.

1. Cooling is a water decision, not just a power one

Most large data centers reject heat by evaporating water in cooling towers — trading water for electricity, because evaporative cooling is far more energy-efficient than refrigeration. The metric is Water Usage Effectiveness (WUE), liters of water per kWh of IT load. The industry average sits around 1.8 L/kWh; best-in-class evaporative designs reach 0.3–0.7; and sealed closed-loop liquid or air cooling use almost no water on-site at all.

That trade-off is why the number that matters isn't just how much a facility draws, but where the draw lands — because the same gallons mean very different things in Virginia and in Arizona.

2. The direct number is big — and the real number is bigger

The hyperscalers do report direct water use, and it is climbing fast. Google's data centers consumed 6.4 billion gallons in 2023 — up about 50% from 2021 — with roughly 95% of that going to cooling. Microsoft used 1.7 billion gallons, up 34% in a single year; Meta reported about 0.81 billion. The largest individual campuses can draw several million gallons a day, on the order of a town of tens of thousands of people.

And these are the companies that disclose. Multiple investigations — Honolulu Civil Beat, Fast Company, the Global Investigative Journalism Network — have found that most operators reveal little or nothing about site-level water use, so the public totals almost certainly understate the real draw.

3. The compute is landing where the water isn't

The siting data is the uncomfortable part. A 2025 Bloomberg analysis found that about two-thirds of new U.S. data centers built or in development since 2022 sit in places already under high water stress — and just five states account for 72% of them. Two of those five, Arizona and Texas, are among the most drought-exposed in the country.

The flashpoints follow the map. In Memphis, xAI's cooling draws on the Memphis Sand aquifer, the city's drinking-water source. In Phoenix and across the Southwest, campuses compete with cities for Colorado River water. A single proposed Utah campus was reported to need up to ~16 billion gallons a year — more than double Google's entire global 2023 direct use — though such headline figures are contested and often reflect worst-case permitting estimates. The pattern is consistent even where the numbers are fuzzy: the cheapest land and fastest power sit in exactly the basins with the least water to spare.

4. Why water, not power, may decide the map

Put the three Gridlas analyses together and a hierarchy of constraints emerges. The grid queue is slow, but it has exits — the whole point of powering around the queue is that a determined operator can generate its own electricity. Water has far fewer exits. The efficient fixes — closed-loop and air cooling, recycled and non-potable supply — are real and spreading (Microsoft now ships chip-level closed-loop designs), but they cost capex, energy, or both, and they don't help a project already permitted for evaporative cooling in a drought.

So water is likely to become the harder gate. It won't stop the buildout, but it will increasingly decide its geography — pushing new campuses toward the Midwest and the Pacific Northwest, toward recycled-water deals, and away from the sunbelt basins that cheap power and fast permitting made attractive in the first place. The map of AI, in the end, may be drawn less by where the electrons are than by where the water is.