The water question: how do you cool AI in a thirsty world?

Every answer your AI gives carries a hidden cost, and it isn't only measured in cents or kilowatts. It's measured in water. A large data center can drink as much as a small city, most of it evaporated into the sky to keep the chips cool. In a desert, that arithmetic looks impossible. Which is exactly why the desert ends up forcing the right answer.

The thirst, in numbers

The figures are sobering. A large data center can consume up to 5 million gallons of water a day, comparable to a town of 50,000 people. US data centers directly consumed about 17.4 billion gallons in 2023, a number projected to reach 38 to 73 billion gallons by 2028 (EESI, Brookings).

One state shows the trajectory in miniature: a study found Texas data centers would use roughly 49 billion gallons in 2025, potentially rising to 399 billion by 2030. Unsurprisingly, water has become a flashpoint, with community backlash reported in the US, Chile, Ireland and the Netherlands (Lincoln Institute).

Why most data centers drink so much

The thirst is a design choice, not a law of nature. The cheapest way to reject heat is evaporative cooling: spray water, let it evaporate, carry the heat away with it. It works, and it's inexpensive, in a place with water to spare. But evaporation means the water is gone, not recirculated, and at the scale of an AI campus that becomes millions of gallons a day vanishing into the air over a community that may have wanted it for something else.

So the water problem is really a cooling-architecture problem. A facility built around evaporative towers trades cheap heat rejection for enormous water consumption. That bargain only looks good where water is abundant and unpriced, and that describes fewer and fewer places every year.

The answer the desert forces: a closed loop

If you can't evaporate water, you have to keep it. That means closed-loop liquid cooling: a sealed circuit that carries heat directly off the chips and rejects it without continuously consuming fresh water. Direct-to-chip and immersion designs do exactly this, and they're also what the densest modern accelerators require anyway, since you cannot air-cool a 150 kW rack. The water-frugal path and the high-density path turn out to be the same path.

Here's the twist the backlash stories miss. A desert facility built on a closed loop can be dramatically more water-efficient than a conventional evaporative plant in a temperate, water-rich region, because it was never allowed to waste water in the first place. The location everyone assumes is worst for water can, designed correctly, be among the best. The constraint produced the better building.