AI Data Centers Are Rising in America's Driest Places

The argument over AI and the grid has been loud all year. The argument over AI and water has been quieter, and it may turn out to matter more. New figures suggest the industry is building its thirstiest infrastructure in exactly the places least able to spare the water.

Around two-thirds of 809 planned AI data-center projects in the United States are set for areas that have seen some of the most severe drought conditions of the past year, according to an analysis covered by Tom's Hardware and EnviroLink. That siting collides with a country where drought has touched close to 63 percent of the land in recent months.

The scale of consumption is real, even if the exact numbers come with caveats. By one estimate from the research firm Mordor Intelligence, AI data centers consumed roughly 264 billion gallons of water in 2025, near a trillion liters, and are drawing on the order of 550 million gallons a day, as Barchart reported. Projections for US AI servers run to 200 to 300 billion gallons a year through 2030.

Where the water goes

Most of it goes to cooling. Many large facilities rely on evaporative cooling, which sprays or evaporates water to carry heat away from racks of hot chips. It is effective and, in hot weather, hungry. Researchers estimate a single large data center can use millions of gallons a day under the right conditions, which is why a cluster of them in a stressed watershed worries utility operators and local governments.

The honest caveat is that water use varies enormously by design. A facility using closed-loop or air cooling, or one sited near reclaimed water, has a very different footprint from one drawing municipal drinking water for evaporative towers. The trouble is that disclosure remains patchy, so communities often cannot tell which kind of neighbor they are getting until the permits are signed.

Not the same as the electricity question

It is worth separating this from the power debate. We recently looked at the claim that AI is single-handedly spiking household electricity bills, and found it overstated. Water is a different problem with a different shape. Electricity can be generated more cleanly or shipped across a grid. Water in a drought-stricken basin cannot be conjured, and a gallon a data center evaporates in Arizona is not available to a farm or a town downstream.

None of this means the build-out stops, or should. AI infrastructure is being treated as strategic, and the data centers are coming. The open question is whether they get sited and cooled with the local water table in mind, or whether the planning catches up only after a few dry summers force it. On current evidence, the siting is running ahead of the caution.