US datacenters consumed three times more water in 2024 than they did in 2014, according to new figures that predate the most aggressive phase of AI infrastructure buildout. That timing is the uncomfortable part: the data captures a period before hyperscalers began deploying the current generation of GPU clusters at scale, meaning the real number today is almost certainly higher — possibly much higher.

Siloed reporting is making the problem harder to track. Water usage data is scattered across different state and local agencies, with no federal standardization for how datacenters report consumption, evaporative cooling losses, or source water impacts. That fragmentation means regulators, researchers, and the public are effectively flying blind while the industry’s footprint expands. A facility in Arizona drawing from an already-stressed aquifer might look fine on paper in one report and alarming in another, depending on who filed what where.

The timing matters because cooling is the primary driver. Traditional datacenters lean on air cooling, which sips water compared to the direct-to-chip and evaporative systems increasingly required for dense AI hardware. As Nvidia’s GB200 and upcoming Rubin platforms push rack densities past 100 kW, operators face a choice: burn more water or invest in closed-loop systems that cost more upfront. The market is already seeing early moves toward liquid cooling that recirculates coolant, but retrofitting existing facilities is expensive and slow.

Meanwhile, the EPA is reportedly moving to drop the requirement for public notice on minor-source permits for polluting datacenters, leaving state and local regulators to decide whether communities get any say at all. Put those two developments side by side — consumption data that’s years out of date and a permitting process becoming less transparent — and you have a regulatory environment that’s moving in exactly the wrong direction at exactly the wrong time. The AI boom didn’t create the water problem; it just made the stakes impossible to ignore.