AI Data Centers Use Water, But the Real Story Is More Complicated

The water question hanging over AI infrastructure has spawned two irreconcilable camps. One warns that data centers will "drain the great lakes" and worsen droughts across the country. The other, led by OpenAI CEO Sam Altman, dismisses concerns as "insane" and disconnected from reality. Neither side is being honest.

The core problem: measuring data-center water use should be straightforward, but it isn't. "We don't have the real ground-truth numbers on data-center water use to put everyone on the same page," according to Eric Masanet, a sustainability researcher at UC Santa Barbara who studies data centers. Data-center operators and utilities release little public information, and what exists can be selectively interpreted to support nearly any conclusion.

On a national scale, the numbers seem manageable. In 2023, U.S. data centers consumed slightly more than 17 billion gallons of water for cooling—equivalent to annual consumption by roughly 500,000 Americans, or less than one-tenth of one percent of total U.S. farm water use. But aggregate figures obscure local vulnerabilities.

A Meta facility under construction in Lebanon, Indiana, could demand 8 million gallons of water daily at peak capacity. That's more than double the town's current peak water demand. In Newton County, Georgia, a local official stated plainly: "we just don't have the water" for all incoming data centers. Yet in the same breath, city officials in Indiana said planned infrastructure upgrades mean they see "nothing to indicate" a future water shortage.

The variability reflects a fundamental truth: every data center is contextual. Water stress depends on local climate, regional water supply, grid composition, and facility design. A data center taxing a drought-stricken region poses genuine risk; the same facility in a water-abundant area with expandable utilities might cause negligible strain. "Outside of water-stressed regions, it's not really a major concern for now," according to Fengqi You, an energy-systems expert at Cornell. "But as the sector grows, who knows?"

Technically, many of AI's largest data centers employ "closed-loop" cooling systems that recirculate water through air chillers rather than expelling it via cooling towers. OpenAI's Stargate facility in Texas, Meta's Hyperion in Louisiana, and Microsoft's Fairwater installations operate this way for most of the year. These systems theoretically don't consume water—it cycles indefinitely without evaporation losses. That's the basis for Altman's argument that water concerns miss the mark.

But closed-loop systems carry a trade-off. Air chillers consume 10 to 65 percent more electricity than cooling towers to dissipate equivalent heat. One proposed nine-gigawatt facility in Utah would demand nearly as much power as New York City. Most of that electricity comes from combustion turbines—many built by tech companies themselves to bypass grid delays—and power generation requires water too.

That's where the indirect water picture shifts dramatically. In 2024, Meta's indirect water consumption tied to power generation reached 19 billion gallons, over 23 times its direct water use, with most routed to data centers. Few industries consistently report these indirect figures, leaving the full environmental footprint opaque.

New York Governor Kathy Hochul's recent statewide data-center moratorium cited "massive amounts" of water as a justification, signaling that policymakers see this as substantive enough to warrant regulation. Yet the evidence remains fragmented, context-dependent, and vulnerable to politicization in both directions.

The honest answer to "Does AI waste a ton of water?" is: not necessarily, sometimes, and it really depends. Sweeping claims—whether apocalyptic or dismissive—flatten a landscape where location, design, and timing determine real impact.