A single 10 MW data center’s cooling system uses roughly 230,000 tons of water a year — about 60 million gallons — and 20–50% of it is dumped down the drain. As AI drives compute (and the heat that comes with it) to new highs, that water footprint is becoming one of the industry’s hardest constraints. This is the AI water problem, and it’s worth understanding in detail.
How much water does a data center actually use?
The water isn’t going into the servers — it’s going into cooling. Most large facilities reject heat with evaporative cooling towers, which work by letting a portion of the water evaporate. That evaporation is what makes the cooling cheap and efficient, but it also means a continuous draw of fresh water.
For a 10 MW facility, that draw adds up to roughly 230,000 tons — about 60 million gallons — every single year. Multiply that across a campus, or across the build-out the AI boom is driving, and the numbers get large fast.
Why cooling towers dump 20–50% of their water
Evaporation leaves the dissolved minerals behind. As the water cycles through the tower, those minerals concentrate, and left unchecked they scale up the equipment. To keep the “cycles of concentration” in check, operators deliberately dump a portion of the loop — called blowdown — and replace it with fresh makeup water.
That blowdown is typically 20–50% of the water moving through the system. It’s not waste in the sense of a leak; it’s an engineered discharge. But it’s still water bought, used once to carry minerals away, and then sent to the sewer.
Why the obvious fixes don’t scale
The two reflexive answers both run into walls:
- “Just stop using evaporative cooling.” Dry and adiabatic cooling use far less water, but they cost roughly 1.9–2.3x more to run and hit efficiency. Evaporative cooling is the cheapest, most efficient way to reject heat — operators aren’t using it by accident.
- “Just use reclaimed water.” Treated wastewater (TSE) works where supply comfortably exceeds demand — but it has hard geographic limits. In the dense data-center corridors where demand is concentrated, the reclaimed supply often isn’t there.
Chasing new water supply is a treadmill. The more compute you add, the more makeup water you need, and the harder each incremental gallon is to source.
The fix: recover the water you already have
There’s a fourth option: recover the cooling water on-site instead of dumping it. If you can continuously pull the dissolved minerals back out of the loop, you no longer need to dump water to control concentration — you keep circulating the water you already have.
That’s what NONA RECOVER does, using a process called ICP — Ion Concentration Polarization, developed at MIT. Unlike reverse osmosis, ICP doesn’t force water through a membrane at high pressure; it pulls the minerals aside with a low-pressure electric field, so there’s no membrane to foul. It bolts onto your existing cooling loop, works with your existing tower chemistry, and recovers up to 90% of the water that would otherwise be blown down. (For the full mechanism, see ICP vs. reverse osmosis.)
ICP isn’t a lab concept: it was developed at MIT, its development was funded by the U.S. Army for field water production, and the first industrial RECOVER pilot is running today with Tracy Renewable Energy.
What this means for site selection
Water is quietly becoming a gating factor for where data centers can be built. Once you can recover most of your cooling water on-site, the calculus changes — circularity beats scarcity. You stop competing for new supply, you cut makeup-water and sewer costs, and you de-risk the sites where water, not power, is the binding constraint.
The AI water problem is real and growing. But the answer isn’t to abandon the cooling that works — it’s to stop throwing the water away.
Want to see the numbers for your site? Estimate your recovery and savings →