NodalZen

Blog · News

Liquid cooling and the water question

· 7 min read · cooling, water, liquid-cooling

Why hotter AI racks pushed liquid cooling into the mainstream, and what the numbers say about how much water data centers use.

Two stories about AI data centers often get mixed up. One is about liquid cooling, the shift from fans to pipes. The other is about water, and whether data centers drink too much of it. They are related, but liquid cooling does not automatically mean more water use. This post untangles them.

Why racks got too hot for air

A rack is the tall cabinet that holds servers. Its power, measured in kilowatts (kW), tells you how much heat it gives off, because nearly all the electricity a chip uses turns into heat.

For years most racks were modest. In its 2025 survey, the Uptime Institute, an industry research group, found that rack densities were rising slowly, driven by more racks between 10 and 30 kW, and that few facilities went above 30 kW (Uptime Institute, July 2025).

AI hardware is in a different league. Nvidia's GB200 NVL72 rack design, for example, draws about 120 kW, which the company described in 2024 as roughly double the rack designs of the time, and it is cooled by liquid piped directly to the chips (eeNews Europe, October 2024). Uptime Institute analysts wrote in 2025 that AI hardware roadmaps promise 200 kW per rack or more within a few years, and that such densities would be difficult, expensive or even impossible to cool with air alone (Uptime Institute Journal, April 2025).

How liquid cooling works

Water and similar coolants carry heat far better than air. It is the same reason a car engine has a radiator full of coolant and not just a fan.

In direct-to-chip cooling, a metal plate called a cold plate sits on each hot chip, and coolant flows through it. The warm coolant goes to a heat exchanger, which passes the heat to the building's cooling system. The loop next to the chips is usually sealed, so the same liquid goes round and round.

Adoption is growing but still partial. In Uptime's 2024 cooling survey, 22% of respondents said their organization used some direct liquid cooling, 61% said they would consider it, and nearly half of the users said it covered fewer than 10% of their racks (Uptime Institute, May 2024). Uptime's analysts expected AI training to lead the shift, with ordinary business computing following more slowly (Uptime Institute Journal, April 2025).

Where the water actually goes

The heat still has to leave the building. Many data centers use evaporative cooling: water is sprayed or trickled through cooling towers, part of it evaporates, and evaporation carries heat away, just as sweat cools your skin. That evaporated water is gone, which is what people mean by water consumption.

There is also water you never see. Many power plants use water for their own cooling, so every kilowatt-hour a data center buys carries an indirect water footprint.

Lawrence Berkeley National Laboratory estimated that US data centers directly consumed about 17 billion gallons of water in 2023, and that the indirect footprint from electricity was about 12 times larger. The lab projected that these totals could double by 2028, as reported by Latitude Media (Latitude Media, August 2026).

The local picture can matter more than the national one. A 2026 study posted on arXiv by Yuelin Han, Pengfei Li, Adam Wierman and Shaolei Ren estimated that if 2024 water-use intensity persisted, US public water systems would need 697 to 1,451 million gallons per day of new capacity through 2030 to serve data centers, compared with about 1,000 million gallons per day for New York City's supply (arXiv, March 2026). That is the authors' scenario, not a measurement.

The trade-off between water and power

Here is the catch: evaporation is an energy-efficient way to cool. Cutting water often means using more electricity, which in turn carries its own indirect water use.

Microsoft offers a concrete example. In December 2024 it described a design for new AI data centers that cools chips with liquid in a closed loop, filled once during construction, with no evaporation for cooling. It said this would avoid more than 125 million liters of water per year per data center, at the cost of a nominal increase in annual energy use. It planned pilots in Phoenix, Arizona, and Mt. Pleasant, Wisconsin, in 2026, with those sites expected online in late 2027 (Microsoft, December 2024).

So liquid cooling by itself is neutral on water. What decides the water bill is how the heat finally leaves the site, and how the electricity is made.

More measuring, still patchy

Uptime's 2026 survey release said more than half of operators now record their water consumption (ITWeb, July 2026). That also means a large share still do not, which is one reason water estimates vary so much.

Try it in NodalZen

In Size, switch your cluster between air and liquid cooling and see how the cooling and total power change. The Learn lessons cover how cooling works in plain language.

Sources

Learn the basics Open the planner