On January 6, Nvidia CEO Jensen Huang told a Las Vegas crowd that "no water chillers are necessary for data centers" running his newest chips. Shares of cooling companies fell the same day.

Ten months later, on November 11, Modine opened a factory in Texas to build chillers for data centers, including high-density AI sites. One of these moves looks wrong. Neither is, and the reason explains how data center cooling actually works.

How do data center chillers work?

A chiller works like a large refrigerator for water. It pulls heat out of the water that cools the building, then pushes that heat outdoors.

Inside a classic data center, cold water flows through coils in big air handlers. Hot air from the servers passes over the coils and warms the water. Pumps send that warm water back to the chiller. SemiAnalysis, in its guide to data center cooling, describes the chiller as running "a refrigeration cycle," which it calls an energy-intensive process.

A refrigerant inside the chiller absorbs the heat from the water. The heat then leaves the building, usually through a cooling tower or a dry cooler. Here is the full trip:

How heat leaves a data center
Step What happens Equipment
1. At the chip Heat moves into air, or into liquid in a metal plate on the chip Fans, or direct-to-chip cold plates
2. In the room That heat moves into the building's water loop Air handlers or coolant distribution units (CDUs)
3. At the plant The water is cooled further if needed Chiller, or a bypass when outside air is cold enough
4. Outdoors Heat is released to the air Cooling tower (uses water) or dry cooler (doesn't)

Step three is the only optional one. Heat always has to reach step four.

Cooling towers vs dry coolers

A cooling tower releases heat by letting some water evaporate. That works well, but SemiAnalysis notes the water use is "substantial." A dry cooler is a big radiator with fans. It uses no water, but it can only cool things down to near the outdoor air temperature.

Why do chillers use so much power?

Chillers run compressors, and compressors are hungry. SemiAnalysis estimates cooling takes 60% to 80% of a data center's power outside the computers themselves.

Chillers lead that list, followed by fans and pumps. That's why operators look for ways to run them less. One common trick is free cooling. When the outdoor air is cold enough, a bypass loop sends water around the chiller's refrigerant. The outdoor equipment does the work alone. The compressor rests.

What changes with liquid-cooled AI racks?

AI racks pipe liquid straight to the hottest chips. Warmer liquid can still cool those chips, and warmer liquid can be cooled outdoors without a chiller.

In direct-to-chip cooling, a liquid flows through a metal plate pressed against each processor. A coolant distribution unit, or CDU, swaps that heat into the building's water. SemiAnalysis says a CDU is a heat exchanger with pumps and controls, typically sized above 1 megawatt.

Nvidia's June blog post on liquid cooling has the numbers. Coolant enters the rack at 45°C, about 113°F, and leaves at roughly 55°C. That outgoing liquid is warm enough to shed its heat into outdoor air on many days. In favorable climates, Nvidia says, that allows "chiller-less operation with dry coolers." The same post cites roughly 2.6 million gallons of water per megawatt per year for systems built on cooling towers. Dry coolers avoid that.