Data Center Cooling Systems: Air, Liquid, Immersion, and Evaporative

Cooling is the largest non-IT energy load in a data center. As rack densities rise with AI workloads, operators are moving from traditional air cooling toward liquid, direct-to-chip, and immersion designs. Each approach has different power, water, capital, and reliability characteristics.

Last updated: 2026-07-09
Reading time: 8 min

Air cooling (CRAC and CRAH units)

Traditional data centers use Computer Room Air Conditioners (CRAC) or Computer Room Air Handlers (CRAH) to circulate cold air through raised floors and hot-aisle/cold-aisle server arrangements. Chilled water or refrigerant loops remove heat from the airflow.

Air cooling is well understood and inexpensive at low densities (under about 15 kW per rack) but becomes impractical above roughly 25 kW per rack, where the volume of air required is no longer economical.

Evaporative and adiabatic cooling

Evaporative cooling passes outside air across water-saturated media, using the phase change from liquid to vapor to reject heat. It is common in hyperscale facilities in dry climates and typically achieves Power Usage Effectiveness (PUE) values near 1.1–1.2.

The trade-off is water consumption. A large evaporatively cooled campus can consume hundreds of millions of gallons per year, which has driven scrutiny of water use in Arizona, Nevada, and Texas.

Direct-to-chip liquid cooling

Direct-to-chip (also called cold-plate) cooling routes a coolant loop directly onto the CPU or GPU package through a metal plate. The liquid absorbs heat closer to the source than air can, allowing much higher rack densities — typically 40–100 kW.

Most new AI training deployments in 2024–2026 use direct-to-chip cooling for GPUs, with a secondary air loop for lower-density components such as network switches and storage.

Immersion cooling

Immersion cooling submerges servers in a dielectric fluid — either single-phase mineral oil or two-phase engineered fluorocarbons that boil at chip operating temperatures. It supports densities above 100 kW per rack and eliminates fans inside the server.

Adoption remains limited by fluid cost, service procedures, and hardware warranty considerations, but interest has grown alongside GPU deployments.

Free cooling and climate selection

In cool climates, operators can use outside air directly (air-side economization) or via a water loop (water-side economization) for much of the year, reducing chiller runtime. Facilities in the Pacific Northwest, upper Midwest, and Nordic countries rely heavily on this approach.

Climate is a major driver of site selection for new hyperscale campuses, alongside power availability and network connectivity.

Frequently asked questions

What is PUE?

Power Usage Effectiveness is total facility energy divided by IT equipment energy. A PUE of 1.2 means 20% overhead for cooling, lighting, and conversion losses.

Does liquid cooling use less energy than air cooling?

Generally yes, particularly at high rack densities. Liquid transfers heat more efficiently than air, reducing the fan and chiller power required per kW of IT load.

Is immersion cooling safe for servers?

Yes, the dielectric fluids used are electrically non-conductive. Adoption is limited more by operational logistics and hardware warranty terms than by safety.

Why does evaporative cooling use so much water?

Water evaporates to reject heat, and the vapor is released to the atmosphere rather than recovered. Consumption scales with cooling load and outside air temperature.