Liquid Cooling vs. Air Cooling for AI-Density Racks

Above roughly 20–30 kW per rack, air cooling usually can’t remove the heat without extraordinary airflow, and GPU racks for AI routinely run at several times that. That is the point where designs move to liquid cooling: direct-to-chip cold plates, immersion or rear-door heat exchangers.

Air cooling works by moving enough cold air across a server to carry its heat away, and it scaled fine for decades of data center growth. The problem is physics, not engineering effort: air has a much lower heat capacity than liquid, so as rack power density climbs, the volume of air — and the fan power to move it — needed to keep pace grows faster than most facilities can practically deliver. Racks running well above roughly 20–30kW commonly exceed what air cooling can remove without extraordinary airflow, and modern GPU racks for AI training and inference routinely run at multiples of that.

Direct-to-chip cooling

The most common liquid cooling approach in new deployments: cold plates mounted directly on the highest-heat components — CPUs and GPUs — circulate coolant that carries heat away at the source. Lower-heat components on the same board (memory, storage, networking) are usually still air-cooled, making this a hybrid approach rather than a fully liquid-cooled server. It's the least disruptive liquid-cooling retrofit path for an existing air-cooled hall, since the room-level air system can often stay in place for everything the cold plates don't cover.

Immersion cooling

Submerges entire servers in a dielectric fluid that doesn't conduct electricity, removing heat from every component at once rather than just the highest-heat ones. Single-phase immersion circulates the fluid in liquid form throughout; two-phase immersion uses a fluid that boils at the operating temperature, carrying heat away as it changes phase and condenses for reuse. Immersion removes heat more thoroughly than direct-to-chip but requires a fundamentally different rack and server form factor — not a bolt-on retrofit to standard rack-mount hardware.

Rear-door heat exchangers

A hybrid that sits between air and liquid cooling: a liquid-cooled heat exchanger mounted on the back of a standard air-cooled rack captures hot exhaust air before it reaches the room, rather than cooling components directly. Servers inside the rack stay conventional and air-cooled internally — only the room-level heat rejection changes. It's a lower-disruption way to raise a facility's effective density ceiling without redesigning server hardware, though it doesn't reach the density levels direct-to-chip or immersion can support.

Where the CDU fits in

A coolant distribution unit (CDU) is the interface between the facility's main cooling loop and the rack-level liquid cooling loop, transferring heat between the two — often through a heat exchanger — while keeping the fluids and pressures on each side independently managed and monitored. It's become a standard electrical/mechanical BOM line on new data center builds rather than a specialty item, tracked on the Data Center Power Equipment pillar page alongside transformers and switchgear.

What this means for PUE

Liquid cooling generally lowers PUE — moving heat with liquid is more efficient than moving it with fans and mechanical chillers — but the mechanism matters: rear-door heat exchangers still lean on the room-level cooling plant behind them, while direct-to-chip and immersion cut the amount of heat that plant has to handle in the first place. The market for liquid cooling overall is forecast to keep growing at roughly 20% a year as direct-to-chip deployments become standard on new high-density builds.

Frequently asked questions