Datacenter · Compare
Air cooling vs liquid cooling for data centers: which to spec
Rack density decides it: air handles conventional loads, liquid is the answer once GPU racks pass what air can physically move.
Short answer
Pick air cooling when your racks stay within the range air can physically carry and pick liquid cooling once density climbs past that. The single deciding factor is rack heat load. Air moves heat at roughly 160 CFM per kW at a 20 F rise, and AI and GPU racks now pull 40, 80, and past 100 kW, more heat than any volume of air you can push through a rack. Below that ceiling, well-managed air is cheaper and simpler; above it, direct-to-chip or immersion liquid is not optional. Most current AI halls run a hybrid: cold plates on the GPUs with contained air handling memory, drives, and power supplies.
Air cooling vs Liquid cooling: side by side
| Factor | Air cooling | Liquid cooling |
|---|---|---|
| Density ceiling | Practical limit per rack; air cannot carry the 40 to 100+ kW of AI and GPU racks | Direct-to-chip handles 40 to 100+ kW; immersion goes higher, cooling densities no air rack reaches |
| How heat moves | Cold supply to server inlets, hot exhaust back to units without mixing | Coolant carries heat off cold plates or a dielectric bath in direct contact with the silicon |
| Upfront cost / complexity | Lower; conventional CRAC/CRAH, tiles, containment; fixes are cheap housekeeping | Higher; CDU, two loops, manifolds, cold plates, leak detection, or a tank and fluid inventory |
| Commissioning | Airflow test-and-balance: inlet map, delta-T, tile tuning; re-run as the floor changes | Flush to a cleanliness target, pressure test, prove leak detection, balance flow per node before hardware goes in |
| Failure consequence | A bad balance leaves a hall inefficient or strands capacity | A leak drips coolant onto GPUs worth six to seven figures per rack; debris chokes a cold plate |
| Energy / free cooling | Bounded by mixing losses; over-provisioning wastes fan energy | Warm-water cooling (ASHRAE W-classes, coolant in the 30s C) lets the plant free-cool or economize much of the year |
| Maintenance | Blanking panels, tile layout, plenum housekeeping, DCIM trending | Coolant chemistry sampling, filter changes, heat-exchanger approach, leak-sensor upkeep the owner inherits |
| Serviceability | Slide a rail-mounted server out dry in a minute | Cold plate uses dripless quick-disconnects; immersion means lifting a hot, dripping node from oil |
| Governing standard | ASHRAE TC 9.9 inlet envelope; AABC/NEBB balance | ASHRAE TC 9.9 water classes, OCP, manufacturer coolant/cleanliness spec; NFPA 75/30 for immersion fluids |
Which should you pick?
Choose Air cooling when
- Rack density sits within what air can carry and there is no near-term jump to AI or HPC loads
- A hall is alarming but the chillers have spare capacity; the fix is usually sealing and airflow, not liquid
- You want the lowest-cost, simplest system that a facility team can service dry with standard skills
- Retrofitting an existing raised-floor or overhead hall where re-plumbing servers is not justified
Choose Liquid cooling when
- GPU or HPC racks pull 40 to 100+ kW, past what any airflow can remove
- Energy is the driver: warm-water supply lets you free-cool or economize for much of the year
- You need the highest density per square foot and can accept the fluid-handling and serviceability regime (immersion)
- The project can fund and commission flushing, leak detection, flow balancing, and coolant chemistry up front
Bottom line
It depends on rack density, and for most halls the honest answer is both. Below the point where air runs out of capacity, air cooling is cheaper, simpler, and serviceable, and many hot-hall problems are airflow mixing that sealing and tile tuning fix without adding cooling at all. Once racks pass what air can physically move, liquid is the only path, and it brings a heavier commissioning and maintenance burden: clean loops, proven leak detection, balanced flow, and coolant chemistry the owner inherits. Large AI deployments today mostly run hybrid, cold plates on the GPUs with contained air on the rest of the rack, so the real question is how much of each, judged against the actual rack heat load and the ASHRAE TC 9.9 targets.
FAQ
Is liquid cooling always better than air cooling?
No. Air cooling is cheaper, simpler, and easier to service when rack density stays within what air can carry, and many hot halls are suffering airflow mixing that sealing and tile tuning fix without adding any cooling. Liquid becomes necessary once racks pull more heat than air can physically move, around the 40 to 100+ kW of AI and GPU racks. Match the method to the rack heat load.
At what rack density do you need liquid cooling?
Air moves heat at roughly 160 CFM per kW at a 20 F rise, and it runs into a physical ceiling as density climbs. AI and HPC racks now pull 40, 80, and past 100 kW, more heat than any volume of air pushed through a rack can remove, which is where direct-to-chip or immersion liquid takes over. The exact crossover depends on the hardware and the design, so verify against the equipment spec and ASHRAE TC 9.9.
Does liquid cooling save energy compared to air?
It can, mainly through warm-water operation. A chip cools fine on coolant in the 30s of degrees C because the junction-to-coolant difference stays large, so the facility can reject heat to outdoor air with an economizer or dry cooler for far more hours of the year and often skip mechanical refrigeration for the liquid load. Air cooling, by contrast, loses efficiency to mixing and over-provisioning. The ASHRAE water class the design targets sets the ceiling.