Air cooling is not obsolete and liquid cooling is not universal. The decision is set by how much heat a single rack must reject, and by whether the building can deliver water to it.
Air is a density question before it is a cost question
Air-cooled halls remain appropriate for storage, networking, and general compute. The constraint is the volume of air a rack can move and the containment discipline required to keep supply and return separated as rack loads climb. Accelerator racks are where that budget runs out.
Liquid moves the heat path, not just the heat
Once a cold plate or a tank captures processor heat, the facility inherits a fluid network: distribution, pumps, filtration, water chemistry, and a plant that has to reject heat at a temperature the servers dictate. That network is the real project, and it is why liquid is planned with the compute rather than after it.
- Rack and row loads at the worst-case processor configuration
- Supply and return temperatures the plant can hold year-round
- Residual air load the room still has to carry
- Service access, leak response, and technician training
Most halls end up hybrid
A mixed hall carries liquid-cooled accelerator racks alongside air-cooled infrastructure, which means two heat paths, two control strategies, and one plant. Treating hybrid operation as the design case rather than a transition state avoids stranding capacity in either direction.
Where the transition usually happens
Conventional air cooling systems with disciplined hot and cold aisle containment are generally comfortable up to the low tens of kilowatts per rack. Between roughly 30 and 50 kW, most data center operators need assistance at the rack, which is where rear-door heat exchangers and in-row units appear. Above that, the volume of air required stops being practical and liquid capture at the processor becomes the default. These are planning bands rather than hard limits, and the honest version of the threshold is site-specific: it depends on ceiling height, containment quality, floor pressure, and how much of the hall is already committed.
Where the energy savings actually come from
Liquid cooling is often justified on efficiency, but the saving is not a single line item. Removing server fans cuts electrical load inside the rack. Warmer supply water raises free-cooling hours, which cuts compressor energy in the plant. Shorter air paths reduce the work the room's air conditioning has to do. A liquid design that still needs cold water and full room air handling captures very little of this, so the efficiency case has to be made against the whole cooling system rather than the cold plate alone.
- Server fan power removed from the rack
- Free-cooling hours gained by raising supply temperature
- Residual air handling still required in the hall
- Pump and distribution energy the liquid loop adds back
The options are a spectrum, not two choices
Framing the decision as air against liquid hides the range of cooling technologies available between them. At one end sit room air conditioners and computer room air handlers, working with hot and cold aisle containment to cool the whole room. Rear-door heat exchangers cool the hot air as it leaves the server racks. Direct-to-chip cooling captures heat at the processor. Immersion cooling removes the air path almost entirely. Each step up that spectrum raises achievable density and cooling efficiency while asking more of the building and the operations team, and most data center cooling solutions in service today combine two or three of them.
What changes for the operations team
Traditional air cooling fails gradually and visibly: temperatures drift and staff have time to react. A liquid cooling system introduces a failure mode that is faster and wetter, so thermal management becomes a procedural discipline. Leak detection, isolation, coolant chemistry, quick-disconnect handling, and server-swap practice all have to exist before the first production rack is energized, and the maintenance contract has to say who performs each of them.
Compare on the same load, climate, and redundancy
An air and a liquid proposal are only comparable when both answer the same schedule: identical rack loads, the same ambient design points, the same one-failure condition, and the same commissioning states. Ask each bidder to name what their number excludes.