Uniform airflow becomes difficult above 40 to 50 kW per rack. AI cabinets now run from 50 kW to 250 kW, with megawatt-scale designs approaching. Thomas Steen, chief technology officer at Nortek Data Center Cooling, told Data Centre Magazine that liquid cooling is the only scalable path for those loads.
The physics support him. Air systems at high rack density fight recirculation, stratification and local hot spots. GPU thermal throttling turns poor heat removal into lost compute. Repeated temperature swings also stress solder joints and boards as semiconductor substrates and copper expand at different rates.
Direct-to-chip systems remove the largest heat sources through cold plates. They leave memory, storage, networking, power conversion and other components to the air. The room still needs airflow. Operators still need humidity control, filtration and heat removal for the residual load.
That residual percentage matters more as total rack power rises. Ten percent of a 250 kW rack is 25 kW. A hall filled with liquid-cooled AI racks can still carry an air-side load that would have looked substantial in the previous generation. The mechanical design must size both loops against actual component allocations, not a liquid-cooled label on the server.
This is why hybrid cooling has become the default architecture for high-density facilities. CDUs, manifolds and facility water move GPU heat. CRAHs, fan walls or in-row equipment handle everything the cold plates miss. The rack roadmap can climb toward 1 MW while the air system remains on the equipment schedule.
A hybrid plant adds interfaces. Operators manage coolant chemistry, leak detection, pump redundancy and approach temperatures on the liquid side. They also manage filters, fan power and airflow paths on the air side. Control sequences must prevent one system from masking a failure in the other.
The procurement question is no longer whether a facility uses air or liquid. It is how much heat each system owns at every rack configuration. Require vendors to provide that split at normal load, peak load and degraded operation. Then commission both systems together.
Liquid cooling carries the AI buildout. Air cooling remains on the invoice. Any design that treats residual air as an afterthought will find its hot spots after the GPUs arrive.