CRAHs: Computer Room Air Handlers

Imagine a server room floor built as a giant hollow plenum, with chilled water arriving from a chiller, covered elsewhere on this site, located far away in a central plant, and a machine sitting inside the room whose only job is to pass that chilled water through a coil, blow the now-cold air out under the raised floor, and let it rise up through vents in front of server racks. That machine is a computer room air handler, CRAH.

A CRAH is deliberately different from an older, related device, the computer room air conditioner, or CRAC, which carries its own self-contained refrigeration compressor and cooling cycle inside each unit, whereas a CRAH simply circulates water that has already been chilled centrally, meaning a single large, efficient central chiller plant can serve dozens of CRAHs across a facility rather than every room running its own small, less efficient refrigeration cycle independently.

Understanding this distinction genuinely matters because it explains why the industry's shift toward liquid cooling, covered elsewhere on this site, has been disruptive rather than incremental, well-designed CRAH-based air cooling tops out at roughly 30 to 40 kW of heat removal per rack, while a modern Nvidia GB200 or GB300 rack now draws upward of 130 to 140 kW, meaning CRAHs and air cooling generally simply cannot keep pace with the newest AI accelerator hardware regardless of how well the room's airflow is engineered.

That single mismatch, air cooling's practical ceiling against AI hardware's actual heat output, is the direct reason liquid cooling penetration inside AI data centres has been projected to jump from around 14% in 2024 to roughly a third of new deployments by 2025 and higher since, a shift covered in full elsewhere on this site, with CRAHs remaining relevant mainly for the lower-density parts of a facility rather than the AI compute halls themselves.