Air cooling vs Direct-to-chip liquid
The rack decides, not a kW rule of thumb. Vendors publish how each rack's heat divides between liquid and air. HPE rates GB200 NVL72 at 132 kW nominal TDP: 115 kW liquid + 17 kW air, and GB300 NVL72 also sends most of its heat to liquid. A rack with a liquid share needs cold plates and a coolant distribution unit (CDU), plus room air for the rest. An all-air rack can stay on air for as long as the room delivers its airflow at its inlet temperature limit.
| Axis | Air cooling | Direct-to-chip liquid |
|---|---|---|
| Which racks it fits | all-air racks whose airflow and inlet temperature limits the room can meet, with the containment and fan power to match | every rack whose spec sends heat to liquid (GB200 and GB300 NVL72 included), and air racks the room can no longer cool |
| PUE band | the Uptime survey average is 1.52 (an average of respondents, not fleet-weighted) | 1.05-1.15 quoted for advanced liquid-cooled facilities in temperate climates with warm-water loops |
| Coolant temperature | supply air within the ASHRAE class the servers are rated for; chillers or economizers reject the heat | water at the rack's supported supply temperature; ASHRAE W-classes rate the facility-water supply, and free-cooling hours depend on climate and approach temperatures |
| Water use | evaporative cooling uses a lot of water; dry coolers save it and pay in PUE | dry heat rejection can bring site water use near zero, and the warm return water becomes usable for heat reuse |
| Failure modes | fans, filters, containment gaps and blocked airflow; how fast the room heats depends on thermal mass and controls | leaks, fittings, coolant chemistry and loss of flow; prove how fast the rack protects itself on your rack and loop |
| Retrofit | nothing to add: air is the incumbent | CDUs, manifolds, facility water, heavier floor loads and trades most data centers have not staffed before |
| Where it wins | storage and CPU rows, network rooms, edge sites, legacy halls, and air-cooled GPU servers whose limits the room meets | every rack whose spec puts heat on liquid, and any hall planned for the next rack generation |
Quantitative cells are the guide's canonical figures — each is date-stamped and sourced in the numbers register and derived in the chapters below.
How the decision falls
Read the rack's spec sheet first. If it sends part of the heat to liquid, as GB200 and GB300 NVL72 do, plan direct-to-chip cold plates, a CDU and enough room air for the remainder. If the rack is all-air, keep it on air while the room can deliver its airflow at its inlet temperature limit; containment and rear-door heat exchangers (30–40 kW typical RDHx; >50 kW with active fans) stretch that range. Whatever you install now, give the shell the floor strength, pipe routes and CDU space the next rack generation needs, because adding them under a running hall costs far more (Chapter 5.1).
What would flip it: No facility water and no time to add it: a rear door or an air-assisted liquid sidecar can run liquid-cooled racks in an air hall as a bridge, provided the room can absorb the heat the sidecar gives back (Chapter 5.3). An all-air rack whose airflow or inlet limit the room cannot meet moves the other way, to a rear door or to liquid.
Model this fork with your own numbers: Rack cooling check →
Full derivations, worked examples, and the numbers behind this matrix: Thermal fundamentals and the density wall (Ch 5.1) · Rear doors and air-assisted liquid: the bridge (Ch 5.3) · Thermal design, leak detection and commissioning (Ch 5.11) · PUE, WUE, ERF and the post-PUE metric stack (Ch 15.1)