Air cooling vs Direct-to-chip liquid
This stopped being a preference around 40 kW per rack: past the air-cooling cliff, moving enough air is physically and acoustically untenable, and every current-generation rack-scale AI system (NVL72-class, ~132 kW) ships liquid-cooled. The real questions are which liquid architecture, and what to do with existing air-cooled halls.
| Axis | Air cooling | Direct-to-chip liquid |
|---|---|---|
| Rack density ceiling | practical to ~20–40 kW/rack (rear-door heat exchangers stretch to ~70) | 130+ kW/rack today; 600 kW-class roadmap racks assume it |
| PUE band | ~1.4–1.6 legacy halls; industry weighted average stuck at 1.54 | ~1.05–1.15 in temperate climates with warm-water loops |
| Coolant temperature | chilled air ~18–27 °C supply; chiller plant does the work | warm water (~45 °C supply / up to ~65 °C return class) — enables free cooling + heat reuse |
| Water consumption | evaporative variants consume heavily; dry variants pay in PUE | closed loop consumes near-zero on site; heat reuse (ERF) becomes feasible |
| Failure modes | fan/filter maintenance; thermal excursions are slow | leaks, fittings, coolant chemistry; excursions are fast at 1+ kW/chip |
| Retrofit story | n/a — it is the incumbent | CDUs + manifolds + floor loading + trades data centers haven't staffed before |
| Where it wins | storage/CPU rows, network rooms, edge sites, legacy estates | every dense AI hall — it is no longer optional at current accelerator TDPs |
How the decision falls
For AI compute the fork has already closed: direct-to-chip liquid is the design basis, and air remains for everything around it (storage, networking, low-density rows). The live decision is the liquid architecture — CDU topology, warm-water temperatures, and how much of the plant becomes heat-reuse-capable.
Full derivations, worked examples, and the numbers behind this matrix: The cooling cliff and the technology ladder (Ch 5.1) · The thermal spine: chip to atmosphere (Ch 5.11) · PUE/WUE/ERF — the post-PUE metric stack (Ch 15.1)