Cooling & Thermal Management

The density of the machine you plan to run dictates the cooling regime, because heat leaves the chip only as fast as your thermal-resistance stack allows — and the air-to-liquid boundary is a cliff.

Air cooling stays viable to roughly 40–50 kW per rack; beyond that the fan-power curve sets an economic ceiling that decides which loads keep air and how far you push it.

Rear-door heat exchangers and air-assisted liquid let a building never plumbed for DLC host ~50–100 kW racks today, bought with dew-point discipline and a ceiling you hit again within one GPU generation.

Past the air ceiling, DLC is the design basis; the open decisions — single- vs two-phase, tray coupling, and the flow, delta-T, and pressure budget — each fix a lifetime serviceability and capex bill.

Immersion wins on PUE and density yet loses on deployment — single-phase survives as a heat-reuse niche, two-phase stalled when its PFAS fluid supply collapsed, and direct-to-chip took the 2026 rack.

The CDU firewalls the clean GPU loop from the building's dirty water; where you place it — in-rack, row, or central — sets blast radius and whether a single pump trip throttles the job.

The facility-water supply temperature — chilled W17/W27 or warm W40/W45 — is the plant's master setpoint, fixing your compressor hours, your water use, and whether waste heat is worth selling.

Heat rejection cashes out your loop temperature against the climate, and the choice among chiller, dry cooler, wet tower, and adiabatic hybrid trades kilowatt-hours against liters every hour it runs.

Whether your waste heat sells as a district-heating commodity or gets dumped depends on its temperature grade, which the facility-water setpoint fixes upstream before construction.

A liquid retrofit is bounded by four quantities the original building fixed — floor strength, plenum volume, electrical headroom, available water — and whichever runs out first sets your density ceiling and strands the rest.

A liquid-cooled hall survives only if its thermal budget closes at the worst-case branch, its leak and loss-of-flow failures trip in seconds, and commissioning finds both before the GPUs do.

A direct-to-chip loop faces two transients — a heat step buffered for tens of seconds, and a flow loss that trips the junction in seconds. Tune the controls against the right one.

Once a hall is plumbed for liquid, cooling becomes a pressure-system problem, where the wrong code, an unmodeled surge, or a galvanic couple can shut down a 132 kW rack through a pipe.