Cooling & Thermal Management

Cooling follows a rack's heat split, heat flux, airflow limits, coolant conditions and climate, not a rack-kW threshold. Size the structure, water and power that are hard to retrofit for the next rack generation, or pay to change them under a running hall.

Air cooling has no universal rack-kW ceiling: it often gives way when fan energy and floor space per kilowatt cost more than a rear door or cold plate. Judge racks by their airflow and inlet limits, not kW, or you strand capacity or overbuild.

A rear door or air-assisted liquid loop can get dense GPU racks into an air-cooled building in months, not years. You pay with a dew-point margin held for the life of the hall and a density ceiling you will hit again within one GPU generation.

Direct-to-chip liquid cooling (DLC) is the design basis when your rack needs liquid at the source. Single- or two-phase, coupling type, coolant, flow, delta-T and the air tail you leave behind then set lifetime capex and operating cost.

Immersion can lead on heat capture and on reuse yet lose on deployment, earning production only where it beats supported cold plates on the same brief. 3M’s end-2025 PFAS manufacturing exit removed one supplier; in February 2026 Chemours qualified a replacement fluid in 2CRSi servers.

An L2L CDU separates the clean GPU loop from building water, whose chemistry and pressure may not match. Its placement, in the rack, the row or a central plant, sets the blast radius and whether one pump trip throttles the job.

The facility-water setpoint, set with the rack's TCS limit, is the master fork. Chilled water holds component margin but costs compressor hours in warm weather; warm water earns economizer hours and possibly sellable heat, but some future tray or memory configuration may not accept it.

Heat rejection converts your loop temperature into a climate bill. The choice among chiller, dry cooler, wet tower and adiabatic hybrid trades kilowatt-hours against liters of water every hour the plant runs.

Waste heat sells as district heat only when its temperature grade, the recoverable duty and the customer's demand line up. Fix the grade upstream, then size the lift and the cooling path that runs with no offtake at design stage, while changes are cheap.

A liquid retrofit is capped by whichever of the building's floor strength, routes, electrical headroom or heat paths runs out first. That limit sets the complete-rack count; capacity above it is stranded, in every migration state as well as at the end.

A liquid-cooled hall works only if its worst branch stays under the GPU throttle limit at design weather, and every leak or loss of flow is caught and contained inside its isolation zone. Miss either and GPUs throttle or a synchronous job restarts.

Direct-to-chip loops see two transients: a heat step that circulating coolant buffers, and a flow loss that only locally coupled fluid and solids can absorb. Tune controls for the first, test protection against the second; heat imbalance and equipment limits decide how many seconds remain.

Once a hall is plumbed for liquid, cooling becomes a pressure-system problem: the wrong code, an unmodeled surge or a galvanic couple can take HPE's 132 kW GB200 NVL72 rack offline.