Cooling & Thermal Management

Cooling is selected from the named rack and facility envelope: rack kW is one input alongside liquid/residual heat split, component heat flux, airflow and inlet limits, TCS/FWS conditions, room rejection, climate, serviceability, redundancy, and the refresh tail.

Air cooling remains viable only where the named rack's airflow, inlet, pressure, heat-flux, containment, acoustics, climate, rejection, service, redundancy, and refresh envelopes close; rack kW alone does not set an economic or physical ceiling.

RDHx and air-assisted liquid can extend a brownfield hall only where the named rack, door or sidecar rating conditions, airflow/inlet envelope, water interface, residual-room rejection, climate, service/redundancy case, and refresh tail all close.

DLC is the design basis where the named equipment requires liquid capture at source; single- versus two-phase, coupling architecture, coolant, flow, delta-T, pressure drop, residual-air duty, and service/redundancy choices then set the lifetime operating and capex consequences.

Immersion can win on heat capture and reuse yet lose on deployment: 3M’s end-2025 PFAS manufacturing exit removed an incumbent source, while Chemours’ February 2026 qualification in 2CRSi servers opened a specific replacement path; the bath earns production floor only when its hardware, fluid and service contract beat the supported cold-plate alternative on the same operating brief.

An L2L CDU firewalls the clean GPU loop from building water whose chemistry and pressure can differ; where you place it — in-rack, row, or central — sets the blast radius and whether a single pump trip throttles the job once surviving flow, external head and heat rejection are counted.

The facility-water setpoint is the master fork: chilled buys component margin and adds compressor duty where the climate demands it; warm buys economizer hours and potentially sellable heat while spending the margin the next silicon generation will want — close the FWS and TCS operating points together against the rack and heat sink.

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 the temperature grade set upstream, the recoverable duty and the customer’s coincident demand; size the lift and the no-offtake cooling path before construction, while those interfaces are still cheap to choose.

A liquid retrofit is bounded by the original building’s floor strength, plenum and routes, electrical headroom, and available liquid and air heat paths; whichever verified capacity runs out first caps the complete-rack count and strands the rest, including during the migration states you must keep running.

A liquid-cooled hall survives only if its thermal budget closes at the worst branch and its leak and loss-of-flow response preserves thermal and containment limits through each required service or fault state; prove the local timing before the GPUs find the gap, then hand that evidence to Chapter 13.5 for installed acceptance.

A direct-to-chip loop faces two transients — a heat step buffered by circulating inventory, and a flow loss that leaves the junction relying on locally coupled mass. Tune the controls against the first and qualify local protection against the second; only the actual heat imbalance and equipment envelope tell you how many seconds remain.

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.