Heat Rejection Architectures

capture · transport · reject

Every watt a site consumes ends as heat that must leave continuously. Liquid capture at the load is now mandatory for AI racks; the open choice is how the site rejects that heat, and water decides it.

Heat rejection as three layers. Capture, where heat leaves the load: room air through CRAH and CRAC units, rear-door heat exchangers (air to water), direct-to-chip cold plates, or immersion in dielectric fluid. Transport: coolant distribution units, the facility water loop, plate heat exchangers and N+1 pumps. Reject, where heat leaves the site: dry coolers using near-zero water, hybrid wet/dry coolers that spray at peaks, evaporative cooling towers with the highest water use, or heat reuse through heat pumps to a district or process heat buyer. CAPTURE: WHERE HEAT LEAVES THE LOAD Room air CRAH, CRAC units Rear-door HX air to water Direct-to-chip cold plates Immersion dielectric fluid Transport: CDUs, facility water loop, plate HX, pumps N+1 REJECT: WHERE HEAT LEAVES THE SITE Dry coolers air only near-zero water Hybrid wet/dry spray at peaks seasonal water Cooling towers evaporative highest water use Heat reuse via heat pump heat sold to buyer
NVIDIA GB200 NVL72 rack
~120-132 kW
Air-cooled rack ceiling
~40-50 kW
Redundancy floor
N+1
Design point
Worst-case ambient
Hub
Supply Chain

Finding

Heat rejection is designed in three layers - capture, transport, reject - and AI density has forced liquid capture while water scarcity is pushing rejection toward dry and hybrid coolers

NVIDIA's GB200 and GB300 NVL72 racks draw roughly 120-140 kW each and ship only with direct-to-chip liquid cooling; room air stops being practical somewhere around 40-50 kW per rack. Rejection is a separate decision. Evaporative towers give the most capacity per dollar on hot days but consume large volumes of water; dry coolers use almost none but lose capacity as dry-bulb temperature rises.

Large AI campuses are converging on closed-loop liquid cooling with dry or hybrid rejection: OpenAI and Oracle's Stargate site in Abilene, Texas uses closed-loop systems filled once rather than evaporative towers. The same three-layer logic applies to megawatt charging depots (liquid-cooled cables and power cabinets), liquid-cooled BESS containers such as Tesla Megapack, CATL TENER and Sungrow PowerTitan, and gigafactory dry rooms. Equipment comes from Vertiv, Schneider Electric (which bought Motivair), Trane, Johnson Controls, Carrier, CoolIT, Boyd and LiquidStack in the West, and Envicool, Inspur and Sugon in China.

Classes

Architecture classes

Heat rejection architecture classes
ClassBest fitKey constraintWatch-outs
Air-cooled (dry coolers, air-cooled chillers)Water-constrained, moderate-density sitesCapacity falls at high dry-bulbFootprint, fan power, noise
Water-cooled (towers plus chillers)High density, large steady loadsWater supply, treatment, permitsBlowdown, plume, Legionella control
Hybrid wet/dryPeak capacity with low annual waterComplexity, higher capexControls tuning between modes
Liquid-to-liquid primary loopDense loads needing stable delta-TLeak management, pump redundancyFiltration, materials compatibility
Direct-to-chip liquidAI racks and HPCSupply temperature control, manifoldsServiceability; water-glycol vs dielectric
ImmersionUltra-dense, tight footprintTooling and vendor ecosystemFluid handling; component compatibility
Heat pump plus reuseSites with a heat buyerNeeds stable heat demandEconomics depend on COP and utilization

Selection

Selection matrix

Design for worst-case ambient and degraded states, not nameplate.

Heat rejection selection matrix
Decision axisTypicalEdge caseArchitecture response
Heat densitykW per zoneMW per zone, high fluxLiquid loops; direct-to-chip or immersion
Water availabilityAbundant, permittedConstrained or uncertainDry or hybrid; reclaimed water
Ambient extremesMild climateHot and humid or hot and dryMore exchanger, chiller and tower capacity
Load variabilitySteadySeconds-to-minutes transientsThermal buffering; faster controls
Expansion speedSlow, plannedRapid modular growthStandard cooling blocks and skids
ReliabilityBusiness-gradeAlways-onN+1 to 2N; separate failure domains

Stack

Subsystems and minimum instrumentation

Thermal system stack
SubsystemCore componentsInstrumentation minimum
Heat captureCold plates, immersion tanks, CRAH/CRAC, power-electronics cold platesSupply and return temperature, flow, differential pressure, leak detection
Primary liquid loopPumps (N+1), strainers, expansion tanks, valves, manifoldsPump status and vibration, delta-T per branch, valve position
Heat exchangePlate-and-frame exchangers, intermediate loops, economizersApproach temperature, fouling, pressure drop
Heat rejectionTowers, dry coolers, chillers, fansWet-bulb and dry-bulb, condenser pressure, fan and pump power
Water treatmentFiltration, softening, dosing, blowdownConductivity, biocide confirmation, makeup and blowdown flow
ControlsBMS/SCADA, chiller and tower sequencing, fault detectionTime-synced logging from load to rejection

Sources

Sources and scope

Scope
Heat Rejection Architectures
Sources
NVIDIA GB200/GB300 NVL72 specifications; operator disclosures; ASHRAE TC 9.9
Updated
October 2026