Liquid cooling can remove most of the heat generated by high-density processors without eliminating the need for air inside the rack. Power supplies, network switches, optical modules, fans, storage devices, and other electrically active components still reject heat into the surrounding air unless engineers provide another dedicated heat path. That creates a rack with two thermal behaviors rather than one uniform environment, even when the compute trays appear almost entirely liquid-cooled. A 120 kW rack can contain substantially different thermal loads because the coolant may capture the dominant processor load while a smaller air-cooled load remains concentrated around specific hardware zones. The engineering problem shifts from moving large quantities of heat to controlling smaller but more localized heat sources with sufficient airflow, temperature margin, and service access.

Dual Thermal Zones Inside a Single Rack Envelope

A high-density rack increasingly behaves like two thermal systems sharing one mechanical enclosure. The lower and middle sections can transfer a large portion of processor heat into liquid loops, while power supplies and network hardware continue to discharge heat through air paths. That arrangement creates a liquid-stabilized zone alongside an air-dependent zone, with the interface determined by component placement, fan direction, cable congestion, and the remaining sensible heat load. PSU assemblies deserve particular attention because their internal switches, magnetic components, capacitors, and conversion stages generate heat across a three-dimensional structure that does not always lend itself to direct liquid attachment. A recent experimental study demonstrated why alternative cooling approaches for high-density PSUs remain technically relevant even as processor cooling moves toward liquid, because heat that escapes into room air still increases the burden on the facility air system.

The top section can create a separate challenge because network hardware may operate in a localized thermal environment that differs from conditions elsewhere in the enclosure. Top-of-rack switches concentrate switching silicon, power conversion, fans, and optical transceivers into a relatively compact volume, while dense cable fields can restrict the intended airflow path around their intake and exhaust surfaces. High-speed optical modules can also contribute meaningful local heat, with current 800G modules commonly carrying double-digit-watt thermal envelopes and specified case-temperature ranges that require controlled conditions. That heat does not disappear simply because the compute trays below transfer most of their power through coolant, and the resulting vertical temperature gradient can become more important than the rack’s average temperature. The design implication is straightforward: liquid loops should establish the primary heat path for liquid-cooled compute, while airflow should remain deliberately engineered around every component that still depends on air.

Containment Inversion: When the Hot Aisle No Longer Drives Airflow

Traditional containment logic assumes that servers generate the dominant air-side heat load and that the hot aisle therefore becomes the principal destination for exhausted air. Liquid cooling changes that relationship by removing a substantial share of compute heat before it reaches the room. The remaining air stream can carry far less total heat while still carrying enough localized heat to create hotspots around PSUs, switches, optics, and other air-cooled components. As a result, the hot aisle may no longer represent the primary source of air-side heat across the entire rack when liquid-cooled compute removes a substantial share of the processor heat before it reaches the room. Instead, localized heat sources can influence local airflow and temperature patterns around components that continue to rely on air cooling after the primary compute heat has moved into the liquid path.

That inversion also changes how CRAH capacity should be evaluated at the room level. A lower total air-side load may justify lower fan energy or warmer supply conditions, but those gains do not automatically protect components that receive air through restrictive or indirect paths. A switch positioned near the rack top can encounter warmer air even when the general room condition remains acceptable, particularly when airflow orientation conflicts with the surrounding equipment. Cable openings, blanking gaps, side clearances, rear doors, and cabinet interfaces can then become thermal control points rather than simple mechanical details. The room system must therefore account for where residual heat enters the air stream, where that air travels, and which component receives it next instead of relying only on aggregate sensible cooling demand. Proper containment still reduces unwanted mixing, but the airflow model must increasingly resolve component-level sources and pathways inside the rack itself.

Optical Thermal Marginality in Residual Air Streams

Optical hardware introduces a particularly narrow thermal-management problem because the modules occupy the front panel while relying on the switch’s airflow architecture for heat removal. Current 800G transceivers can consume roughly 15 W or more depending on form factor and reach, while some higher-power variants can approach 30 W, creating a substantial aggregate load when dozens of ports operate simultaneously. Their specified operating ranges do not mean that every temperature within those ranges produces identical engineering margin, because module power, heatsink performance, inlet temperature, airflow direction, and port density interact continuously. A switch can therefore remain within a published component limit while losing useful thermal headroom as local inlet temperatures rise or airflow weakens. The remaining network load still requires controlled airflow after liquid cooling removes a substantial portion of the compute heat, making local airflow and inlet temperature important design variables.

Low-velocity residual air creates another issue because heat removal depends on both temperature difference and mass flow across the heated surfaces. When airflow falls too far, the same wattage can produce a larger local temperature rise because the air spends longer near the heat source and removes less energy per unit time. That condition can develop around optical cages, switch ASIC heatsinks, PSU exhausts, and fan trays even when the broader aisle temperature looks controlled. High-radix switches can present a concentrated thermal-management challenge because dense optical and switching hardware places multiple heat-producing components within the same chassis airflow environment. Consequently, engineers should evaluate the inlet temperature and airflow available at the optical cage itself rather than infer its condition from a nearby rack sensor or room-average value.

Low-Velocity Air Management and Particulate Settlement Risk

Reducing air volume can improve cooling-system efficiency when liquid removes most of the rack’s high-power compute heat, but reduced airflow changes particle transport as well as heat transport. Air velocity influences how particles move through equipment, where they settle, and how effectively filtration and local airflow patterns remove them from sensitive surfaces. Research into electronics cooling environments shows that particulate deposition depends strongly on aerodynamic conditions, including fan-generated turbulence and local flow behavior. That finding matters in a rack where airflow is concentrated around switch bays, PSU intakes, and other restricted ventilation paths rather than distributed uniformly across the equipment volume. Lower flow should therefore not automatically translate into simply reducing CRAH airflow without evaluating the resulting particle residence and deposition patterns. The objective becomes controlled low-volume airflow rather than minimum airflow, with sufficient movement to prevent stagnant pockets around equipment that still depends on air.

Serviceability adds another layer because filters, fans, cable bundles, and removable PSUs gradually change the pressure and flow characteristics of the rack. A configuration that performs well when new can develop localized restrictions as cable density increases or dust accumulates on surfaces, particularly where airflow already operates at a reduced velocity. Engineers should consequently monitor pressure, temperature, and airflow near the actual air-cooled components instead of relying exclusively on CRAH return temperature as the control variable. Thermal sensors near the rack top and rear can identify hotspots that remain invisible in a room-level measurement scheme, while component-level telemetry can reveal whether optical or switching temperatures rise as airflow conditions change. The facility then gains a control loop that can adjust airflow to the residual load instead of maintaining a legacy air volume designed around heat that the liquid system already captures.

Reclassifying Residual Heat as a Primary Design Load

Residual component heat should enter the design process at the same time as rack power, liquid-flow requirements, electrical distribution, and network architecture. A rack that transfers most processor heat into liquid still needs an engineered air path for PSUs, optics, switches, fans, storage devices, and other components that reject heat into the cabinet environment. The critical design question is no longer how much air the entire rack needs in aggregate, but where the remaining watts originate and which airflow path removes them without creating local thermal accumulation. That requires computational fluid dynamics, rack-level thermal testing, component inlet measurements, and explicit consideration of airflow orientation before the mechanical design becomes difficult to change. It also changes CRAH sizing because the facility can potentially reduce bulk air movement while preserving targeted airflow where the remaining heat load actually resides.

The strongest design outcome comes from separating the heat-removal responsibilities without separating them from the same reliability model. Liquid cooling should carry the heat that the liquid path can remove efficiently, while controlled airflow should protect the components that still depend on air and provide predictable conditions around service interfaces. That means the concept design should define thermal zones, airflow direction, component-level temperature limits, leakage paths, filtration requirements, sensor locations, and maintenance access before the site commits to final rack and room geometry. It also means commissioning should validate actual airflow behavior at the PSU and network layer rather than confirming only the liquid loop and room supply conditions. Once residual heat becomes a governing load case, CRAH controls, containment geometry, rack architecture, and service procedures can respond to the real thermal map instead of an averaged rack number.