Rear-Door Heat Exchangers & Air-Assisted Liquid Cooling (The Bridge)

Chapter 5.1 set out the overlapping ranges where each cooling method works, and Chapter 5.2 showed where air runs out: airflow, inlet limits, containment, fan power, heat flux and serviceability. Chapter 5.4 explains why direct-to-chip liquid is the supported design for frontier racks that put most of their heat into liquid. This chapter covers the ground in between. It is the brownfield hall where the existing air system cannot cool the rack you want to install, but full DLC is overkill or, more often, impossible in the building you already own: no facility water to the white space, and no slab, plenum or electrical headroom for a CDU and a manifolded liquid rack. Rear-door heat exchangers (RDHx) and air-assisted liquid cooling (AALC) bridge that gap. They are easy to underrate because they are transitional, and transitional is exactly what a large, time-pressured class of deployments needs.

The bridge is not free. Each kilowatt you put on a door gets a brownfield deployment done in months rather than a greenfield one in years. In exchange you commit to a dew-point margin for the life of the hall, a density ceiling you will hit again within one GPU generation and, with active doors, fan power and noise that eat into the PUE advantage that made liquid attractive. So know which generation will hit that ceiling before you buy the door. GB300 NVL72 shipped in 2025 and is deploying through 2026 alongside GB200; Lenovo's profile puts it at 135 kW TDP and 155 kW peak. VR200 NVL72 has been ramping into full production since May 2026, and NVIDIA reported racks running at partners in late August; it runs at 188 kW in Max Q and 228 kW in Max P, with a 330 kW facility design basis. Rubin Ultra/Kyber is a ~600 kW planning point for NVIDIA's H2 2027 target; SemiAnalysis reported a slip to 2028 in July 2026, which NVIDIA disputes. The choices below are passive or active, RDHx or AALC, chilled or warm water, and the hybrid rack that uses DLC and a door together. Each has a downstream cost.

A rear-door heat exchanger is a finned liquid-to-air coil that replaces the perforated back door of a standard cabinet. Server fans push hot exhaust air, typically 35–45 °C off a dense GPU rack, through the coil before it re-enters the room. Chilled or tempered water (or a glycol mix) in the coil absorbs the heat. Sized correctly, the door returns air at or near room temperature, so the rack is thermally room-neutral and adds essentially zero net heat to the white space. You have not changed the servers, touched a cold plate or added a quick-disconnect to a single GPU. You have wrapped an ordinary air-cooled rack in a liquid jacket at the one place all its heat must pass through anyway. The OCP Door Heat Exchanger project and ASHRAE TC 9.9 treat this as its own cooling class because it sits cleanly between air and direct liquid: liquid economics with air-cooled servers.

The distinction vendor literature most often blurs is RDHx vs AALC. A plain RDHx cools the room's air. The server is still an air-cooled server, and the door neutralizes its exhaust. Air-assisted liquid cooling works the other way round. It runs a closed liquid loop inside the rack: cold plates on the hottest chips, a small in-rack manifold and pump, and a rack-integrated or sidecar heat exchanger that rejects the heat, with no facility-water drop at the chip. Where that exchanger sends the heat varies by vendor, and you need to check it before you classify anything. Some units reject into the room's air and need no building water at all. Motivair's ChilledDoor AALC arrangement instead returns the captured heat to the facility chilled-water loop through the door. The room-air version gives you direct-to-chip cold-plate cooling in a building with no facility water whatsoever: the loop never leaves the rack and the room is the final heat sink. In the door version, the coil hands the heat to facility water, and the room carries only what the door misses, plus the heat released if the door fails. So RDHx intercepts exhaust, while AALC cools the silicon and then rejects the heat to room air or facility water, depending on the design. The two share hardware vocabulary and vendors, but they answer different questions, and confusing them is how operators end up with a door that cannot follow them to the next GPU generation.

The first real design choice inside the RDHx family is whether the door is passive or active, and it depends on where the energy to move the air comes from.

A passive RDHx has no fans. It relies on the static pressure the server fans already produce to push exhaust through the coil. That is its great virtue: no door-fan power, no door-fan failures, and a PUE contribution that is little more than the pump and the heat-rejection plant. The server fans still respond to the extra pressure drop, though, and that is where capacity runs out. Server fans are sized to move air across the servers, not to force it through a dense finned coil as well. Push a passive door past its pressure budget and the server fans spin up to compensate, IT fan power climbs, the servers run hotter, and the energy you saved at the chiller has moved into the servers' own fans: a PUE shell game.

Published passive-door ratings are hard to pin down. Vertiv's Liebert DCD sheet rates the DCD35 and DCD50 at 35 and 50 kW with 21 °C room air, 12 °C entering water and 50% RH. Those are not clean passive ratings: the same rating table covers models sold only with the active fan module, and Vertiv's current CoolLoop sheet lists 550–700 W of door-fan power at nominal capacity, plus higher maxima (66 and 85 kW) at 100 cfm/kW of server airflow. No published source gives a general kW range for passive doors, so ask the vendor for the passive schedule itself: airflow and static pressure, water flow and pressure drop, fan state and capture target.

An active RDHx adds its own fans (often EC fans) to the door, so the coil is no longer limited by the server fans' static pressure. That can raise the rated duty, but only for that model at its published entering-air and entering-water temperatures, airflow, water flow and pressure, fan state, containment and capture target. A 75 or 100 kW rating does not transfer to other conditions. The price is a real parasitic load. Door fans draw power continuously, need their own N+1 redundancy and are a single point of failure. If the fans trip, the heat the door stops capturing goes into the room; if the water is lost, that can be the door's full rated duty. Unless a local response cuts the load, the room has to absorb it, and the controls must warn you that the cooling path is gone. Active doors also bring back the noise liquid cooling was supposed to remove. A wall of EC fans on dense racks is loud, and in some jurisdictions it becomes an occupational-noise constraint. The right choice depends on the product and the site: passive saves door-fan energy but spends server-fan pressure; active supplies pressure but adds fan energy, noise, controls, maintenance and a failure response you have to plan.

The four rows run roughly in order of capacity and of how irreversible the plumbing is. Passive and active RDHx leave the servers air-cooled and need only a chilled-water riser to the door, a comparatively cheap retrofit. The L2A form of AALC asks less of the building (no facility water at all, though room cooling must reject its full heat) and the most of the rack (an in-rack closed loop with cold plates). Full DLC asks the most of the building (a CDU on facility water and a manifolded warm-water loop). In return it can follow you up the density ramp without refitting the building, provided its water temperature and flow, pressure and distribution can handle each refresh you fund. The bridge technologies let you postpone the building's irreversible plumbing decision by one generation. The consequence is that you make it later, under more time pressure, at higher density, when it is harder.

Any liquid-to-air coil that runs below the room's dew point condenses water out of the air passing through it. That is not a defect; a window air conditioner does the same. But in a white space, water dripping onto power-dense IT is a problem you must engineer away. The dew-point margin is the second discipline the bridge imposes, and it forces a hard tradeoff between capacity and condensation risk.

The aggressive choice is to run the coil cold, with chilled water at, say, 7–12 °C, at or below a typical 10–13 °C white-space dew point, and accept condensation wherever the coil surface falls below the measured local dew point. A colder coil means a bigger temperature difference between air and coil, more heat transfer and more kilowatts per door. The consequence is active condensate management: a drip tray under every door, condensate drains routed and pitched correctly, and room humidity controlled tightly enough that you know your dew point and keep a safe margin from it. A blocked drain or a humidity excursion becomes a water-on-IT event.

The conservative choice, and the one CDU-fed and warm-water designs have converged on, is to hold the coil's supply temperature above the white-space dew point. The door then runs 100% sensible: it removes heat but condenses nothing, because no surface in the airstream is ever cold enough to reach saturation. The same dew-point margin governs the secondary loop in Chapter 5.6, where a CDU feeding an RDHx deliberately keeps its supply a few degrees above the measured dew point. Here the tradeoff runs the other way. A warmer coil has a smaller approach temperature to work with, so the same capacity needs more coil area, more flow or more fan power. At a given room temperature, that is why warm-water RDHx tends to sit at the lower end of the active-door capacity range. A cold coil buys capacity and demands condensate discipline; a warm coil buys condensation-free operation and leaves less thermal headroom.

AALC is worth considering where cold-plate equipment has to run without facility water at the rack, and a rack-integrated or sidecar unit rated for the job can reject the rack's full heat into the room. Some leased, brownfield, edge and metro sites fit that description, but the lack of rack water alone does not make AALC feasible.

The best-documented AALC deployments are hyperscale. Meta's GB200-based Catalina pod runs in air-cooled halls built for about 20 kW per rack: its two compute racks (reported at about 120 kW each) are flanked by four air-assisted liquid-cooling racks, so cooling takes four of the pod's six rack positions. Microsoft's Maia Sidekick does the same job with a fan-cooled sidecar. AALC trades floor space for water, so budget the cooling positions per compute rack and prove the room air plant can reject the whole pod's heat.

The cold plates, manifold, pump and local heat exchanger form a factory-built loop, but the building still receives all the heat through room air. Get your rack's heat split and heat flux, the coolant, the pump and heat-exchanger ratings, the entering-air conditions, fan state, service clearances and failure response. Then confirm that the existing room air and heat-rejection plant can carry the full rack load at design climate, with the redundancy you have committed to.

AALC's limit is the whole system (the local unit plus the room), not a portable ~40 kW ceiling. A future rack at the same kW can still fail on cold-plate heat flux, pump or heat-exchanger capacity, fan and air conditions, room rejection, serviceability or redundancy. If your refresh roadmap may outgrow what the unit and the room can handle, keep a practical path to convert to a TCS/FWS liquid loop.

The most important use of the rear door in 2026 is as the air-side partner of a direct-liquid rack, not as a standalone bridge. A GB200/GB300-class DLC rack does not put all of its heat on the cold plates. The GPUs, and often the highest-power switch ASICs, are liquid-cooled, but a meaningful share, commonly ~10–20% of rack power, still leaves as hot air (Chapter 5.4 lists what stays on air). On the 132 kW GB200 NVL72 reference, that residual is ~17 kW (~13%). The share varies by product: Lenovo documents a 90/10 liquid/air split for its GB300 NVL72, so read the residual off the rack you are actually buying. Either way it is modest, but it is concentrated in one dense cabinet and repeated down every row.

This is where the door earns its place in the liquid era. An RDHx on a DLC rack neutralizes the residual air at the rack, so the entire rack, liquid and air, becomes room-neutral. The payoff is in plant sizing and stranded capacity. If the residual goes to the room, the hall's CRAH and containment plant must be sized for the sum of every rack's residual. In a dense hall that becomes a real, separately redundant cost competing for the same floor space and power. If instead each rack's residual is captured at its own door, fed from the same plant that serves the cold plates, you can shrink the room's mechanical air-cooling plant. You can eliminate it only where another heat sink, or a proven load reduction, covers every failure and maintenance state you must survive. The result is a white space whose heat goes almost entirely to liquid. The catch is water temperature. A door on 40 °C warm water cannot return 24 °C air to the room, so either the door gets a colder branch or the room is designed for warmer air (Chapter 5.7). Vendors now sell this pairing: HPE, for example, offers Motivair rear doors alongside its direct-liquid-cooled racks to make the rack room-neutral. The design is DLC for the silicon, RDHx for the residual, one facility-water loop for both, and a CRAH plant reduced to ventilation and ride-through rather than primary cooling.

Decide in order of evidence, not by density band. First, pin down your rack's heat split, heat flux, airflow and inlet limits, and the thermal interfaces it supports. Second, test what the existing hall can deliver: airflow, water availability and temperature, capacity to reject the heat left in the room, design climate, structure, service access and redundancy. Third, check each candidate against its model-specific rating: entering-air and water temperatures, air and water flow, pressure, fan state, containment and capture target. Published door ratings are taken with room (server supply) air from 21 to 35 °C and water 3 to 10 K colder than that air, so ask each vendor to restate capacity at your server inlet temperature, entering-water temperature and server airflow (cfm/kW) before comparing headline kW. Fourth, compare today's answer with your refresh roadmap and migration plan.

RDHx stays on the list only where the server or door airflow and the water supply match the door's rating. AALC stays only where the local unit and the room can reject the full heat. DLC stays only where the TCS/CDU/FWS chain and the residual-air path both carry their load. A bridge is worth it when it passes those checks and beats the alternative on lifecycle cost, not because the rack falls inside some portable kW band.

Water flow = 20.0 × 60/(4.18 × 10.0) = 28.7 L/min. The cold-end difference is 24.0 − 20.0 = 4.0 K and the hot-end difference 36.0 − 30.0 = 6.0 K, so neither end crosses and the temperatures are feasible. Get the supplier's airflow, static-pressure and coil selection before buying the door. A 35.0 °C TCS supply cannot produce 24.0 °C air: even an ideal coil needs water no warmer than 24.0 °C, and a real one needs it colder. Either keep the colder water branch or carry the 20.0 kW on room cooling.

One lost door releases 20.0 kW, inside the assumed 25.0 kW reserve. Four lost doors release 80.0 kW and exceed it by 55.0 kW. The common failure needs another sink or a proven heat reduction of at least 55.0 kW. Use the heat balance in 5.1 and the ASHRAE heat-path guidance; the installed failure test belongs to 13.5.

Choose the door only when its water and air ratings deliver the outlet temperature you need and the room plant can carry the worst failure you must survive. A warm-water mismatch costs you a second temperature branch. Removing room capacity without a proven way to reduce load turns a door outage into stranded compute.