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

Chapter 5.1 drew the cooling cliff and Chapter 5.2 pushed air to its honest limit — roughly 41 kW per rack with perfect containment, and a hard physical ceiling not far beyond. Chapter 5.4 will show that past ~100 kW direct-to-chip liquid is not a preference but a requirement. This chapter is about the country in between: the 40–75 kW band where air has run out but full DLC is either overkill or, more often, impossible in the building you already own because it has no facility water plumbed to the white space and no slab, plenum, or electrical headroom to land a CDU and a manifolded liquid rack. Rear-door heat exchangers (RDHx) and air-assisted liquid cooling (AALC) are the bridge across that gap. They are the most under-respected technologies in the cooling stack precisely because they are transitional, and being transitional is exactly what makes them the right answer for a specific, large, and time-pressured class of deployment.

The bridge is not free. Every kilowatt you carry on a door buys you a brownfield deployment in months instead of a greenfield in years, but it also commits you to a dew-point margin you must hold for the life of the hall, a density ceiling you will hit again in one GPU generation, and, if you choose active doors, a fan-power and acoustic budget that erodes the PUE advantage that made liquid attractive in the first place. The forks ahead are passive vs active, RDHx vs AALC, chilled-water vs warm-water coil, and the hybrid rack where DLC and a door coexist, and each one carries 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) circulating inside the coil absorbs the heat. Sized correctly, the air leaving the door is at or near room temperature: the rack is thermally room-neutral, contributing essentially zero net heat to the white space. That is the elegant trick — you have not changed the servers, not touched a cold plate, not added a quick-disconnect to a single GPU. You have wrapped a conventional air-cooled rack in a liquid jacket at the one place all of its heat must pass through anyway. The OCP Door Heat Exchanger project and ASHRAE TC 9.9 treat this as a distinct cooling class precisely because it sits cleanly between air and direct liquid: liquid economics, air-cooled servers.

The critical conceptual fork — and the one most often blurred in vendor literature — is RDHx vs AALC. A plain RDHx cools the room's air: the server is still an air-cooled server, and the door simply neutralizes its exhaust. Air-assisted liquid cooling inverts this. AALC runs a closed liquid loop inside the rack — cold plates on the hottest chips, a small in-rack manifold and pump, fed not by facility water but by a self-contained loop that rejects its heat through a rack-integrated or sidecar air-to-liquid heat exchanger back into the room's airflow. In other words, AALC delivers direct-to-chip cold-plate cooling to a chip in a facility that has no facility water whatsoever — the loop never leaves the rack, and its ultimate heat sink is room air (or, in the door-HX variant, a coil that hands the heat to chilled water at the door). RDHx cools the room; AALC cools the silicon and uses the room as a heat sink. They share hardware vocabulary and they are sold by the same vendors, but they are answers to different questions, and confusing them is how operators buy a door that cannot follow them to the next GPU generation.

The first hard design fork inside the RDHx family is whether the door is passive or active, and it turns entirely on where the air-moving energy comes from.

A passive RDHx has no fans of its own. It relies entirely on the static pressure the server fans already generate to push exhaust air through the coil. This is its great virtue: zero added electrical load, no fan failure mode, nothing to fail acoustically, and a PUE contribution that is essentially just the pump and the heat-rejection plant. The cost is capacity. Server fans are sized to move air across the servers, not to overcome the substantial pressure drop of a dense finned coil on top. Passive doors therefore top out around 10–30 kW per rack in practice — enough to neutralize a high-density air rack or take the edge off a CRAH-cooled hall, but not enough to reach the GPU-rack band. Push a passive door past its pressure budget and the server fans spin up to compensate, the IT-side fan power climbs, the servers run hotter, and you have quietly moved the energy you saved on a chiller into the servers' own fans — a PUE shell game.

An active RDHx adds its own fans (often EC fans) to the door, so the coil is no longer limited by server static pressure. This unlocks the band that matters for AI: ~40–75 kW per rack, with leading active doors (e.g. Motivair's ChilledDoor) rated to around 75 kW and 100% heat capture. The price is a real parasitic load — door fans are a continuous power draw, an N+1 redundancy question of their own, and a single point of failure that, if it trips, dumps the rack's full exhaust into the room with no warning. Active doors also reintroduce the acoustic problem that liquid cooling was supposed to solve: a wall of EC fans on dense racks is loud, and in some jurisdictions becomes an occupational-noise constraint. The fork is therefore: passive buys you efficiency and simplicity but caps you below the GPU band; active buys you the GPU band but reintroduces fan power, a redundancy obligation, and noise.

The four rows are roughly ordered by both capacity and irreversibility of the plumbing decision. Passive and active RDHx leave the servers air-cooled and demand only a chilled-water riser to the door — a comparatively cheap retrofit. AALC demands even less of the building (no facility water at all) but the most of the rack (an in-rack closed loop with cold plates). Full DLC demands the most of the building (a CDU on facility water, a manifolded warm-water loop) and is the only option that follows you up the density ramp without a re-fit. The bridge technologies exist to let you skip the building's irreversible plumbing decision for one more generation — and the consequence is that you make that decision later, under more time pressure, at a higher density, when it is harder.

Every liquid-to-air coil that runs below the room's dew point will condense water out of the air passing through it. This is not a defect — it is the same physics as the coil in a window air conditioner — but in a data center white space dripping liquid onto power-dense IT it is a problem you must engineer away, not hope away. The dew-point fork is therefore the second discipline the bridge imposes, and it sets a hard tradeoff between capacity and condensation risk.

The aggressive choice is to run the coil cold — chilled water at, say, 7–15 °C — which is below the typical white-space dew point. This maximizes the temperature difference between air and coil, which maximizes heat transfer, which maximizes the kilowatts the door can carry. The consequence is active condensate management: a drip tray under every door, condensate drains routed and pitched correctly, and humidity control in the room tight enough that you know your dew point and stay above the panic threshold. A blocked drain or a humidity excursion becomes a water-on-IT event.

The conservative choice — and the one the CDU-fed and warm-water world has converged on — is to hold the coil's supply temperature above the white-space dew point, so the door operates 100% sensibly: it removes heat but condenses nothing, because no surface in the airstream is ever cold enough to reach saturation. This is the same dew-point margin doctrine that governs the secondary loop in Chapter 5.6 — a CDU feeding an RDHx will deliberately keep secondary supply a few degrees above the measured dew point. The consequence runs the other way: a warmer coil has a smaller approach temperature to work with, so for the same capacity you need more coil area, more flow, or more fan power — which is precisely why warm-water RDHx tends to live at the lower, not the upper, end of the active-door capacity band. The fork is real: cold coil buys capacity and owes you condensate discipline; warm coil buys condensation-free operation and owes you a tighter thermal budget.

The single best reason to reach for AALC rather than a plain RDHx is the facility that has no facility water at all — and cannot get it on the timeline that matters. This is more common than the greenfield-centric literature admits: leased colocation halls where the landlord's lease and mechanical plant simply do not contemplate chilled water at the cabinet; brownfield enterprise rooms with a chiller plant sized and piped for a 5–10 kW-per-rack world; edge and Tier-2 metro sites where running a wet loop to the white space is a non-starter. In all of these, even a chilled-water RDHx is blocked, because the door's coil still needs to be fed by something.

AALC removes that dependency. The cold plates, the small manifold, and the pump live inside the rack; the loop is sealed and factory-charged; and its heat is rejected through a rack-integrated or sidecar air-to-liquid heat exchanger straight back into the room's existing airflow — which the room's existing CRAH plant then handles as it always has. From the building's perspective, an AALC rack is just another air-cooled rack with a (high) air heat load; from the silicon's perspective, it is getting genuine cold-plate cooling. That is the bridge in its purest form: you deliver direct-liquid thermal performance to a chip that needs it, in a building that cannot supply a drop of facility water, and you defer the wet-plumbing decision entirely.

The consequence — and the reason AALC is a bridge and not a destination — is the ceiling. Because the loop's ultimate heat sink is room air, AALC inherits air's room-side limits: the room must still be able to absorb and reject the rack's full heat through its air-side plant, so AALC deployments cluster around ~40 kW per rack and the in-rack HX and pump are sized for a TDP that next-generation GPUs will exceed. When the chips' heat load outgrows what a rack-integrated air-to-liquid HX can shed into room air, AALC has no more headroom and you are back at the facility-water decision — only now the silicon is already on cold plates, which at least makes the cut-over to a CDU-fed DLC loop (Chapter 5.4) less violent than starting from air.

The most important 2026 use of the rear door is as the air-side partner in a direct-liquid rack, not as a standalone bridge. Even a fully DLC-cooled GPU rack does not put 100% of its heat on the cold plates. The GPUs and often the highest-power switch ASICs are liquid-cooled, but a meaningful residual — commonly 15–30% of rack power — still leaves on air: NICs, DIMMs, voltage regulators, power-supply losses, optics, and any component without a cold plate (the full inventory of 'what stays on air' is treated in Chapter 5.4). On a 132 kW NVL72-class rack, 15–30% residual is ~20–40 kW of air load — itself above the comfort zone of an ordinary contained hall, concentrated on a single dense cabinet.

This is where the door earns its place in the liquid era. An RDHx on a DLC rack neutralizes the residual air load at the rack, making the entire rack — liquid plus air — thermally room-neutral. The payoff is in plant sizing and stranded capacity. If the residual air goes to the room, the hall's CRAH/containment plant must be sized for the sum of every rack's residual, and in a dense hall that air-side plant becomes a real, separately-redundant cost that competes for the same floor and power. If instead each rack's residual is captured at its own door — fed off the same CDU and facility-water loop already serving the cold plates — you can shrink or, in the limit, eliminate the room's mechanical air-cooling plant, converging on a near-room-neutral white space whose entire heat load goes to liquid. That is the architecture several hyperscale liquid halls are moving toward: DLC for the silicon, RDHx for the residual, one facility-water loop doing both, and a CRAH plant reduced to ventilation and ride-through rather than primary cooling.

Putting the forks together yields a clean decision order. First, is the density target inside the bridge band (roughly 41–75 kW) — if it is below, Chapter 5.2's air-and-containment toolkit is cheaper and you do not need a door; if it is well above, Chapter 5.4's DLC is mandatory and a door is at most the air-side partner. Second, does the building have facility water at the white space — if yes, an active RDHx is the simplest bridge; if no, AALC is the only bridge that reaches cold-plate performance. Third, what is the dew-point posture — a cold coil for maximum capacity with condensate discipline, or a warm coil held above dew point for sensible-only operation at a tighter thermal budget. Fourth, and underneath all of it: is this hall a bridge (a brownfield substrate you already own, where the door buys you a generation) or a destination (a greenfield where bridging is just a costlier route to the DLC you will build anyway).

The recurring consequence to keep in view is the density ramp. The bridge band is a moving target shrinking from below: as GPU TDPs climb, the residual-air share of a DLC rack stays roughly constant in percentage but grows in absolute kilowatts, and the standalone-RDHx ceiling that comfortably hosted a Hopper-class air rack will not host the air-cooled servers of two generations hence. The bridge is real, useful, and current — but it is, by construction, temporary for any deployment on the training or dense-inference trajectory. Buy it for what it is: the fastest way to put high-density racks into a building that was not built for them, with your eyes open about when you will hit the far bank.