The Rack Door Buying AI More Time

One of AI's more consequential infrastructure upgrades looks almost comically ordinary: a server-rack door.

The usual story about AI data centers jumps straight from today's air-cooled rooms to new campuses full of direct-to-chip liquid cooling. That is directionally right at the frontier. It is also an incomplete picture of how infrastructure actually changes. Buildings, power systems, cooling plants, and colocation leases turn over much more slowly than GPUs do. The installed base does not disappear because the newest rack became harder to cool.

That mismatch creates room for a bridge. A rear-door heat exchanger, or RDHx, replaces the back of a server rack with a water-fed coil. Hot exhaust air passes through the coil, which moves the heat into the facility's chilled-water system. The server can remain air-cooled; the building captures the heat much closer to where it is created.

The idea is simple. The consequence is not. If RDHx works across a meaningful band of AI inference workloads, some supposedly obsolete data-center capacity may remain useful for another hardware cycle. The scarce asset is no longer just the door. It is the building's ability to supply water, power, controls, and operating discipline at the rack.

A surprisingly literal back door

The transition to liquid cooling is real, but it is not uniform. In the Uptime Institute's 2025 cooling survey, 75% of respondents said they used perimeter air cooling. When operators were asked what makes direct liquid cooling viable, the most common answer was ease of retrofit into existing infrastructure, selected by 46%.

That is the practical constraint hiding beneath the AI-infrastructure boom. Operators are not choosing cooling methods on a blank sheet of paper. They are trying to fit new compute into buildings that already have pipes, chillers, maintenance procedures, customers, and uptime commitments.

For air-cooled servers in a facility with an outdoor chilled-water system, Schneider Electric describes RDHx as a bolt-on option. Hot server exhaust crosses coils in the rear door, transfers heat into a closed liquid loop, and returns to the building's heat-rejection system. Schneider says configurations such as its Motivair ChilledDoor can cool up to 72 kilowatts per rack, depending on the size of the exchanger and the surrounding system.

The qualifier matters more than the headline number. A product rating is not a promise that any old room can suddenly absorb a 72-kilowatt rack.

Vertiv's reference design for a 50-kilowatt RDHx rack makes the dependencies concrete. Its stated capacity assumes 12°C chilled water flowing at 90 liters per minute, a 23°C room, and 55% relative humidity. The design includes power distribution and monitoring, and it calls for installation, fluid fill, startup, acceptance testing, and ongoing service.

In other words, the rack door is the visible object. The product is the working system around it.

The retrofit is a systems business

That distinction changes where value may accrue.

The obvious market is the hardware: heat exchangers, coolant distribution units, pumps, valves, hoses, sensors, and controls. But a facility owner cannot treat these pieces as interchangeable accessories. Flow rates, pump curves, water temperatures, filtration, controls logic, serviceability, and redundancy have to work together. Schneider's own liquid-cooling reference-design guidance calls liquid cooling "a system, not a SKU."

The more interesting venture opportunity may therefore sit one layer above the component. A retrofit platform that can assess a building, model the thermal envelope, configure equipment, commission the loop, and monitor performance over time is selling something more valuable than cooling capacity: it is selling certainty that a brownfield deployment will work.

That also shifts power toward facility owners and colocation operators. A tenant may choose the servers, but the building controls the chilled-water connection, operating envelope, maintenance access, and approval process. A facility that can qualify and deliver a repeatable high-density zone may monetize existing capacity differently from one that offers only generic power and floor space.

Our read is that RDHx is less a new cooling category than a way to segment the installed base. Some facilities will have the water system and operating headroom to become credible AI-inference sites. Others will discover that the retrofit exposes a deeper constraint in power delivery, heat rejection, or controls. The door does not erase those differences; it reveals them.

A bridge, not an endpoint

There is a clean countercase. RDHx still cools an air-cooled server. It does not remove heat directly from the chip, and it cannot make every workload fit every building.

Equinix notes that rear-door and in-row approaches can improve air-cooling effectiveness with relatively limited disruption, but also that their effectiveness declines as facility water temperatures rise. At the leading edge, the thermal burden moves beyond what augmented air should be asked to do.

The frontier makes that limit visible. Schneider's 2026 reference designs for NVIDIA Grace Blackwell GB300 NVL72 systems are engineered for up to 142 kilowatts per rack. Those densities demand a holistic liquid-cooling architecture rather than a clever attachment to yesterday's room.

So the thesis is not that the future stays air-cooled. It is that the migration will be tiered. Purpose-built training clusters will move fastest toward direct liquid cooling. A broader set of inference, enterprise, and mixed-density workloads may occupy the middle, where RDHx can extend the useful life of compatible facilities.

The bridge ends if two things happen together: mainstream AI servers stop offering air-cooled configurations, and the workloads buyers want routinely exceed the practical capacity of brownfield water and power systems. It lasts longer if model efficiency improves, inference stacks are right-sized to the application, and operators can standardize retrofits across many sites.

That is what we would watch: not the largest rack announced at a conference, but the share of deployed AI workloads that actually need it; not the nameplate capacity of a door, but its sustained performance inside a live facility; and not the number of cooling products launched, but whether operators can turn retrofit projects into a repeatable service.

The future of AI cooling is still liquid. It may simply arrive through the back door.