An operator with a full air-cooled hall and a rack of accelerators to install has two ways forward. Run pipework to the rack, which is a construction project. Or use a unit that captures heat in liquid and hands it back to the room as air, which is a delivery. The second is faster and more limited, and knowing where the limit falls is the whole decision.
What liquid-to-air actually does
Heat is captured in liquid at the chip, carried a short distance, and given up to air through a heat exchanger inside a self-contained unit, which vents into the hot aisle. The loop never leaves the rack or its neighbor, so the building needs no new pipes, no additional plant, and no shutdown to install one. Google's Brazos design is a recent example: a sidecar rack that cools the rack beside it, published with a reported capacity of around 60 kW and intended to be installed one rack at a time.
The heat does not leave the room
This is the trade, and it is the whole trade. A liquid-to-air unit changes where heat is captured, not where the building finally rejects it. The existing air handling still has to absorb every watt, and the hot aisle runs hotter than it did. A hall with spare cooling capacity and good containment absorbs that comfortably. A hall already at its limit gains far less than the rack figure suggests, and may gain nothing at all.
- Spare air handling capacity, measured at the worst hour
- Hot aisle containment good enough to keep the hotter air out of the intakes
- Electrical capacity at the rack, which is often the binding limit
- Floor space for a sidecar, which occupies a rack position
Facility water costs more and scales further
Connecting to facility water means pipework, a coolant distribution unit, and the plant to reject the heat outside. It is slower, more expensive, and more disruptive to install. In exchange, the heat genuinely leaves the room, the density ceiling is set by the plant rather than by the air handling, and the loop can run warm enough to avoid a chiller for much of the year. Every design that reaches beyond a handful of dense racks ends up here.
How to choose between them
Count the racks and count the years. A few dense racks in a hall with cooling headroom, needed this quarter, is exactly what liquid-to-air is for. A row of them, or a hall being repositioned for AI workloads over several years, will hit the air handling ceiling and should be plumbed properly the first time. The expensive mistake is using sidecars as a strategy rather than as a bridge, and discovering the ceiling only after the fourth deployment.
- Under roughly five dense racks, with cooling headroom: liquid-to-air
- A row or more, or a multi-year plan: facility water
- No electrical headroom: neither, until that is solved
- Leased space with no rights to alter the building: liquid-to-air by default
Where rear-door heat exchangers sit
A rear-door heat exchanger is a third option and is easy to confuse with the first, because both hand heat to air or water at the rack. The difference is what captures the heat. A rear door still relies on server fans moving air across components, then cools that air on the way out, whereas liquid-to-air captures heat directly at the chip with cold plates. That makes the rear door less invasive, since nothing inside the server changes, and less capable at the densities that make people consider liquid in the first place.