Liquid cooling for GPU racks: above what kW per rack you need it, and what 50 to 120 kW racks demand

Where air cooling stops: 20 to 50 kW per rack

Liquid cooling is typically used above about 50 kW per rack, according to Uptime Institute, and the other published limits are close to its figures. Uptime puts perimeter air cooling at 20 to 25 kW per rack and close-coupled air, meaning in-row units with containment, at up to 50 kW. Vertiv and Schneider both draw the practical line for ordinary air around 20 kW. ASHRAE’s own paper gives the physical reason: a good raised floor delivers about 1,900 CFM per tile, while a 40 to 50 kW rack can need up to 5,000 CFM by its figure, or 6,400 to 8,000 at the 11 °C rise our rack guide assumes. We build AI servers to order with air or liquid cooling chosen by GPU count and site: tell us the load you plan per rack and our team will help.

A single GPU purchase can take a rack to that limit. Four eight-GPU air-cooled servers draw about 40 to 50 kW depending on the GPUs (four DGX H100 at up to 10.2 kW each are about 41 kW), with fans consuming up to a tenth of the load; the 64-GPU racks Uptime describes run 80 to 90 kW. As racks approach 150 kW, Uptime expects total or close to total liquid cooling. Its survey of July 2026 shows how recent this is: on average the typical rack has only just passed 11 kW, about one operator in four runs any rack at 30 kW or above, about one in seven runs racks above 50 kW, and some of those run racks above 100 kW.

Rear-door heat exchangers, cold plates and immersion compared

Sources: vendor datasheets (Vertiv, Motivair, Legrand ColdLogik, GRC), Schneider Electric WP133, Uptime Institute, DCD.

The rear door is the route that changes least. The servers stay air-cooled; the door takes the exhaust through a coil and returns cooled air to the room. Passive doors rely on the server fans, active ones add their own: Vertiv’s active DCD50 moves 9,000 m³/h. Capacity depends on water temperature, which is why one product quotes 93 kW at 14 °C and 74 kW at 21 °C. NVIDIA’s DGX B300 guidance does not recommend passive doors, and four systems on a 54 V busbar in one rack need an active door sized for 58 kW average and 76 kW peak at 6,200 CFM.

Cold plates take the heat where it is made. Vertiv puts the share captured by the plates at 70 to 75 % and Schneider at 60 to 90 %. Supermicro claims up to 92 % of the heat of its 4U liquid-cooled servers, and up to 98 % of the system heat for its DLC-2 generation. Whatever the figure, the remainder is still air: Uptime’s figures for a 132 kW-rated GB200 NVL72 are about 100 kW to the plates and more than 25 kW into the hall. A liquid-cooled rack therefore still needs some air cooling.

Immersion removes the fans and most of the airflow problem, but it needs servers built for the bath, a different service procedure and, historically, warranty terms agreed separately with the vendor. We found no accepted standard for the fire behaviour of single-phase fluids, and two-phase immersion became much less available after 3M stopped producing PFAS coolants, according to Uptime Institute.

Facility water: ASHRAE classes, flow rates and CDUs

ASHRAE classifies facility water by its maximum supply temperature: W17, W27, W32, W40, W45 and W+ above that, all with a 2 °C floor. Class W45 therefore means supply water from 2 to 45 °C, and compliance means full performance of the equipment across the class. The industry is converging on 32 °C supply (class W32), with 40 and even 45 °C as the targets. The hotter the water the equipment accepts, the less refrigeration the site needs: NVIDIA’s page for the Vera Rubin NVL72, read in October 2026, lists an inlet temperature of 45 °C and says the platform is ramping into full production.

Plan about 1.5 litres per minute of flow per kilowatt, which is what water at a 10 K rise carries; a 9 kW server therefore needs about 13 L/min. The coolant in the secondary loop is usually PG25, three parts water to one part propylene glycol, filtered to 50 µm, and always held above the room’s dew point so that nothing condenses on a cold plate.

Sources: Schneider Electric WP133, Vertiv CoolChip datasheets.

Liquid-to-air units need no plumbing to the building, because they return the heat to the hall. That works for a rack or two, but the heat then has to be removed by the room’s air conditioning again. Anything from a row upwards needs a facility loop, a heat-rejection path outside the building and, in a modern design, dry coolers rather than chillers.

A 120 kW rack: the GB200 NVL72

The GB200 NVL72 is the usual reference for this density. NVIDIA rates the rack at approximately 120 kW; Supermicro’s version lists 132 kW. The eighteen compute trays and the nine NVLink switch trays are liquid-cooled; only the power shelves and the management networking stay on air. Eight power shelves deliver 33 kW each in an N+N arrangement, and leak detection is built in. At GTC 2024, where it introduced the rack, NVIDIA described the loop as 25 °C in, 45 °C out, about 2 litres per second. Its successors, the GB300 NVL72 and the Vera Rubin NVL72, keep the liquid-cooled rack design. There is no air-cooled version; the same GPU silicon in air-cooled form is the DGX B200, at about 14.3 kW for ten rack units, and Supermicro fits 32 air-cooled B200 GPUs in a rack against 96 with cold plates.

Retrofitting liquid cooling into an existing data hall

In an existing hall, a liquid-cooled rack needs building work before the IT work can start. The first item is the floor. TIA-942 rates the floor of a Rated-3 data hall at 12 kPa, about 1,220 kg per square metre, and NVIDIA added more than 100 lb of steel to the NVL72 rack frame to withstand the mating force of the trays against the backplane. Next is the water loop, with supply and return piping to every rack position, and drip trays with leak detection and piped drains under connections that run above live equipment, which ASHRAE’s 2021 paper says should be installed. After that come heat rejection outside the building, commissioning of the CDUs with redundant pumps, and a service procedure for a rack that cannot be unplugged and rolled out like an air-cooled one.

Vertiv publishes a six-step roadmap for exactly this: workload, site audit, thermal modelling, site impacts, sustainability review, design; commissioning comes on top. We found no vendor timeline. If you need liquid cooling deployed quickly, the rear door is the route that changes least: it leaves the servers and the room as they are but still needs water piped to the rack, and its capacity ends somewhere between 50 and 93 kW, depending on how cold your water is. A 120 kW rack is beyond the doors listed here; racks of that density, such as the GB200 NVL72, ship with cold plates, so the work there starts with a CDU and the water that feeds it.

Which cooling method for which rack density

Up to about 20 to 25 kW per rack, room air is enough. Up to about 50 kW, use containment with close-coupled cooling, in-row units or a rear door. From 50 to about 80 kW, an active rear door on cold water is usually the least disruptive option; NVIDIA asks for one for four busbar-powered DGX B300 per rack. From about 80 kW upwards, or wherever the servers ship with cold plates, use direct-to-chip cooling with a facility loop. At 120 kW, racks such as the GB200 NVL72 are liquid-cooled by design.

Efficiency matters, but less than density in this decision. The average PUE (power usage effectiveness) in Uptime Institute’s 2026 survey was 1.52; liquid cooling mainly cuts fan and chiller energy, and Vertiv argues that PUE understates what direct-to-chip saves in fan power inside the server.

The RTX PRO 6000 Server Edition comes in an air-cooled dual-slot and a liquid-cooled single-slot version, so we build AI servers to order with the cooling chosen by GPU count and site, and check rack, power and airflow before we quote. Outline the rack and its cooling in the form below.

FAQ