The GB300 NVL72 is NVIDIA's rack built on Blackwell Ultra. It follows the GB200 NVL72 and precedes Vera Rubin, which NVIDIA has been shipping since the second half of 2026. Before choosing a site to host it, three figures matter: electrical power, the share of heat removed by liquid, and the water temperature.
What a GB300 NVL72 rack contains
A rack holds 72 Blackwell Ultra GPUs and 36 Grace CPUs, spread over 18 compute trays and connected by 9 NVLink switch trays. NVIDIA announces 20 TB of GPU memory, 130 TB/s of NVLink bandwidth and 1,440 petaflops of FP4 compute (NVIDIA product page).
The rack takes 48U, 600 mm wide and 2,294 mm high (Lenovo product guide).
135 kW nominal, up to 155 kW at peak
Lenovo gives a power of 135 kW per rack, with peaks up to 155 kW depending on the workload. Estimates published elsewhere range from 132 to 142 kW (ModulEdge). The rack is powered by 6 or 8 power shelves of 33 kW, on three-phase 346 to 480 V.
Ten racks therefore represent about 1.35 MW of IT power, excluding cooling. How many racks a room can host depends on the site's grid connection, the power distribution to the row and the margin kept for peaks.
90% of the heat into water, 10% into air
GPUs and CPUs are cooled by cold plates. According to Lenovo, about 90% of the heat goes into the liquid loop and 10% into the air, which leaves about 13 to 16 kW per rack for the room to remove, through its air conditioning or a cooled rear door.
Water can enter the compute trays at up to 45 °C. The warmer the water, the higher the flow: Lenovo gives 59 L/min with 25 °C water and 177 L/min at 45 °C.
The next generation removes this air share: Vera Rubin NVL72 trays have no fans, and the rack reaches about 190 to 230 kW (our article on Vera Rubin).
Where to host GB300 NVL72 racks at DC2SCALE
150 kW per rack with direct liquid cooling is a common density on our sites: a GB300 NVL72 rack fits within it in nominal operation. Higher densities are possible after a study, and the technical study validates the configuration with you, peaks included.
For a cluster of several megawatts, available power decides the site:
PAR9, in the Paris region, connected at 36 MVA, which accepts racks well above 150 kW;
Castres, connected at 10 MVA;
Lille, connected at 5 MVA.
The first tranche of a cluster, up to 1 MW, is delivered in 4 to 8 weeks. Beyond that, allow 3 months.
The configurator counts 140 kW per GB300 NVL72 rack: from the number of GPUs, it works out the power, the rack count and a budget range. For a project of 1 MW or more, give us the target power and timeline in the large-account form: your sales contact calls you back within 48 hours.
