NVIDIA presents Vera Rubin as the generation that follows Blackwell. The Vera Rubin NVL72 rack is in production and integrators have started shipping it: Supermicro announced its first deliveries on 23 September 2026. For the datacenter that hosts it, this rack changes three parameters: power per rack, cooling water temperature and, for the next generation, supply voltage.
What a Vera Rubin NVL72 rack contains
A rack holds 72 Rubin GPUs and 36 Vera CPUs, connected by NVLink 6. NVIDIA announces 20.7 TB of HBM4 memory and 3.6 exaflops of FP4 inference compute (NVIDIA product page).
Integrators deliver these racks in units of 16. The Supermicro unit contains 1,152 GPUs and comes with in-row CDUs rated at 1.8 MW each, in N+1 redundancy (Supermicro announcement).
About 190 to 230 kW per rack
NVIDIA does not publish a single power figure for the rack. Published estimates range from 190 to 230 kW, compared with 132 to 142 kW for a GB300 NVL72 rack (ModulEdge). A 16-rack unit therefore represents 3 to 3.7 MW of IT power.
At this density, floor space matters little. Whether a site can host the hardware depends on the power distribution and cooling capacity available at each rack.
100% liquid cooling with 45 °C water
The Vera Rubin NVL72 rack is fully liquid-cooled, directly on the chips: its compute and switch trays have no fans. NVIDIA designs its MGX racks for a 45 °C water inlet (NVIDIA technical blog).
The datacenter has to bring water to every rack, with CDUs sized for the power of the row, and its cooling loop has to accept this inlet temperature.
The next generation: Vera Rubin Ultra in the Kyber rack
NVIDIA plans the Kyber rack for the second half of 2027. It will host Vera Rubin Ultra in an NVL144 configuration. When it was presented in March 2025, NVIDIA announced it at about 600 kW per rack, powered by 800 VDC (Tom's Hardware). The design has changed since then (GTC 2026 recap): the final power figure will be known when the product ships.
An 800 VDC supply requires a dedicated power distribution, to be planned when the room is designed.
Hosting Vera Rubin at our PAR9 site
Our PAR9 site in the Paris region is connected at 36 MVA. It accepts racks well above 150 kW, and in particular:
Vera Rubin NVL72 racks;
a 45 °C water inlet on the liquid cooling loop;
Kyber racks of about 600 kW, powered by 800 VDC.
The first tranche of a cluster, up to 1 MW, is delivered in 4 to 8 weeks. Beyond that, allow 3 months.
The configurator includes the Vera Rubin NVL72 and the Vera Rubin Ultra NVL144 in the Kyber 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.
