Vera Rubin NVL72 / Vera Rubin NVL72 specs
NVIDIA Vera Rubin NVL72 specs
What's in one rack, what each number means for your workload, and where each figure comes from.
The spec table
| Item | Per rack | Source |
|---|---|---|
| GPUs | 72 NVIDIA Rubin | nvidia.com |
| CPUs | 36 NVIDIA Vera (88 custom Olympus cores each) | nvidianews.nvidia.com |
| GPU memory (HBM4) | 20.7 TB at 1,580 TB/s | nvidia.com |
| CPU memory (LPDDR5X) | 54 TB | nvidia.com |
| NVLink 6 bandwidth in the rack | 260 TB/s (3.6 TB/s per GPU) | nvidianews.nvidia.com |
| NVFP4 inference | 3,600 PFLOPS | nvidia.com |
| NVFP4 training | 2,520 PFLOPS | nvidia.com |
| FP8 / FP6 training | 1,260 PFLOPS | nvidia.com |
| Scale-out networking | 28.8 TB/s, ConnectX-9 SuperNICs | nvidia.com |
| DPU | BlueField-4 | nvidianews.nvidia.com |
| Rack power (reported) | About 190 to 230 kW | hashrateindex.com |
| Cooling (reported) | 100% liquid, 45°C inlet water | hashrateindex.com |
NVIDIA labels its figures preliminary and "up to" (NVIDIA). Power and cooling numbers are not on NVIDIA's product page; the ones here are reported by Hashrate Index and should be confirmed with your operator.
Six new chips, one system
NVIDIA designed the platform as a set: the Rubin GPU, the Vera CPU, the NVLink 6 switch, the ConnectX-9 SuperNIC, the BlueField-4 DPU and the Spectrum-6 Ethernet switch (NVIDIA, Jan 2026). A Vera Rubin NVL72 rack combines 72 Rubin GPUs and 36 Vera CPUs with NVLink 6, ConnectX-9 and BlueField-4.
The Rubin GPU
Each Rubin GPU delivers 50 petaflops of NVFP4 inference compute (NVIDIA). Across the rack that is 3,600 PFLOPS for inference and 2,520 PFLOPS for training. Memory moves to HBM4: 20.7 TB per rack at 1,580 TB/s, nearly three times the memory bandwidth of the Blackwell generation.
The Vera CPU
Vera has 88 custom Olympus cores (NVIDIA), and the rack carries 54 TB of LPDDR5X. Agentic workloads lean on the CPU for orchestration and tool calls, so this matters more than CPU specs did a generation ago.
NVLink 6
3.6 TB/s per GPU and 260 TB/s per rack, double GB300. This is what lets 72 GPUs act as one for models that don't fit in a single server.
Built to be serviced
The trays are modular and cable-free, and NVIDIA says assembly and servicing are up to 18x faster than Blackwell (NVIDIA). For a renter that shows up as faster repairs and less downtime, so ask what the operator commits to.
Questions
How many GPUs are in a Vera Rubin NVL72?
How much memory does a Vera Rubin NVL72 have?
What does NVL72 mean?
RackShore runs this site and may be paid a referral fee by the provider you choose. Buyers pay us nothing. Spot a mistake? Tell us.