(858) 613-1818Quote list 0
Coming soon

Vera Rubin NVL72 multi rack AI factory, coming soon

Multi rack Vera Rubin NVL72 AI factory with 144 or more Rubin GPUs and 20.7 TB of HBM4 per rack, designed now for delivery when Rubin ships.

GPU: NVIDIA Vera Rubin NVL72

Illustration of three racks of servers
Racks
2 or more x Vera Rubin NVL72 racks, 72 Rubin GPUs and 36 Vera CPUs each, one NVLink domain per rack
GPU memory
20.7 TB HBM4 per rack at 1,400 TB/s, per NVIDIA
Performance
3,600 PFLOPS sparse NVFP4 inference and 2,520 PFLOPS dense NVFP4 training per rack, per NVIDIA
Scale out fabric
ConnectX-9 SuperNICs and BlueField-4 DPUs into Quantum-X800 or Spectrum-X fabrics, 32.4 TB/s per rack
Storage and control
1 TB usable flash per GPU under a parallel file system, 6 x EPYC 9005 control plane nodes
Power and cooling
Liquid cooled; per GPU and rack power not yet published, planned against GB300 NVL72 as the baseline

Overview

The Vera Rubin NVL72 AI factory is the multi rack successor to the GB300 NVL72 factory. NVIDIA's published figures give each rack 72 Rubin GPUs with 20.7 TB of HBM4 at 1,400 TB/s, 36 Vera CPUs, sixth generation NVLink at 216 TB/s and 3,600 PFLOPS of sparse NVFP4 inference, with ConnectX-9 SuperNICs and BlueField-4 DPUs for scale out. Several racks join over that scale out fabric into one system.

NVIDIA states that Vera Rubin is ramping into full production with its system partners, but per GPU power, rack power and reseller delivery dates are not yet published, so we list the factory as coming soon and take expressions of interest rather than firm orders. A Rubin factory will need at least the liquid cooling and power density of a GB300 NVL72 factory, which is the planning baseline we use today.

If you are planning a Rubin deployment, we can reserve a queue position, run the site survey against the GB300 NVL72 requirements, design the fabric, storage and control plane now, and quote an interim GB300 NVL72 factory so the facility and software stack are ready when Rubin ships.

Use cases

  • Planning next generation AI factories
  • Roadmap alignment for sovereign compute
  • Reasoning model inference at extreme scale

Specifications

Compute
Racks2 or more x Vera Rubin NVL72, 72 Rubin GPUs and 36 Vera CPUs per rack, per NVIDIA
GPU memory20.7 TB HBM4 per rack at 1,400 TB/s
CPU memoryUp to 54 TB LPDDR5X per rack
NVFP4 performance3,600 PFLOPS sparse inference, 2,520 PFLOPS dense training per rack
Control plane6 x 1U single socket EPYC 9005 control plane servers, or Vera based equivalents when published
Interconnect
NVLinkSixth generation NVLink, 216 TB/s per rack, 65 TB/s NVLink-C2C
Scale outConnectX-9 SuperNICs and BlueField-4 DPUs, 32.4 TB/s bi-directional per rack, into Quantum-X800 InfiniBand or Spectrum-X Ethernet
Fabric designRail optimized leaf and spine sized to the rack count, designed with the order
Storage
Capacity1 TB usable per GPU minimum, 144 TB and up, expandable
Nodes2U 24 bay all NVMe storage servers, count sized to the rack count
SoftwareParallel file system with GPUDirect Storage, specified with the order
Facility and availability
CoolingLiquid cooled, 45 C inlet temperature per NVIDIA; CDUs and facility water loop required
PowerNot yet published; planned against GB300 NVL72 as the baseline
AvailabilityNVIDIA states Vera Rubin is ramping into full production; reseller delivery dates not yet published
Interim platformGB300 NVL72 racks on the same site design, deliverable now

Figures describe the system class. Exact configuration, weights and power draw are confirmed on the quotation.

What's included

  • Expression of interest and queue position
  • Site survey against GB300 NVL72 liquid cooling and power requirements
  • Fabric, storage and control plane design ahead of delivery
  • Specification updates as NVIDIA publishes them
  • Firm quotation when configurations are released

Compliance

Compare it with the next class up or down

The nearest classes we quote are the NVIDIA Vera Rubin NVL72 rack and the GB300 NVL72 multi rack AI factory. If neither is quite right, tell us the workload and we specify the system between them.

Talk to an engineerRequest a quotation