Description
Overview
The NVIDIA Vera Rubin NVL72 is the flagship rack-scale system of NVIDIA’s next-generation Rubin platform, unveiled at CES 2026. Built on the third-generation MGX modular reference design, it unifies 72 Rubin GPUs and 36 Vera CPUs in a single liquid-cooled rack connected by the NVLink 6 switch fabric.
Rubin NVL72 is purpose-built for agentic AI, advanced reasoning, and mixture-of-experts (MoE) workloads. Compared to Blackwell GB300 NVL72, NVIDIA cites a 4x reduction in GPUs required to train the same MoE model, with 5x higher inference throughput and 3.5x faster training.
Key Features
- Rubin GPU: Up to 288 GB HBM4 per GPU with 22 TB/s memory bandwidth.
- Vera CPU: Custom Arm “Olympus” core, 88 cores / 176 threads via NVIDIA Spatial Multi-Threading, 1.5 TB LPDDR5x.
- NVLink 6 Switch Fabric: 3.6 TB/s bidirectional GPU-to-GPU bandwidth, 260 TB/s scale-up rack bandwidth.
- ConnectX-9 SuperNIC and BlueField-4 DPU: Integrated for east-west and north-south traffic offload.
- Quantum-X800 InfiniBand / Spectrum-X Ethernet: Scale-out across thousands of nodes.
Technical Specifications
| Specification | Details |
|---|---|
| GPUs | 72 x NVIDIA Rubin |
| CPUs | 36 x NVIDIA Vera (Arm Olympus) |
| HBM4 Capacity | 20.7 TB total |
| HBM Bandwidth | 1.6 PB/s aggregate |
| LPDDR5x Capacity | 54 TB total |
| NVFP4 Inference | 3.6 EFLOPS |
| NVFP4 Training | 2.5 EFLOPS |
| Scale-Up Bandwidth | 260 TB/s (NVLink 6) |
| Networking | Quantum-X800 InfiniBand or Spectrum-X Ethernet |
| Form Factor | Liquid-cooled rack, MGX gen-3 |
Ideal Use Cases
- Training trillion-parameter foundation models and MoE architectures
- High-throughput agentic and reasoning model inference
- Scientific simulation and digital twin workloads at scale
- Sovereign AI factories and hyperscale cloud deployments
Why Choose Vera Rubin NVL72?
Rubin NVL72 sets the new ceiling for AI factory performance with co-engineered compute, memory, and networking that no single-vendor stack can match. Our team helps you design rack power and cooling, plan the migration path from Blackwell GB300 NVL72, and build an end-to-end AI factory blueprint with NVIDIA-certified partners.
Interested? Contact us for personalized pricing, lead times, and configuration options.
