The Vera Rubin Era of AI is Shaping
AI progress is no longer defined by how fast models can be trained, but by how efficiently intelligence can be produced, reasoned, and deployed at massive scale. As AI factories are taking shape, reasoning and agentic workflows demand computing platforms that can sustain multi-step inference, with extremely long context windows, low latency, and energy efficient token generation across GPUs, racks and data centers.
The NVIDIA Vera Rubin platform is built specifically to meet this challenge. Architected as a rack-scale, fully-liquid cooled AI supercomputer, it brings together six chips — NVIDIA Vera CPU, NVIDIA Rubin GPU, NVIDIA NVLink™ 6 Switch, NVIDIA ConnectX®-9 SuperNIC™, NVIDIA BlueField®-4 DPU and NVIDIA Spectrum™-6 Ethernet Switch — through extreme co-design to operate as a unified system, enabling massive bandwidth for both scale-up and scale-out workloads. By optimizing how tokens, model state, and context flow across the AI factory efficiently, NVIDIA Vera Rubin NVL72 delivers astonishing improvements over the Blackwell generation in training efficiency and cost-per-token, setting the foundation for the next AI frontier.
Highlights
- 36 NVIDIA Vera CPUs
- 72 NVIDIA Rubin GPUs
- Support 6th generation NVIDIA NVLink™ for Scale-up
- Support NVIDIA Quantum-X800 InfiniBand and Spectrum-X™ Ethernet for Scale-out and Scale-across
- Cableless Compute Tray Design
- Built for the Age of Agentic Reasoning
Rack-scale Architecture
-
Design Details
Key Components of NVIDIA Vera Rubin NVL72
NVIDIA Vera Rubin Superchip
The NVIDIA Vera Rubin Superchip, combining one Vera CPU and two Rubin GPUs with HBM4 memory, forms the core engine for large-scale intelligence production. Delivering a significant boost in communication and memory movement, the superchip enables next-generation inference in the era of agentic AI.
NVIDIA Vera Rubin NVL72 Compute Tray
The Rubin NVL72 compute tray is a liquid-cooled, high-density module integrating two NVIDIA Vera CPUs, four NVIDIA Rubin GPUs, high-bandwidth memory, and seamless sixth-generation NVIDIA NVLink™ connectivity. Designed with a focus on rapid deployment, it features a cable-free modular design that greatly reduces assembly time while delivering outstanding performance to accelerate the AI industrial revolution.
