NVIDIA Vera Rubin POD
Type: Platform Architecture Tags: NVIDIA, Vera Rubin POD, Rubin, Vera CPU, Groq 3 LPX, BlueField-4 STX, Spectrum-6 SPX, MGX, DSX, AI factory Related: NVIDIA-Vera-Rubin, NVIDIA-Vera-CPU, NVIDIA-Groq-3-LPX, NVIDIA-Spectrum-6-SPX, NVIDIA-BlueField-4, NVIDIA-STX, NVIDIA-CMX, NVIDIA-ConnectX-9, NVIDIA-Spectrum-X, NVIDIA-Quantum-X800-InfiniBand, NVIDIA-Silicon-Photonics, NVIDIA-Enterprise-AI-Factory, NVIDIA-DGX-SuperPOD, NVIDIA-Dynamo Sources: https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/, https://nvidianews.nvidia.com/news/nvidia-vera-rubin-platform Last Updated: 2026-05-09
Summary
NVIDIA Vera Rubin POD is NVIDIA’s POD-scale AI factory architecture for the agentic AI era. It unifies five purpose-built rack-scale systems on the third-generation NVIDIA MGX rack architecture: Vera Rubin NVL72 compute racks, NVIDIA-Groq-3-LPX inference accelerator racks, NVIDIA-Vera-CPU racks, NVIDIA-BlueField-4 STX storage/context-memory racks, and NVIDIA-Spectrum-6-SPX networking racks.
Detail
Purpose
Vera Rubin POD moves AI factory infrastructure from separate server and rack-scale systems into a co-designed POD-scale supercomputer. It is meant for agentic workloads that need large context windows, low-latency inference, dense CPU sandboxing, shared context memory, resilient scale-out networking, and high token throughput per watt.
Architecture notes
- NVIDIA describes the POD as 40 racks with 1,152 Rubin GPUs, 60 exaflops, and 10 PB/s total scale-up bandwidth.
- The architecture uses extreme co-design across compute, networking, and storage chips.
- Vera Rubin NVL72 is the core rack-scale compute engine, integrating 72 Rubin GPUs and 36 Vera CPUs with NVLink.
- NVIDIA-Groq-3-LPX adds a low-latency inference accelerator path with 256 LPUs per rack.
- NVIDIA-Vera-CPU racks provide dense CPU capacity for reinforcement learning, sandboxing, and agent execution environments.
- NVIDIA-BlueField-4 STX racks host NVIDIA-CMX context memory storage for AI-native KV-cache sharing.
- NVIDIA-Spectrum-6-SPX racks connect the POD with Spectrum-X Ethernet or NVIDIA-Quantum-X800-InfiniBand switching.
- NVIDIA Vera Rubin DSX provides a broader AI factory reference-design platform that ties chips, systems, software libraries, APIs, partners, power, cooling, and facility controls into one deployment framework.
NVIDIA context
Use this page when the question is about all of the hardware that composes a next-generation NVIDIA AI factory rather than a single server, GPU, or networking component. It should bridge NVIDIA-Vera-Rubin, NVIDIA-Groq-3-LPX, NVIDIA-CMX, NVIDIA-Spectrum-6-SPX, NVIDIA-STX, NVIDIA-Dynamo, and NVIDIA-Enterprise-AI-Factory.
Connections
- NVIDIA-Vera-Rubin - platform family and compute architecture foundation.
- NVIDIA-Vera-CPU - CPU rack and superchip component.
- NVIDIA-Groq-3-LPX - low-latency inference accelerator rack in the POD.
- NVIDIA-Spectrum-6-SPX - networking rack connecting the POD.
- NVIDIA-BlueField-4 - storage/context processor used by STX and CMX.
- NVIDIA-STX - AI-native storage reference architecture used by the BlueField-4 STX rack.
- NVIDIA-CMX - context memory storage layer hosted in the STX rack.
- NVIDIA-ConnectX-9 - SuperNIC component in next-generation AI factory networking.
- NVIDIA-Spectrum-X - Ethernet fabric option for MGX ETL and SPX networking.
- NVIDIA-Quantum-X800-InfiniBand - InfiniBand fabric option for SPX networking.
- NVIDIA-Silicon-Photonics - CPO/optical direction used in Spectrum-6 SPX.
- NVIDIA-Enterprise-AI-Factory - strategic enterprise AI factory context.
- NVIDIA-DGX-SuperPOD - predecessor/current scale-out AI factory pattern.
- NVIDIA-Dynamo - serving framework that makes heterogeneous GPU/LPU/context-memory inference practical.
Source Excerpts
- NVIDIA’s technical blog describes Vera Rubin POD as five purpose-built rack-scale systems that work together as one AI supercomputer.
- NVIDIA lists key rack systems as NVL72, Groq 3 LPX, Vera CPU, BlueField-4 STX, and Spectrum-6 SPX.