NVIDIA Vera Rubin Platform: Seven Chips, Five Racks, One AI Supercomputer
The headline hardware announcement at GTC 2026: the Vera Rubin platform combines seven new chips and five rack-scale systems into one coherent AI supercomputer. Jensen claims 40 million times more compute in 10 years since DGX-1.
Key Takeaways
- 01
Seven new chips and five rack-scale systems designed to function as one massive AI supercomputer, supported by 80+ NVIDIA MGX ecosystem partners.
- 02
Vera Rubin NVL72 trains large MoE models with 1/4 the GPUs vs. Blackwell and delivers 10x inference throughput per watt at 1/10 cost per token.
- 03
Groq 3 LPX rack delivers 35x higher inference throughput per megawatt — Samsung manufactures the LP30 chip, shipping Q3 2026.
- 04
First rack already running at Microsoft Azure. AWS deploying 1M+ NVIDIA GPUs plus Groq LPUs.
The Lineup
- Vera Rubin NVL72 GPU rack — 72 Rubin GPUs + 36 Vera CPUs connected by NVLink 6, with ConnectX-9 SuperNICs and BlueField-4 DPUs. Trains large MoE models with 1/4 the GPUs vs. Blackwell; delivers 10x higher inference throughput per watt at 1/10 the cost per token.
- Vera CPU rack — 256 Vera CPUs in a liquid-cooled rack, purpose-built for reinforcement learning and agentic AI. 2x efficiency, 50% faster than traditional CPUs. Uses LPDDR5 — the only data center CPU to do so.
- Groq 3 LPX inference rack — 256 LPU processors with 128GB on-chip SRAM and 640 TB/s scale-up bandwidth. Delivers up to 35x higher inference throughput per megawatt. Samsung manufactures the LP30 chip; shipping Q3 2026.
- BlueField-4 STX storage rack — AI-native storage with 5x token throughput, 4x energy efficiency, and 2x faster data ingestion.
- Spectrum-6 SPX Ethernet rack — NVIDIA's first co-packaged optics switch in full production, co-developed with TSMC.
Key Numbers From the Keynote
- 3.6 exaflops of compute, 260 TB/s all-to-all NVLink bandwidth
- 100% liquid cooled with 45°C hot water
- Installation time: 2 days → 2 hours
- First rack already running at Microsoft Azure (confirmed by Satya Nadella)
- AWS deploying 1M+ NVIDIA GPUs plus Groq LPUs
What's Next
Rubin Ultra (taping out now): 144 GPUs in one NVLink domain via the new Kyber rack. Beyond that, the Feynman generation: new GPU, LP40 LPU, Rosa CPU, BlueField-5, CX10.
