Vera Rubin
#2 in KI-Trainings-Hardwarenvidia · siet Zweite Jahreshälfte 2026 (Partnersysteme); volle Serienproduktion seit Anfang 2026 · 13× · tolest 30. Juni 2026
NVIDIA Vera Rubin is NVIDIA's next-generation rack-scale AI infrastructure platform and successor to Blackwell. It comprises seven co-designed chips (including the Rubin GPU, Vera CPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch, and Groq 3 LPU) combined into five rack-scale systems forming one AI supercomputer. The platform has been in full production since early 2026, with first partner systems (including AWS, Google Cloud, Microsoft Azure, OCI, and CoreWeave) expected in the second half of 2026. The core component is the Vera Rubin NVL72 rack with 72 Rubin GPUs and 36 Vera CPUs, built on HBM4 memory and TSMC's 3nm process.
Features
| Manufacturing Process (nm) | TSMC 3 nm process (Rubin GPU and Vera CPU) |
| Platform | Vera Rubin NVL72: 72 Rubin GPUs + 36 Vera CPUs, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6/Quantum-X800 |
| Price | No official NVIDIA list price published (third-party sources estimate DGX rack at approx. $3.5–4 million) |
| Compute Performance (FLOPS/TOPS) | 3.6 EFLOPS NVFP4 inference and 2.5 EFLOPS training per NVL72 rack; 50 PFLOPS NVFP4 per GPU |
| Release Date | Partner systems from H2 2026; full-scale production since Q1 2026 |
| Memory | 288 GB HBM4 per GPU (22 TB/s bandwidth); 20.7 TB HBM4 + 54 TB LPDDR5X per NVL72 rack |
| Availability | Full-scale production since early 2026; cloud availability (AWS, Google Cloud, Azure, OCI, CoreWeave, and others) from H2 2026 |
