NVIDIA Unveils Vera Rubin POD Cluster: Five Racks, Seven Chips, 1.2 Quadrillion Transistors

NVIDIA announced the Vera Rubin POD, a cluster of five specialized rack systems based on third-generation MGX architecture, designed for the era of agentic AI. The co-designed platform integrates seven chips covering compute, networking, and storage, making it the most complex AI platform from NVIDIA. The complete Vera Rubin POD includes 40 racks, 1.2 quadrillion transistors, nearly 20,000 NVIDIA dies, and 1152 Rubin GPUs, achieving 60 EFLOPS AI performance and about 10 PB/s aggregated interconnect bandwidth. The five rack platforms are Vera Rubin NVL72 (72 Rubin accelerators and 36 Vera CPUs per rack, connected via NVLink 6), Groq LPX (256 LPU accelerators for low-latency inference), Vera CPU rack (up to 256 processors for RL environments), BlueField-4 STX storage rack with CMX context memory, and Spectrum-6 SPX networking rack. The NVL72 rack delivers up to 4x better training performance and 10x better inference performance per watt compared to Blackwell. All MGX racks use hot-water cooling at 45°C inlet, support dynamic power redistribution, and feature intelligent power smoothing to reduce peak loads by 25%. The Vera Rubin NVL72 is scheduled for release in the second half of 2026.