NVIDIA Unveils Vera Rubin: A Leap in Performance Designed for Agent AI

At the 2026 GTC conference, NVIDIA unveiled the Vera Rubin platform, designed specifically for agent AI. The Rubin GPU utilizes TSMC's 3nm process, integrating 336 billion transistors, equipped with 288GB of HBM4 memory, and boasts a bandwidth of 22TB/s. Its FP4 inference capability reaches 50 PFLOPS (5 times that of Blackwell), and its training capability is 35 PFLOPS (3.5 times). The entire Vera Rubin NVL72 rack achieves 260TB/s bandwidth via NVLink 6, exceeding the total bandwidth of the entire global internet.

Compared to its predecessor, Blackwell, Vera Rubin represents a generational leap: single-token inference cost is reduced to 1/10, inference throughput per watt is increased tenfold, and the GPU resources required for training large hybrid expert models are reduced to only a quarter of what they were before. The platform is currently in full production and is expected to ship in the second half of 2026.