Nvidia CEO Jensen Huang kicked off GTC 2026 in San Jose on Monday before 30,000 attendees from 190 countries, officially launching the Vera Rubin platform — the successor to Blackwell. The Rubin GPU packs 336 billion transistors on TSMC's 3nm process, nearly doubling Blackwell's 208 billion, and delivers 50 petaflops of FP4 inference performance — a 5x leap over its predecessor.
Each Rubin GPU carries 288GB of HBM4 memory with 22 TB/s bandwidth, nearly tripling Blackwell's 8 TB/s. The platform pairs with the new 88-core Vera CPU built on Arm v9.2 architecture and connects via NVLink 6 at 3.6 TB/s per GPU. Nvidia says the platform cuts inference token costs by 10x and reduces the number of GPUs needed to train mixture-of-experts models by 4x.
Rubin entered full production in early 2026, with 200,000 to 300,000 units expected this year and partner products shipping in the second half. A more powerful Rubin Ultra variant is confirmed for 2027. Nvidia also announced a multiyear partnership with Thinking Machines Lab to deploy at least one gigawatt of Vera Rubin systems for frontier model training.
