Ship a New Architecture Every Year

At Computex in June 2024, secondary reports say, Huang described a one-year rhythm and laid out a roadmap: Blackwell Ultra in 2025, Rubin and Vera in 2026 and Rubin Ultra in 2027. Feynman was added for 2028 at GTC in March 2025. The wording comes from secondary coverage, so we paraphrase it here. So far the rhythm has held: Nvidia unveiled the Vera Rubin platform at GTC 2026 (16 to 19 March), and its August 2026 10-Q says Vera Rubin “began production shipments in the third quarter of fiscal year 2027” (August to October 2026). Rubin Ultra and Feynman remain plans.

Why annual. On the Lex Fridman Podcast (#494, March 2026) Huang said AI model architectures “are being invented about once every six months.” A hardware cadence of two or three years cannot keep pace with that. A faster rhythm means customers always have a reason to buy the newest generation, and competitors are measured against a chip that is already being replaced.

Doesn’t transfer without the machine behind it. An annual cadence needs tens of thousands of engineers (about 31,000 of Nvidia’s roughly 42,000 staff are in R&D), suppliers who can keep up, and customers who will buy every year. A roadmap is a plan, not a guarantee. For most businesses, shipping faster than customers can absorb just creates inventory.