NVIDIA GTC 2026 Keynote: Jensen Huang Unveils Vera Rubin Platform and Six New Chips

NVIDIA CEO Jensen Huang took the stage at the SAP Center in San Jose on Monday to deliver the GTC 2026 keynote, formally unveiling the complete Vera Rubin GPU platform to a packed arena of over 30,000 attendees from more than 190 countries. True to his pre-conference promise, Huang introduced "several chips the world has never seen before" — including the first architectural details of the next-generation Feynman platform designed for AI agent reasoning workloads.

The Vera Rubin Platform

Vera Rubin, named after the astronomer whose work confirmed the existence of dark matter, represents NVIDIA's most ambitious GPU platform to date. The flagship chip features 336 billion transistors, 288 GB of HBM4 memory with 13 TB/s bandwidth, and delivers 50 PetaFLOPS of FP4 performance — a generational leap over the current Blackwell architecture. NVIDIA confirmed that Vera Rubin has entered full-scale mass production, with samples already shipping to customers including Microsoft, Google, Amazon, Meta, and Thinking Machines Lab.

The platform encompasses six new chips across the compute, networking, and switching stack. The NVL144 configuration — 144 Vera Rubin GPUs connected via NVLink in a single rack-scale system — delivers aggregate performance equivalent to hundreds of current-generation systems, enabling training runs that were previously impractical.

Feynman: The Agent Architecture

Perhaps the biggest surprise was the first look at Feynman, NVIDIA's next-generation architecture after Vera Rubin. Named after the physicist Richard Feynman, the architecture is designed specifically for AI agent reasoning and long-term memory workloads. Huang described it as "silicon designed for thinking, not just computing" — hardware optimized for the kind of multi-step, branching reasoning that agentic AI requires.

Feynman will feature a redesigned memory hierarchy optimized for the massive context windows and persistent memory stores that AI agents need to maintain state across long-running tasks. NVIDIA provided no timeline for Feynman beyond confirming it is "in active development," suggesting a 2028 or 2029 production date.

Open Frontier Models Panel

Wednesday's agenda includes a panel discussion on open frontier models featuring leaders from Cursor, Mistral, AI2, and Thinking Machines Lab — Mira Murati's new venture. The panel reflects NVIDIA's strategic interest in maintaining a diverse AI ecosystem where multiple model providers compete, rather than a world dominated by two or three closed-source providers.

GTC 2026 runs through March 19, with over 700 sessions covering AI factories, agentic AI, physical AI, robotics, and vertical applications. The keynote replay is available on NVIDIA's website.