Watch Nvidia's Computex 2026 Keynote Live (GTC Taiwan)

Author: Jensen Huang, Founder and CEO, NVIDIA Collection: Computex 2026

Summary

Jensen Huang delivered NVIDIA's 134-minute GTC Taiwan keynote at Computex 2026, organized around a single thesis: "useful AI has arrived." Agentic AI — AI systems that observe, reason, plan, use tools, and act autonomously — has crossed from research into production, validated by GitHub data showing 90 million pull requests merged, 1.4 billion commits, and 20 million new repositories per month in early 2026, with commits nearly tripling year-over-year (Image 205, 22:55). The core architectural shift is that the application pattern of the future is not software binaries but agents: a large language model (LLM) sitting inside a harness that orchestrates memory management, tool use, and multi-agent coordination. Every product announcement in the keynote — Vera Rubin, Vera CPU, RTX Spark, Nemotron 3 Ultra, Cosmos 3, Isaac GR00T — was framed as infrastructure for this agentic computing pattern.

NVIDIA announced Vera Rubin is in full production as a five-rack-system ("pod scale") supercomputer explicitly built to run agents rather than just run AI inference. The Vera Rubin NVL72 compute tray contains 6 trillion transistors with 18,000 components on a single board; the full pod comprises NVL72 compute, an NVLink switch tray, a Vera CPU tray (256 CPUs), a Groq 3 LPX tray for low-latency inference, and a Vera BlueField-4 STX storage/security tray. The Vera CPU — NVIDIA's first CPU built for agents rather than humans — features 88 Olympus cores at 10-wide instruction decode, 1.2 TB/s LPDDR5X ECC memory, 3.4 TB/s core-to-core bandwidth, PCIe Gen 6, and NVLink chip-to-chip, delivering 3x faster SQL processing (2.2 min vs 0.7 min vs x86) and 6x faster real-time stream processing (18 ms vs 3 ms) in NYSE benchmarks (Images 763, 767).

The PC was declared reinvented for the first time in 40 years via a new RTX Spark product family: a laptop SoC (N1X) combining a Blackwell RTX GPU (6,144 CUDA cores, 1 petaFLOP FP4 AI) with a 20-core Grace CPU co-designed with MediaTek, fused by NVLink, with 128 GB unified memory on TSMC 3nm. The DGX Station for Windows delivers 20 petaflops, 8 TB/s memory bandwidth, and 768 GB unified memory — enough to run a trillion-parameter model locally. A roadmap slide revealed the next generation PC chip family: Rubin (2027-2028) using Vera CPU and Vera Rubin Spark with LPDDR6, and Feynman (2029-2030) using Rosa CPU and HBM Next (Image 1059, 112:00).