Nvidia’s COMPUTEX 2026 Blitz: Vera Rubin in Full Production, a CPU “Built for Agents,” and a PC You Just Ask

Jensen Huang took the stage at GTC Taipei during COMPUTEX 2026 with a thesis he repeated like a drumbeat: agentic AI is here, it works, and it makes money — every token is now a revenue unit. The keynote that followed was Nvidia’s densest announcement run of the year, spanning a new flagship platform in full production, the company’s first CPU designed for AI agents rather than humans, an AI-native PC push, and a frontier model for physical AI.

The centerpiece: Vera Rubin is in full production. Huang called the platform the most ambitious project in Nvidia’s history, with some 40,000 engineers involved and a supply chain twice the size of Grace Blackwell’s. Assembly that took two hours per Grace Blackwell rack now takes five minutes, and the system unifies five purpose-built racks into one integrated AI supercomputer aimed squarely at agentic workloads. Shipments to AWS, Google Cloud, Microsoft Azure and Oracle begin this summer, with Samsung, SK Hynix and Micron supplying HBM4 memory — and OpenAI, Anthropic and SpaceX named among early adopters.

Alongside Rubin came Vera, Nvidia’s first CPU purpose-built for AI agents, which the company says completes tasks 1.8x faster than comparable x86 processors. "This CPU is built for agents," Huang said. "All the CPUs of the past we built for humans." It is a pointed claim: the CPU — the one part of the data center Nvidia historically left to Intel and AMD — is being redefined around agentic workloads Nvidia intends to own end to end.

On the consumer side, Huang introduced RTX Spark, a superchip pairing a Blackwell RTX GPU with 6,144 CUDA cores and a 20-core Grace CPU, built with Microsoft to make Windows machines agent-native. "For forty years, you launched apps," Huang said. "With RTX Spark, you ask — and the PC does the work." Laptops and compact desktops arrive this fall from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, with Acer and GIGABYTE to follow.

The physical-AI layer got its own headline act: Cosmos 3, an open world foundation model built on a mixture-of-transformers architecture that can act as a vision-language model, world model, simulator or robot policy — understanding and generating text, images, video, sound and actions with physics accuracy, and cutting robot training cycles from months to days. Nvidia paired it with the Isaac GR00T reference platform for academic research — a six-foot, 150-pound humanoid with 31 degrees of freedom running on Nvidia’s Thor computer — plus an open-source toolkit of agents and skills, and robotics partnerships spanning the U.S., Europe, South Korea and China’s Unitree.

The keynote also lands days after Nvidia committed to pouring $150 billion a year into Taiwan and building a Taipei headquarters — a reminder that the island Huang calls the epicentre of the AI revolution is now both Nvidia’s supply chain and its stage.