Nvidia chief executive Jensen Huang used his GTC Taipei keynote on June 1 to declare that the age of autonomous agents has arrived, and he backed the claim with new hardware across the data center, the desktop and the physical world. Huang announced that the Vera Rubin platform, Nvidia’s next data center system, has reached full production, and he framed nearly every product the company revealed in Taipei around software agents that observe, reason, plan and act with little human input.
The keynote at the Taipei Music Center reflected a change in how Nvidia describes itself. Huang said the company now sells AI infrastructure rather than chips alone, and he argued that compute has become a direct source of revenue for the businesses that buy it. He pointed to coding platforms where developer commits have nearly tripled in the first months of 2026 as evidence that agents are already doing useful work. For technology leaders, the announcements map where data center spending, enterprise software and personal computing are heading.
Vera Rubin Reaches Full Production
Vera Rubin is a five-rack system that Nvidia treats as one large computer for agentic workloads. The platform combines Vera Rubin NVL72 systems, the new Vera CPU, Groq 3 LPX inference trays, Spectrum-6 Ethernet racks and Vera BlueField-4 STX storage. Each NVL72 rack links 36 Vera CPUs and 72 Rubin GPUs through the sixth-generation NVLink switch, with ConnectX-9 network cards and BlueField-4 data processing units handling traffic and security.
Nvidia says the rack delivers up to 10 times higher inference performance per watt and 10 times lower cost per token than its prior generation, and that pairing it with Groq 3 LPX raises throughput per watt as much as 35 times for trillion-parameter models. The cable-free, hose-free and fanless tray design cuts assembly from two hours to five minutes per compute tray, and a fully liquid-cooled design runs at 45 degrees Celsius so it fits existing data centers.
Huang said the Vera Rubin supply chain is twice the size of the prior Grace Blackwell effort, spanning 150 partners in Taiwan and more than 350 factories across 30 countries. Production shipments begin this fall.
A CPU And Software Stack Built For Agents
The Vera CPU is Nvidia’s first standalone data center processor, with 88 cores and an on-chip fabric the company designed for agents rather than human operators. Huang argued that billions of agents will run continuously and demand far lower latency than people do, creating a processor market that did not exist before.
Nvidia also moved its Spectrum-X Ethernet Photonics networking into production, describing it as the first 200 gigabit per second Ethernet switch with co-packaged optics and naming CoreWeave, Lambda and Oracle Cloud Infrastructure among early adopters. A separate framework called DSX helps operators design and run AI factories, and Nvidia said one configuration fits 40% more GPUs within the same power budget.
On the software side, Nvidia introduced an Agent Toolkit that bundles models, an agent harness and an enterprise runtime, alongside a secure runtime called OpenShell that isolates each agent and enforces policy. The company released Nemotron 3 Ultra, a 550-billion-parameter mixture of experts model it says runs inference five times faster and costs about 30% less than leading open alternatives. Verified Nvidia agent skills are now available inside the Claude Code plug-in marketplace and the Hermes Skills Hub.
Nvidia Enters The PC Market
Huang announced RTX Spark, a chip built with MediaTek that brings 1 petaflop of AI performance to Windows laptops and compact desktops. It pairs a Blackwell RTX graphics processor that has 6,144 CUDA cores with a 20-core Grace CPU, and Nvidia positioned it as the foundation for personal computers that run agents locally rather than calling a cloud server.
The company unveiled a Windows lineup that includes a laptop, an always-on desktop agent box and a deskside DGX Station for Windows capable of running frontier models up to 1 trillion parameters on the desk. Partners including Asus, Dell, Gigabyte, HP, MSI and Supermicro begin shipping DGX Station systems this month. Adobe is rebuilding Photoshop and Premiere for RTX Spark, with versions Nvidia says run twice as fast and work with agents.
Physical AI Moves To The Foreground
Nvidia extended its agent message into robotics and vehicles. It launched Cosmos 3, an open world foundation model built on a mixture-of-transformers design that learns from teleoperation, simulation and re-projected video so robots can reason about their surroundings. The company said its Drive Hyperion vehicle platform now reaches services representing about 97% of the world’s mobility market, and it introduced Alpamayo 2 Super, an open reasoning model for self-driving research paired with a reinforcement learning trainer and a scenario generator. For robotics labs, Nvidia released an open humanoid robot reference design built on its Jetson Thor module. A set of media tools rounded out the day, including a synthetic video detector Nvidia says flags AI-generated footage with about 92% accuracy in 22 milliseconds.
The Limitations
Most of the performance numbers come from Nvidia and have not been independently tested, and Vera Rubin will not ship in volume until the fall, so buyers cannot yet validate the cost-per-token claims in their own workloads. The new Windows machines and RTX Spark systems also arrive later this year, which leaves their software ecosystem and agent tooling unproven outside controlled demos. Enterprise agent runtimes raise governance and security questions that products such as OpenShell address in principle but have not faced at production scale. Competition is intensifying as well, with AMD pushing its Instinct accelerators and cloud providers expanding custom silicon such as AWS Trainium, Google Ironwood and Microsoft Maia.
For technology decision makers, the keynote sharpened a choice that will define AI budgets. Huang argued that performance per watt and the runtime that surrounds the model now matter as much as the chip itself, which means architecture decisions made over the next year will shape both capability and cost long after the hardware lands.
