News

NVIDIA’s GTC Keynote Puts AI Agents, Inference, and Infrastructure at Center Stage

• By John K. Waters
• 03/16/2026

NVIDIA Chief Executive Jensen Huang used his GTC 2026 keynote on Monday to argue that the center of the AI market is shifting from training large models to running them, securing them, and turning them into software agents and physical machines. In a presentation streamed online from the San Jose McEnery Convention Center, Huang paired that message with a familiar playbook: more chips, more systems, and more software meant to bind customers more tightly to the company’s platform.

This year's conference, which runs from March 16-19, is large even by current AI-event standards, attracting more than 30,000 attendees from over 190 countries, the organizers said. (The keynote is available online without registration.) That scale helps explain why Huang’s annual appearance now functions as both a product launch and a market signal for the broader AI industry.

The clearest theme of this year’s keynote was inference, the stage during which a trained AI model handles real-world tasks by generating answers, making predictions, or taking action. “The inflection point of inference has arrived,” Huang said, arguing that AI is moving from the training phase into broader deployment. He said demand for computing had risen sharply and that Nvidia expects at least $1 trillion in revenue between 2025 and 2027. He also positioned Nvidia as a full-stack supplier for what the company calls “AI factories,” offering not only chips, but also the networking, software, and system design needed to run them.

On hardware, Huang pointed to the Vera Rubin platform as the next full-stack system for agentic AI and previewed a later architecture called Feynman. GTC’s live blog said Vera Rubin will comprise seven chips, five rack-scale systems, and one supercomputer, while Feynman is expected to add a new CPU called Rosa, along with updated networking and scale-out components. The company also used the event to further extend Blackwell's life, including a Blackwell Ultra platform that is designed for reasoning workloads and large-scale inference.

If the chip roadmap was the spine of the keynote, the more consequential message may have been NVIDIA's effort to move higher into the software stack. Huang highlighted OpenClaw, an open-source project for AI agents, and said it would support it across its platform. He also introduced OpenShell and NemoClaw, which are designed to add policy controls, network guardrails, and privacy routing, enabling those agents to be used within enterprises. “Every company in the world today needs to have an OpenClaw strategy," he said.

That matters because it shows how NVIDIA is trying to make itself harder to replace. For years, the company’s grip on AI came largely from accelerators and CUDA. On Monday, Huang described a broader ambition in which supplies not just processors but also the operating environment for enterprise agents and the infrastructure to deploy them safely. Keep in mind that these claims come from, and many of the product timelines remain forward-looking. Still, the direction is clear.

The same logic showed up in the systems business. The company used the keynote and related GTC materials to push DGX Spark and DGX Station as deskside machines for developers and enterprise teams building autonomous agents locally before scaling them to data centers. Huang said DGX Station can run models of up to 1 trillion parameters and that systems will ship in the coming months from partners including ASUS, Dell, GIGABYTE, MSI, and Supermicro, with HP to follow later in the year.

Huang also spent time on what he calls physical AI, which is effectively the company’s umbrella term for robotics, autonomous vehicles, and real-world systems that use AI models to perceive and act. He said that new automaker partners for its robotaxi platform include BYD, Hyundai, Nissan, and Geely, and he linked that effort to industrial robotics partners such as ABB, Universal Robots, and KUKA. The message was that the company wants the same stack that runs digital agents to extend into warehouses, factories, and vehicles.

Some of the presentations stretched further than that. Huang announced that the company was "going to space," tying the idea to future data center designs built around Vera Rubin. That line will attract attention, but it was peripheral to the more grounded commercial point of the keynote, which was about locking in the next phase of enterprise AI spending around inference, agents, and the infrastructure needed to manage them.

For investors and enterprise buyers, the practical takeaway is less theatrical. NVIDIA is still selling chips, but it's increasingly selling a market theory in which AI becomes persistent infrastructure rather than a series of isolated applications. In that version of the future, demand does not stop at model training. It extends to the continual running of agents, the securing of those agents, and the deployment of AI into physical systems. Monday’s keynote was the latest attempt to show that NVIDIA intends to supply every layer of that stack.