Nvidia opened its annual GTC conference in San Jose on March 16 with a keynote from CEO Jensen Huang that put AI agents at the center of the company's strategy. The biggest announcement was NemoClaw, an enterprise-grade reference stack built on OpenClaw, the open-source agent framework that Huang called the operating system for personal AI and compared in significance to HTML and Linux.
NemoClaw adds three security layers to OpenClaw: OpenShell runtime sandboxing, a privacy router, and network guardrails. The platform is open-source and hardware-agnostic, designed to plug into existing enterprise compliance systems so organizations can govern autonomous agent behavior on their own terms. In practice, it turns OpenClaw from a developer tool into something enterprises can deploy with a single command.
On the hardware side, Huang projected that combined purchase orders for Blackwell and Vera Rubin chips would reach $1 trillion through 2027 - double the $500 billion estimate from last year's conference. The Vera Rubin platform, scheduled to ship later in 2026, is built from 1.3 million components and promises 10x better performance per watt than Grace Blackwell.
The conference underscored a strategic shift at Nvidia. While GPUs remain central to AI training, agentic workloads are creating new bottlenecks in CPU orchestration, memory management, and inference speed. Nvidia is positioning itself as the full-stack infrastructure provider for the agent era, not just a chip vendor.