Nvidia unveils Feynman GPU architecture and NemoClaw at GTC 2026

Jensen Huang walked onto the GTC 2026 stage in San Jose on March 16 and did what he usually does: spent two hours redefining what Nvidia is building toward. This year, the headline was the Feynman architecture, a next-generation GPU design built from the ground up with agentic AI and inference workloads in mind. It is not a speed bump on the existing roadmap. It is a rethink of how chips should behave when the workload is not just training a model, but running one continuously, making decisions, and talking to other systems.

The conference drew developers from 190 countries. That number alone says something about where GTC sits in the industry calendar right now. It has quietly become the event that enterprise AI teams watch most closely, more so than any consumer tech launch.

What the Feynman architecture actually changes

Nvidia named the architecture after physicist Richard Feynman, which fits the theme. Feynman was known for building mental models of complex systems from first principles. The GPU architecture carrying his name is designed around the assumption that AI inference is now a continuous, multi-agent process rather than a one-shot calculation. Previous architectures optimized heavily for training throughput. Feynman shifts that balance.

The specific details Huang shared point to improvements in how the chip handles context switching between agent tasks, lower latency for token generation, and better energy efficiency per inference operation. These are not abstract claims. For companies running large-scale AI deployments, inference costs have been growing faster than training costs for the past 18 months. A chip that reduces the cost per query by even 20 percent changes budget conversations at the enterprise level.

NemoClaw: Nvidia's open-source agent platform

The second major announcement was NemoClaw, an open-source platform for building and deploying enterprise AI agents. The name combines Nvidia's existing NeMo framework with a new runtime layer designed for multi-step autonomous tasks. Where earlier AI tools handled single prompts or simple chains, NemoClaw is built to manage agents that plan, execute, check their own outputs, and loop back when something goes wrong.

Huang described use cases ranging from automated code review pipelines to supply chain monitoring agents that can query multiple data sources and generate action recommendations without a human in the loop for each step. The open-source release is significant. It means enterprise teams can inspect the code, modify it, and deploy it on their own infrastructure without being locked into Nvidia's cloud offerings. That said, NemoClaw is clearly optimized to run well on Feynman-class hardware.

The platform includes a tool-calling layer that lets agents connect to external APIs, a memory module for maintaining context across sessions, and an evaluation harness so teams can measure whether their agents are actually completing tasks correctly. That last part matters more than it might seem. One of the persistent problems with agentic AI in production is that agents fail quietly, completing a task in a way that looks correct but produces wrong outputs. Built-in evaluation tooling addresses that directly.

Why this keynote was different from previous years

GTC has historically been a developer conference where Nvidia announced hardware and frameworks that would ship months later. This keynote felt more like a product launch event. Both Feynman and NemoClaw appear to be closer to production-ready than Nvidia's typical conference reveals. The company has been under pressure to show that its hardware advantage extends into the inference era, not just the training era that made it dominant.

There is also competitive context here. AMD and Intel have both been pushing harder on AI inference chips over the past year. Custom silicon from Google, Amazon, and Microsoft has taken real workloads off Nvidia hardware in data centers. The Feynman architecture is Nvidia's clearest answer yet to the question of whether the company can stay relevant as AI workloads shift from training runs to always-on inference.

What comes next

Nvidia has not announced a shipping date for Feynman-based products, but the architecture reveal at GTC typically precedes availability by six to twelve months. NemoClaw is available now on GitHub under an Apache 2.0 license. Developer documentation and example agent templates were published alongside the keynote. Nvidia's next major hardware announcement is expected at Supercomputing 2026 in November, where the company typically shares performance benchmarks against competing silicon.