NVIDIA proposes stopping AI agents with its own chips

NVIDIA unveiled the OpenShell Reference Architecture on Wednesday at GTC Berlin, an open design for putting limits on AI agents, the programs that use tools and data without someone dictating each step. The company says it is available now but has published no price.

What is new is where the limits are enforced. OpenShell is an open-source runtime (Apache 2.0 license) that restricts what an agent can do. NVIDIA now wraps it in its own hardware: models run on Vera Rubin systems, the agent on Vera CPUs, and BlueField-4 chips run NVIDIA Sentry, a monitor that watches from outside the main system and enforces policy in hardware.

NVIDIA says Sentry cuts off an agent that crosses its perimeter "in milliseconds" and raises by up to 20% the number of agents that fit on the system. It also claims 30 times more performance per megawatt with Vera Rubin NVL72. These are the manufacturer's figures, with no independent testing.

"AI capability and security must advance together," said Jensen Huang, the company's founder and CEO.

The design spans NVIDIA's own desktop machines (RTX Spark, DGX Spark, DGX Station) up to the AWS, Google Cloud and Microsoft Azure clouds, with Linux from Canonical, SUSE and Red Hat, plus Windows. Models that can run on it include those from Anthropic, OpenAI, Google and Meta, along with Cohere, Perplexity, Reflection and SpaceX Super Intelligence. Among software companies, SAP already includes it in Joule Studio, and ServiceNow's Arc agent builds on it. The banks BBVA and Santander are only exploring it.

OpenShell began in March as open software inside NVIDIA's Agent Toolkit. Wednesday's step turns it into an architecture that adds the company's chips. For anyone who gives an agent access to email, files or internal systems, that solves the control problem but ties the user to NVIDIA hardware; the open runtime does not require it.