Nvidia Launches OpenShell Architecture for Silicon-Level AI Agent Containment

Nvidia announced the general availability of its OpenShell Reference Architecture at GTC Berlin, deploying a full-stack framework to secure and govern autonomous AI agents across enterprise and edge environments. The architecture integrates Vera Rubin infrastructure for model inference, Vera CPUs for agent tool execution, and BlueField-4 DPUs that run the Sentry security framework for in-silicon policy enforcement. Nvidia claims the Vera Rubin NVL72 systems achieve 30x higher throughput per megawatt for agentic workloads, while Sentry on BlueField-4 quarantines out-of-bound agents in milliseconds and increases agent capacity by up to 20%. Enterprise operating systems from Microsoft, Red Hat, and Canonical support the deployment, with native integration underway from major cybersecurity vendors including CrowdStrike, Cisco, and Palo Alto Networks. SAP and ServiceNow are directly integrating OpenShell into their enterprise platforms, with ServiceNow's Arc agents now governed at runtime by the new Nvidia infrastructure. "As enterprises prepare to deploy billions of AI agents, safety and control must be built into the computing architecture." — Jensen Huang.