NVIDIA introduced an OpenShell reference architecture at GTC Berlin on Oct. 21, pairing software controls with an optional BlueField-4 hardware layer for monitoring enterprise AI agents.
The design assigns model workloads to Vera Rubin systems and agent harnesses and tool use to Vera CPUs. OpenShell sets runtime permissions. Enterprises can add BlueField-4 data processing units (DPUs) running NVIDIA Sentry as a separate safeguard. NVIDIA’s technical blog says the hardware-isolated system examines exchanges between agents and models, along with network and file operations. NVIDIA says Sentry can isolate an agent and halt its activity within milliseconds of a boundary violation.
NVIDIA claims Vera Rubin NVL72 systems deliver 30 times the throughput per megawatt for agent workloads. Its release also says Vera CPUs provide twice the performance of x86 processors and Sentry can increase agent capacity by up to 20%. The release does not give test setups for those three headline figures. In a technical blog, NVIDIA reports that Vera CPUs delivered up to twice the performance on critical agent tasks, with a 1.8-times average across 20 Terminal Bench tasks in a CoreWeave comparison with a leading x86 CPU.
The design covers Linux and Kubernetes infrastructure offered by Canonical, SUSE and Red Hat, and Microsoft Windows deployments using Microsoft Execution Containers. NVIDIA also lists Amazon Web Services, Google Cloud and Microsoft Azure, and names RTX Spark, DGX Spark and DGX Station for edge deployments.
ServiceNow said ServiceNow Arc is generally available and governed by its Super Intelligence Control Tower with OpenShell enforcement. It said AI Passport, including enforcement through NVIDIA OpenShell and BlueField DPUs, is expected in a future release.
NVIDIA says the reference architecture is available now. Some products and features described in the announcement remain in development.
