NVIDIA’s new OpenShell reference architecture puts model inference, tool execution and security enforcement in separate parts of enterprise infrastructure. BlueField processors run security software outside the main workload, while other processors handle models and agent tools. The design extends to data centers and edge systems. Some partner products and integrations remain in development.
NVIDIA has announced a reference architecture extending OpenShell, its secure environment for running AI agents, across enterprise infrastructure. The announcement at the Berlin GTC event places model inference, agent tools and security enforcement in separate parts of the design. OpenShell software already existed; the new development connects that runtime to data-center and edge hardware.[1], [2]
Vera Rubin supplies model computing, while Vera central processing units run agent tools. BlueField data processing units run Sentry security software in a separate domain. NVIDIA says this lets observation and policy enforcement operate outside the main workload, constraining behavior and network access according to agent permissions. The company aims to avoid relying solely on the agent or its host operating system for security boundaries.[1]
NVIDIA names RTX Spark, DGX Spark and DGX Station among systems covered beyond data centers. Enterprise software company SAP is connecting its Joule Studio runtime, while ServiceNow describes policy decisions through its control tower with enforcement through OpenShell. Operating-system, cloud and server partners are also involved. NVIDIA says some products and features remain under development, with delivery schedules subject to change.[1]
