NVIDIA makes OpenShell Reference Architecture available

NVIDIA frames the reference architecture as infrastructure for both agent development and operations at industrial scale.[1] Its stack combines Vera Rubin infrastructure for model workloads, Vera CPUs for agent harnesses, OpenShell software and BlueField DPUs running Sentry.[1] NVIDIA also lists RTX Spark, DGX Spark and DGX Station systems as part of the architecture.[1] The company says the design is intended to build, contain, observe and enforce safe agents across enterprise infrastructure and edge systems.[1] The architecture is available on Linux and Kubernetes infrastructure from Canonical, SUSE and Red Hat, as well as Windows with Microsoft Execution Containers.[1]

OpenShell treats agent workloads as untrusted and grants access to resources only when needed, according to NVIDIA.[2] A control plane manages sandbox lifecycles and policies, while filesystem and process access are restricted inside the isolated workload.[2] Outbound requests are checked against policy, and credential access is mediated outside the sandbox.[2] Formal verification checks whether permissions stay within an organization-defined ceiling.[2] NVIDIA says middleware can also inspect, block or change agent network traffic as an additional enforcement layer.[2] Together, these controls give platform teams several places to define and enforce boundaries, though the evidence does not specify how an organization should configure them.

NVIDIA says BlueField-4 provides a hardware-isolated trust domain outside the agent’s reach, where Sentry monitors behavior and enforces runtime policy.[2] Sentry can monitor requests and responses, network and file activity, delegated authority and selected host memory without relying on an agent’s self-reporting.[2] NVIDIA says BlueField-4 can quarantine an agent within milliseconds if Sentry detects a threat or policy violation.[2] NVIDIA also claims Sentry on BlueField-4 can boost agent capacity by up to 20%.[1] Separately, CoreWeave measured sandbox startup on Vera at more than three times the speed of x86 and 1.8-times faster performance than x86 across agentic tasks on Terminal Bench.[2] These are vendor-reported figures, not independent comparisons in the evidence.

NVIDIA says the architecture is now available and can run with agents from Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence.[1] Cloud providers including AWS, Google Cloud and Microsoft Azure can run it, according to NVIDIA.[1] The company says BBVA and Santander are exploring the architecture, SAP is integrating it with Joule Studio, and ServiceNow made its Arc personal long-running AI agent generally available with actions governed by its control tower and enforced by OpenShell.[1] The Next Web reported that NVIDIA did not provide prices.[3] The Register reported that telemetry is collected by default, with an opt-out, and that OpenShell had security-related issues listed on GitHub.[4]