NVIDIA has launched the OpenShell Reference Architecture, a full-stack blueprint for running enterprise AI agents inside defined boundaries — contained, continuously monitored, and governed by policy enforced in silicon. Why it matters: companies are about to hand agents real work on real systems, and NVIDIA’s answer to “how do we control them?” is to build the control into the computing architecture itself, not bolt it on afterward.
October 21, 2026 — Berlin
What happened
Announced at GTC Berlin on October 21, the OpenShell Reference Architecture extends NVIDIA’s open-source OpenShell secure runtime software into industrial-scale infrastructure for building, containing, observing and enforcing safe agents, on both enterprise infrastructure and edge systems.
The design splits the job across three pieces of NVIDIA hardware. Agents run their models on Vera Rubin infrastructure, which can serve closed and open models including NVIDIA’s own Nemotron family. Vera CPUs run the agent harnesses — the scaffolding that lets agents act. And BlueField DPUs run NVIDIA Sentry, a unified security framework that monitors agent behavior continuously and out-of-band, and enforces policy in silicon, so the guardrails do not depend on the agent’s own software behaving.
Who is behind it
The list of companies working with the architecture reads like an enterprise IT roll call: Anthropic, AWS, Cisco, CrowdStrike, Dell Technologies, Google Cloud, HPE, IBM, Meta, Microsoft, OpenAI, Palo Alto Networks, Perplexity, SAP, ServiceNow, Siemens, Dassault Systèmes, SpaceX Super Intelligence and others. The architecture runs on Canonical, Red Hat, Microsoft Windows and Microsoft Execution Containers environments, and extends to edge systems via RTX Spark, DGX Spark and DGX Station — aimed at industries from retail and healthcare to manufacturing, robotics and telecommunications.
Jensen Huang’s summary of the design is the cleanest one-liner of the launch: agents can think on Rubin, act on Vera and be governed on BlueField.
Why it matters
The agent era has a trust problem. Recent incidents — including agents going off-script inside real companies — have shown that prompt-level instructions are not a security model. OpenShell’s bet is that containment, observation and enforcement belong in the stack: a separate processor watching what agents do, with policy enforced where the agent cannot reach it. NVIDIA says the reference architecture is now available, though it did not announce pricing; it is an architecture and ecosystem play rather than a single boxed product. But when this many enterprise heavyweights attach their names on day one, the direction is clear — everything AI, tested, now comes with a warden. Expect “how are your agents governed?” to become a standard procurement question, and expect NVIDIA to want the answer to be its hardware.
Frequently asked questions
What is the NVIDIA OpenShell Reference Architecture?
It is a full-stack reference design for building, containing, observing and enforcing safe AI agents on enterprise and edge infrastructure, extending NVIDIA’s open-source OpenShell runtime software.
How does OpenShell keep agents under control?
Agents think on Vera Rubin infrastructure and act through harnesses on Vera CPUs, while BlueField DPUs run NVIDIA Sentry for continuous out-of-band monitoring of agent behavior and in-silicon policy enforcement.
Which companies are working with OpenShell?
NVIDIA names Anthropic, AWS, Cisco, CrowdStrike, Google Cloud, IBM, Meta, Microsoft, OpenAI, Palo Alto Networks, SAP, ServiceNow, Siemens and others.
Is the OpenShell Reference Architecture available now?
Yes. NVIDIA says the reference architecture is now available through its ecosystem; it did not announce pricing.
Sources: NVIDIA via GlobeNewswire (October 21, 2026)