NVIDIA launches OpenShell reference architecture for safe AI agents

NVIDIA on October 21, 2026, announced the OpenShell Reference Architecture at GTC Berlin, saying it is designed to help enterprises build, contain, observe and enforce safe AI agents across data center and edge environments. The company said the architecture is now available and extends its OpenShell secure runtime software into a full-stack system for agent development and operations.

Evidence and context

According to NVIDIA’s announcement, the architecture combines NVIDIA Vera Rubin infrastructure for model inference, Vera CPUs for agent harnesses and tool use, and BlueField-4 DPUs running NVIDIA Sentry for in-silicon enforcement. NVIDIA said the design is intended to separate governance from agent execution and to support enterprise and edge deployments running on systems such as RTX Spark, DGX Spark and DGX Station. The company also said the reference architecture runs on Linux and Kubernetes environments from Canonical, SUSE and Red Hat, and on Microsoft Windows using Microsoft Execution Containers.

NVIDIA chief executive Jensen Huang said in the company’s release that “AI capability and safety must advance together,” and added that OpenShell is intended to give agents “the freedom to do useful work within clearly defined boundaries — contained, monitored and under control.” He said that with the OpenShell Reference Architecture, “agents can think on Rubin, act on Vera and be governed on BlueField.”

The company said Vera Rubin NVL72 systems deliver 30x higher throughput per megawatt for agentic workloads, while Vera CPUs deliver 2x the performance of x86 processors with greater energy efficiency. NVIDIA also said Sentry, powered by DOCA and running on BlueField-4, can quarantine agents within milliseconds and boost agent capacity by up to 20%. Those figures are vendor-reported claims and were not independently verified in the source material.

NVIDIA said the architecture is being worked on with a broad set of partners across infrastructure, model development, cybersecurity and enterprise software, including Anthropic, AWS, Cisco, CrowdStrike, Dassault Systèmes, Dell Technologies, Google Cloud, HPE, IBM, Meta, Microsoft, OpenAI, Palo Alto Networks, Perplexity, SAP, ServiceNow, Siemens and SpaceX Super Intelligence. The company also said banks including BBVA and Santander are exploring the framework, while SAP and ServiceNow have announced integrations tied to their own agent platforms. The announcement did not provide adoption figures, revenue impact or customer deployment counts.

Separate background material from NVIDIA dated September 28, 2026 describes OpenShell as Apache-2.0 licensed and outlines controls over files, networks, tools, processes and credentials, with a Policy Prover, Supervisor, Sandbox and Gateway architecture. That background also describes Sentry as an independent hardware security layer. The latest announcement did not supply a timetable for broader customer rollout beyond saying the reference architecture is available.