NVIDIA has introduced its OpenShell Reference Architecture, a hardware-and-software platform designed to provide runtime security, isolation, monitoring, and policy enforcement for autonomous AI agents operating across enterprise data centers, cloud infrastructure, and edge systems. Announced at GTC Berlin, the architecture combines NVIDIA Vera Rubin infrastructure for AI inference, Vera CPUs for agent execution, and BlueField-4 DPUs running NVIDIA Sentry security software. The architecture separates AI agent execution from security enforcement, allowing operators to monitor and restrict agent behavior through controls outside the agent’s software environment.

The architecture builds on NVIDIA OpenShell, an open-source secure runtime that establishes execution boundaries for autonomous agents. Vera Rubin NVL72 systems provide accelerated computing for inference and model workloads, while Vera CPUs execute agent harnesses, orchestration logic, and tool interactions. BlueField-4 DPUs run NVIDIA Sentry, a security framework built on NVIDIA DOCA that monitors agent behavior and enforces policies. According to NVIDIA, Sentry can quarantine and stop an agent within milliseconds if it attempts to operate outside its authorized software boundary. NVIDIA also claims its BlueField-4 and Sentry implementation can increase agent capacity by up to 20%, without detailed benchmark conditions in the announcement.

NVIDIA is positioning OpenShell as a common infrastructure architecture for enterprise AI agent deployments, with support spanning Linux, Kubernetes, Windows, and Microsoft Execution Containers. The company identified collaborators including AWS, Google Cloud, Microsoft, Cisco, CrowdStrike, Palo Alto Networks, Dell Technologies, HPE, IBM, SAP, and ServiceNow. Dell, Cisco, HPE, Lenovo, and Supermicro are among system manufacturers building infrastructure around the architecture, while CoreWeave, Crusoe, IREN, Lambda, Nebius, and Oracle Cloud Infrastructure are working on OpenShell-ready environments. The reference architecture is available now; individual hardware capabilities and deployments have separate schedules.

Key Technical Highlights

Executive Perspective

“With the OpenShell Reference Architecture, agents can think on Rubin, act on Vera and be governed on BlueField.” — Jensen Huang, NVIDIA founder and CEO

Analysis: BlueField Moves Into AI Agent Security Enforcement

The most significant architectural development is NVIDIA’s use of BlueField-4 as an independent enforcement layer for autonomous AI workloads. DPUs have traditionally offloaded networking, storage, virtualization, and infrastructure security from host CPUs. OpenShell extends this model to agents that autonomously access applications, invoke tools, and interact with enterprise data. Moving enforcement to the DPU introduces an additional control point, although effectiveness depends on policy coverage, identity integration, and visibility into agent interactions.

The broader significance is a three-layer model for enterprise AI agents: GPU-based inference, CPU-based agent execution, and DPU-based security enforcement. The architecture connects agent security to Kubernetes, cloud platforms, and cybersecurity systems. NVIDIA has announced the reference design and ecosystem participation, but has not quantified production deployments or provided independently validated security benchmarks. Open questions include policy distribution across agent fleets, application-level authorization, and controls spanning multiple clouds.

NVIDIA Company Profile

We’re tracking AI infrastructure developments. Follow ongoing coverage at Converge Digest AI Infrastructure.