At GTC Berlin, NVIDIA announced the NVIDIA OpenShell Reference Architecture, a full-stack infrastructure design engineered to build, contain, observe, and enforce safety policies across autonomous artificial intelligence agents operating on enterprise networks and edge environments.
The newly unveiled architecture builds directly upon the foundation of NVIDIA OpenShell secure runtime software, broadening its capabilities into an industrial-scale framework for both agent development and ongoing operational management.
A Specialized Hardware and Software Stack
The OpenShell Reference Architecture structures agent governance across dedicated compute layers. At its core, the framework integrates NVIDIA Vera Rubin infrastructure designed to run demanding enterprise workloads, including open and proprietary models built with NVIDIA Nemotron. Complementing this, NVIDIA Vera CPUs manage the execution of agent harnesses, while the open-source NVIDIA OpenShell software handles runtime containment.
Hardware-level oversight is delivered through NVIDIA BlueField DPUs running the NVIDIA Sentry unified security software framework. This configuration enables continuous, out-of-band monitoring of agent operations alongside in-silicon policy enforcement, preventing unauthorized actions before they reach production networks.
Addressing the dual requirements of capability and governance, NVIDIA founder and CEO Jensen Huang emphasized the necessity of embedding constraints into computing infrastructure. Huang stated that as enterprises prepare to deploy billions of AI agents, safety and control must be built directly into the computing architecture. He noted that OpenShell was designed to grant agents operational latitude within defined parameters, allowing them to think on Rubin, act on Vera, and remain governed on BlueField.
Extending Safe Intelligence to the Edge
Beyond central enterprise data centers, the OpenShell Reference Architecture extends governance to distributed deployments at the network edge. Organizations operating in sectors such as retail, healthcare, manufacturing, robotics, software development, and telecommunications can deploy the framework using NVIDIA RTX Spark, NVIDIA DGX Spark, and NVIDIA DGX Station systems.
To support diverse computing environments, the architecture operates across major enterprise operating systems and container orchestration platforms. This includes Linux and Kubernetes configurations managed by Canonical, SUSE, and Red Hat, alongside Microsoft Windows deployments utilizing Microsoft Execution Containers (MXC). The reference design also supports cloud operating platforms across Amazon Web Services, Google Cloud, and Microsoft Azure.
Ecosystem Integration and Enterprise Adoption
Enterprises running OpenShell infrastructure can host AI agents developed by prominent model builders, including Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection, and SpaceX Super Intelligence. NVIDIA is also collaborating with model partners to deploy OpenShell within advanced model research environments.
A broad array of cybersecurity providers is incorporating support for the OpenShell Reference Architecture into their agent security portfolios. These partners include 1Password, Check Point, CrowdStrike, Cisco, F5, Fortinet, Microsoft, Okta, Palo Alto Networks, TrendAI, and Zscaler.
In the financial services sector, major European banking institutions such as BBVA and Santander are actively exploring the OpenShell Reference Architecture as an operational framework to bolster agent security, safety, and regulatory compliance.
Enterprise software developers are also building dedicated agent platforms aligned with the architecture. Cadence, Dassault Systèmes, IBM, SAP, Siemens, and Synopsys are working with OpenShell, while Palantir is collaborating with NVIDIA to deploy secure agent systems directly for commercial clients. SAP is integrating OpenShell into the Joule Studio runtime within its SAP Business AI Platform to link corporate governance rules with agent execution environments.
Instant Policy Enforcement and Hardware Availability
Highlighting production rollouts, ServiceNow announced the general availability of ServiceNow Arc, a personal, long-running AI agent. Actions taken by ServiceNow Arc are monitored at runtime by the ServiceNow Super Intelligence Control Tower and strictly enforced via NVIDIA OpenShell. Furthermore, ServiceNow's AI Passport authority layer relays administrative policies directly to OpenShell and NVIDIA Sentry, allowing enterprises to revoke an agent's access across hardware and software within seconds if policy anomalies arise.
System manufacturers including Dell Technologies, Cisco, HPE, Lenovo, and Supermicro are developing hardware platforms built to OpenShell specifications. Cloud service providers such as CoreWeave, Crusoe, Firebird, Firmus, GMI Cloud, IREN, Lambda, Nebius, Nscale, Oracle Cloud Infrastructure, and Vultr are preparing OpenShell-ready environments.
Additionally, inference and post-training service providers—including Baseten, Deepinfra, Fireworks, and Together AI—are embedding OpenShell integrations directly into their model development and serving pipelines. The NVIDIA OpenShell Reference Architecture is available immediately for deployment.
