Confirmed — Wednesday: Nvidia announced its OpenShell Reference Architecture at GTC Berlin and said it is available now. The design explains how AI agents can operate within set limits across enterprise infrastructure and edge machines. Nvidia says Anthropic, OpenAI and other companies are working with it.
What is OpenShell in this design?
OpenShell is Nvidia’s open-source runtime software, which sets a boundary around what an agent can do. Jensen Huang said in his keynote that the software is released under the Apache 2.0 licence. Nvidia says OpenShell is also part of its agent safety platform launched in September.
How are the hardware roles divided?
Nvidia says models run on Nvidia Vera Rubin systems, while Vera CPUs run the software that drives agents and their tools. BlueField-4 data processing units run Nvidia Sentry, a separate monitoring component that watches agent behaviour from outside the main system. This division assigns model computation, agent software and monitoring to distinct parts of the design.
The architecture also covers edge machines, including RTX Spark, DGX Spark and DGX Station systems. Nvidia showed RTX Spark laptops built to run agents locally last week. Its design therefore describes operation both on broader infrastructure and on local machines.
What happens if an agent crosses its boundary?
Nvidia says Sentry quarantines and stops an agent in milliseconds if it attempts to move beyond its software boundary. The company also claims Sentry increases agent capacity by up to 20%. The supplied announcement does not provide independent test results or explain how those outcomes were measured.
Huang described the intended division as agents thinking on Rubin, acting on Vera and being governed on BlueField. The available details identify the components and the claimed response to boundary violations. Nvidia did not give prices for the design.
Which companies are connected to OpenShell?
Nvidia said agents from Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence can run on OpenShell infrastructure. Security companies named as adding support include 1Password, Check Point, CrowdStrike, Cisco, Fortinet, Okta, Palo Alto Networks and Zscaler. These descriptions indicate different forms of participation rather than one uniform partnership.
Nvidia said BBVA and Santander are exploring the design for stronger agent safety and governance. SAP is building it into Joule Studio runtime, while ServiceNow made Arc generally available on Wednesday with actions enforced by OpenShell. The information supplied does not specify when the explorations will become deployments.
Which operating platforms and infrastructure are included?
The architecture runs on Linux and Kubernetes from Canonical, SUSE and Red Hat, and on Windows through Microsoft Execution Containers. Nvidia also lists AWS, Google Cloud and Microsoft Azure as supported cloud platforms. Dell, Cisco, HPE, Lenovo and Supermicro are building infrastructure with the design.
Nvidia said Claude Managed Agents from Anthropic have been paired with OpenShell since September. Anthropic said the arrangement allows companies to limit Claude agents’ actions and log every action. The announcement describes this as one example of how the runtime can be used with an agent service.
Nvidia has made the architecture available, but pricing and commercial terms remain undisclosed. The announcement identifies technical components and company activity, while leaving independent performance evidence open. Further deployment details would show how the design works in practice.