Red Hat and Nvidia are expanding their collaboration around Red Hat AI Factory with Nvidia. They are working together on the OpenShell Reference Architecture for more secure AI agents, presenting a joint reference design for inference, and promising Day 0 support for Nvidia Vera Rubin.
According to John Fanelli, VP of Enterprise Software at Nvidia, an agent should not be responsible for monitoring its own boundaries. “Those controls belong outside the agent, in the runtime and the infrastructure beneath it,” he says.
That is why both companies are working on the OpenShell Reference Architecture. OpenShell is an open-source sandbox runtime for agents. Red Hat and Nvidia are among the key contributors.
How does this work in practice? Each agent runs in isolation within its own sandbox, without direct network access. An agent is granted access only if the policy explicitly permits it (deny-by-default). Credentials are managed outside the agent. In addition, a policy prover verifies that proposed permissions stay within the limits the organization sets. Red Hat ultimately aims to integrate OpenShell as native functionality into Red Hat AI and AI Factory.
Inference under control
Agents generate a significant amount of inference traffic. That’s why the companies are presenting a joint reference design based on the Nvidia Inference Hub design. It combines MaaS access management and an AI Gateway with vLLM, llm-d, NVIDIA Dynamo, NIM microservices, and GPU orchestration. Organizations gain self-service model discovery, granular access policies, rate limiting, and subscription-based entitlement tracking. Access management is decoupled from execution, enabling quota management and cost allocation.
The latest Nemotron models, including Nemotron 3.5 Super VL, are also fully supported in vLLM and AI Factory. Using Red Hat’s LLM compressor, they created FP8 and DSpark speculative-decoding checkpoints for that model. According to the companies, this significantly reduces inference latency while maintaining accuracy.
Vera Rubin and Vera CPUs
On the hardware front, Red Hat promises Day 0 support for upcoming NVIDIA architectures, including Vera Rubin. Red Hat Enterprise Linux for NVIDIA now also supports Vera CPUs. Nvidia positions Vera as the go-to CPU for agents, claiming it delivers twice the agentic CPU performance. Agentic workloads require continuous reasoning loops, tool orchestration, and sandbox execution. To support this, Vera offers faster sandbox performance and greater memory bandwidth than traditional processors.
Partners such as Cisco, Dell Technologies, and Lenovo offer joint use cases, reference designs, and agentic capabilities.