Red Hat has expanded Red Hat AI Factory with NVIDIA with a reference architecture for running artificial intelligence (AI) agents in isolation, a reference design for inference with access control, and day-one support for NVIDIA’s Vera Rubin platform. The announcement was made on October 21, 2026, during NVIDIA GTC in Berlin, and will reach customers through partners such as Cisco, Dell, and Lenovo.
Red Hat AI Factory with NVIDIA in 20 seconds
- New OpenShell Reference Architecture to isolate AI agents.
- Inference design with models as a service, AI Gateway, vLLM, llm-d, Dynamo, and NIM.
- Nemotron 3.5 Super VL arrives on vLLM in FP8.
- Day-one support for Vera Rubin and the Vera CPU.
- Sold through Cisco, Dell, and Lenovo.
The package addresses a recurring concern among infrastructure teams: an agent that executes code, calls tools, or accesses internal data needs clear boundaries, and those boundaries have to work the same way in an on-premises data center as in a public cloud. Red Hat builds its proposal on OpenShift as a common layer for both environments.
Isolated agents and controlled access to models
The OpenShell Reference Architecture describes how to deploy AI agents in isolated environments within the platform, so that each agent operates with scoped permissions. It’s a reference architecture — a validated, documented design, not a new product with its own license.
The reference design for inference brings together several pieces: a models-as-a-service (MaaS) access model with control over who consumes what, an AI Gateway to manage traffic to the models, the vLLM and llm-d inference engines, and NVIDIA’s Dynamo and NIM tools. The idea is for a platform team to be able to offer models internally with quotas and traceability, without every developer standing up their own inference server.
Red Hat has also added NVIDIA’s Nemotron 3.5 Super VL model to vLLM in FP8 precision, and offers DSpark checkpoints generated with LLM Compressor, its tool for compressing models and reducing GPU memory usage.
Vera Rubin from day one
On the hardware side, Red Hat announces day-one support for the Vera Rubin platform, and Red Hat Enterprise Linux (RHEL) for NVIDIA already supports the Vera CPU. NVIDIA attributes to these CPUs double the performance on agent workloads, a figure from the manufacturer itself for which the announcement doesn’t detail methodology or comparison system.
The announcement features comments from Joe Fernandes, of Red Hat, and John Fanelli, of NVIDIA. The announcement doesn’t include pricing, and availability depends on the partners integrating the solution: Cisco, Dell, and Lenovo. No specific dates are given for each one.
The move follows Red Hat AI 3.5, focused on security, observability, and GPU control, and joins the AI factories other manufacturers have introduced with NVIDIA, such as Dell’s effort to take agentic AI from the desktop to the data center.
Frequently asked questions
What is Red Hat AI Factory with NVIDIA?
It’s the joint offering from Red Hat and NVIDIA for deploying AI infrastructure on OpenShift and RHEL with NVIDIA hardware and software.
What does the OpenShell Reference Architecture provide?
A validated design for running AI agents in isolated environments with scoped permissions.
How can it be acquired?
Through partners such as Cisco, Dell, and Lenovo. Red Hat hasn’t published pricing.
Sources:
This article is also available in Spanish: “Red Hat y NVIDIA refuerzan la seguridad de los agentes de IA en nube híbrida,” on revistacloud.com.
