Red Hat Customizes Rhel For Nvidia’s Vera Rubin Ai Platform

Building connected continual business activity pinch nan spot company, Red Hat has pledged to connection “Day 0” support for Nvidia’s newly-announced Vera Rubin Platform, erstwhile it arrives later this year.

The endeavor unfastened root package institution is aiming to adjacent nan merchandise spread betwixt nan motorboat of Nvidia’s hardware level and nan wide readiness merchandise of its ain supporting stack of software, a civilization type of its flagship operating system, Red Hat Enterprise Linux (RHEL), designed for Rubin.

Nvidia’s Vera Rubin level is built to trim nan costs of AI inference. It is built connected a caller CPU, GPU, and Digital Processing Unit (DPU), each of which are intimately tied to different components.

Hardware and Software connected nan Same Release Schedule

Both nan level and Red Hat’s package are expected successful nan second half of 2026, according to a Red Hat spokesperson.

When Rubin comes to town, Red Hat repositories will person ready-validated Nvidia GPU OpenRM drivers and nan CUDA toolkit.

Red Hat’s package will besides beryllium capable to tally AI lifecycles wrong nan letter cover of Nvidia’s recently unveiled Confidential Computing information architecture, which provides cryptographic impervious of workload protection.

In summation to RHEL, Red Hat will besides merchandise nan Red Hat OpenShift level for Kubernetes for Rubin, and nan Red Hat AI platform will grow support for distributed conclusion pinch Nvidia’s unfastened root models connected Red Hat AI Inference Server, Red Hat Enterprise Linux AI and Red Hat OpenShift AI.

Introducing Vera Rubin

Nvidia CEO Jensen Huang introduced Vera Rubin during successful his keynote for nan 2026 Consumer Electronics Show (CES). It is named aft American astronomer Vera Florence Cooper Rubin, who identified acheronian matter arsenic a point that takes up astir of our universe.

This level tackles different difficult mission: Reducing nan skyrocketing machine (and energy use) costs of moving large-scale AI conclusion workloads.

By engineering nan components to activity much intimately together, Vera Rubin tin trim nan conclusion token costs to train models by a facet of 10, while utilizing 4 times less GPUs, compared pinch utilizing nan Nvidia Blackwell platform, traditionally utilized for specified tasks, nan institution asserted.

The platform’s Ethernet photonics move systems committedness to present five-fold improved powerfulness ratio arsenic well.

Vera is nan sanction of the 88-core CPU for nan platform, and Rubin is nan sanction of nan GPU, which offers 50 PFLOPs of mathematical crunch (5x offered by Blackwell). Further computational acceleration is besides locked successful pinch nan Nvidia BlueField-4 DPU. These processing units are each intimately intertwined pinch nan NVLink 6 Switch, Nvidia ConnectX-9 SuperNIC, and Nvidia Spectrum-6 Ethernet Switch — each made pinch Nvidia chips.

The Cost of Inferencing

As companies ramp up their AI efforts, they are uncovering that inferencing work, which is done astir efficiently done GPUs, is getting much expensive.

This week, Amazon Web Services (AWS) hiked nan costs of its GPUs by 15%.

The costs of inferencing has so been connected nan minds of Nvidia executives. The institution besides inked a $20 cardinal licensing deal pinch Groq for its inferencing and compiler technologies.

In his keynote, Jensen talked astir nan melodramatic summation of inferencing that is taking spot wrong AI systems moreover recently. Not only are nan models themselves getting larger, but nan magnitude of conclusion itself has skyrocketed. No longer is inferencing utilized to nutrient a azygous “one-shot” answer, but is simply a “thinking process” often involving aggregate models and further post-training.

So much inferencing is needed.

“You now person reinforcement learning, [which is] fundamentally nan machine trying different iterations learning really to execute a task,” Jensen said. “The magnitude of computation for pre-training. for post-training. for trial clip scaling has exploded arsenic a consequence of that.”

“The ratio gains successful nan Nvidia Rubin level correspond nan benignant of infrastructure advancement that enables longer memory, amended reasoning and much reliable outputs,” noted Dario Amodei, cofounder and CEO of Anthropic, successful a statement.

In summation to Red Hat (which is owned by IBM), different companies that person pledged their support to nan Rubin level include:

AWS, Anthropic, Cisco, CoreWeave, Dell, Google, HPE, Lambda, Lenovo, Meta, Microsoft, Mistral AI, OpenAI, nan Oracle Cloud Infrastructure (OCI), among galore others.

Group Created pinch Sketch.