Official Ubuntu support for the NVIDIA Rubin platform, including the NVIDIA Vera Rubin NVL72 rack-scale systems, announced at CES 2026
CES 2026, Las Vegas. – Canonical, the publisher of Ubuntu, is pleased to announce official support for the NVIDIA Rubin platform and the latest distributions of the new NVIDIA Nemotron 3 open models.
As AI workloads transition from small-scale proof of concepts to massive-scale AI factories, the need for a stable, high-performance operating system substrate has never been greater. Ubuntu provides that foundation, unifying the NVIDIA Vera CPU, NVIDIA Rubin GPU, and NVIDIA BlueField-4 DPU into a cohesive execution environment.
“AI democratization requires more than powerful silicon; it needs a secure, frictionless environment from cloud to edge,” said Justin Boitano, Vice President, Enterprise AI Products, NVIDIA. “Our work with Canonical ensures enterprises can deploy the NVIDIA Rubin platform, including the rack-scale Vera Rubin NVL72, on the Ubuntu foundation they already trust, streamlining the path from development to production with enterprise-grade security, and removing friction so organizations can innovate with confidence.“
“Our collaboration with NVIDIA reinforces Canonical’s mission to democratize AI,” said Cindy Goldberg, VP Cloud and Silicon Partnerships at Canonical. “By unifying the CPU, GPU, and DPU into a single, cohesive Ubuntu execution environment, we are making it easier to build private or sovereign clouds and the next generation of intelligent applications. Enterprises can build for the future on a secure foundation they trust.“
First-class Arm support for Vera Rubin
NVIDIA Vera Rubin NVL72 rack presents a scalable architecture for integrated AI compute, pairing the custom Arm-based Vera CPU with the Rubin GPU. In Ubuntu 26.04, Arm is a first-class citizen with x86 performance parity. Canonical is integrating critical upstream features including Nested Virtualization and MPAM (Memory System Resource Partitioning and Monitoring). These tools enable providers to partition memory bandwidth and cache at the hardware level, ensuring predictable performance for multi-tenant AI workloads. This infrastructure will be further reinforced with native Arm support in OpenStack Sunbeam and Apache Spark, allowing data engineers to run end-to-end pipelines on Arm-native silicon.
In addition, Ubuntu serves as the host OS for NVIDIA Mission Control software, which accelerates every aspect of infrastructure operations — from configuring deployments and integrating with facilities to managing clusters and workloads.
Streamlined inference with Nemotron inference snaps
Deploying large language models (LLMs) often involves complex dependency management and version conflicts. To solve this, Canonical has introduced inference snaps. In a recent livestream with the NVIDIA Developer community, Canonical’s team presented inference snaps for DGX Spark, showing how a single snap install command brings silicon-optimized AI models to every Ubuntu machine. To further this effort, Canonical will be collaborating with the NVIDIA team on packaging and distribution of the NVIDIA Nemotron-3 family of open models, starting with Nano models. Packaging these models using inference snaps provides a containerized, immutable environment that includes all necessary libraries and runtimes, ensuring that it is optimized for the new NVIDIA platforms.
BlueField-4 and scalable storage
Performance of the Rubin platform is further amplified by its data pipeline. Canonical is recommitting to the NVIDIA BlueField-4 DPU as an important networking and security component of NVIDIA platforms. With 64 NVIDIA Grace CPU cores and 800G/s throughput, BlueField-4 makes scaling AI Factories even more efficient. Ubuntu serves as a foundational environment for NVIDIA DOCA – allowing for seamless offloading of networking, storage, and security tasks from the Vera CPU to the DPU.
In collaboration with storage partners, Canonical is ensuring that Ubuntu’s storage subsystem is optimized for the GPUDirect Storage capabilities of BlueField-4. This enables high-speed data access between NVMe storage and Rubin GPU memory, eliminating bottlenecks. Whether deploying on-premises or in a sovereign cloud, the combination of Ubuntu and BlueField-4 provides the bare-metal isolation and hardware-root-of-trust necessary for secure, multi-tenant AI infrastructure.
