Canonical has published a technical guide on running agentic AI workloads using Arm processors alongside Ubuntu. The reasoning is straightforward: AI agents chain together model calls, tools and APIs, which means sustained, power-efficient compute. That profile maps well onto recent Arm silicon.
The hardware in question
The article uses as its reference an Arm CPU with up to 136 Neoverse V3 cores laid out in a dual-chiplet design built on TSMC’s 3nm process. The figures Canonical gives are specific: a 300W TDP for maximum throughput, memory latency below 100ns, 12 DDR5 channels and 6GB/s of bandwidth per core. At the rack level, they cite 272 dedicated cores per 1OU dual-node server unit, which says a lot about the density they are targeting for data centres.
These aren’t hand-wavy claims. They are the numbers that explain why Arm keeps coming up in AI conversations, where power draw and density per rack matter as much as raw performance.
What Ubuntu brings to the table
The software side is where Ubuntu earns its place. Ubuntu 26.04 LTS ships with the Linux 7.0 kernel and enterprise support of up to 15 years through Ubuntu Pro. For anyone running fleets of Arm servers, two details stand out:
- Kernel livepatching on Arm64: applying kernel security patches without rebooting. This existed on other architectures already, and it now covers Arm64 servers too, which matters when you have agents running continuously and can’t afford downtime windows.
- Arm64 features at the system level: support for MPAM (control over shared resources such as cache and memory bandwidth) and nested virtualization.
The post also points to Rust entering system components, with sudo-rs and coreutils rewritten in the language for memory safety, plus alignment with MTE (Memory Tagging Extension), present in Neoverse V3. This is the direction Ubuntu has been moving in for a while: cutting whole classes of memory bugs out from the base of the system.
The rest of the stack
Canonical takes the chance to remind readers that its infrastructure tooling has native Arm64 support: MAAS for bare-metal provisioning, OpenStack, MicroCloud, Ceph for storage, LXD for containers and virtual machines, and Canonical Kubernetes. The message is that you can build the full stack, from metal to orchestrator, on Arm without stepping outside official support.
Cindy Goldberg, VP of Cloud and Silicon Partnerships at Canonical, frames the announcement within a long-running collaboration with Arm. Eddie Ramirez, VP of Go-to-Market at Arm, sums up the motivation: agentic AI calls for a new class of infrastructure where performance, efficiency and scalability are engineered together.
Why it matters to you
If you run AI agents in production, or you are weighing up where to host them, the practical takeaway is clear: Arm is no longer a niche option for this kind of workload. Pairing many efficient cores with Ubuntu’s long support and hot kernel patching cuts two costs that hurt in operations: the electricity bill and maintenance downtime.
Canonical also notes that Ubuntu was the first enterprise distribution to support Arm servers, back with Ubuntu 12.04 LTS in 2012. Fourteen years on, that track record shows in how polished the architecture support has become.
Source
Based on the article by Youssef Eltoukhy published by Canonical on 26 May 2026: Run agentic workloads on Arm and Ubuntu.