Microsoft has expanded AKS on bare metal to support Ubuntu as a host operating system, and the new option is in public preview now. The original preview paired the service with Azure Linux 3.0 running on validated Azure Local small form factor hardware, which was a sensible but narrow starting point. If your racks were full of Ubuntu machines and no Azure Local hardware in sight, you were out of luck. That restriction is gone.
Modest requirements, customer hardware
The Ubuntu path runs on hardware you provide yourself. According to the system requirements on Microsoft Learn, a host needs Ubuntu 24.04.3 LTS or 24.04.4 LTS, an x86_64 processor, at least two physical cores and 4 GB of RAM (four cores and 8 GB recommended), and 256 GB of free disk space. One single-node cluster per host during the preview. Ubuntu 24.04.5 LTS is not supported yet; Microsoft says support is planned for a future update.
The host also has to be connected to Azure Arc, and during the preview everything lands in East US. That is a real constraint for anyone running on-prem infrastructure outside that region, though it is the same East US-only limitation the Azure Linux flavor carries.
You own the OS, Microsoft owns Kubernetes
The responsibility split is the part worth reading twice. AKS does not reimage the host and does not manage its operating system lifecycle. You install Ubuntu, you patch the kernel and system packages, you maintain drivers, DNS, proxy, firewall, and capacity. AKS installs and manages the host-level packages, agents, and networking components required for the Kubernetes layer itself.
This is a different deal from cloud AKS, where Microsoft owns everything below your pods, and it is a different deal from the Azure Linux flavor, where the Azure Local platform handles host lifecycle. On Ubuntu hosts, the operational burden of OS patching sits squarely with you. If your team already runs Ubuntu fleet management at scale, that is familiar work. If you were hoping for a fully managed edge Kubernetes experience, this is not it.
What the preview actually supports
The preview limitations page is honest about scope, which I appreciate. Ubuntu hosts get single-node clusters only, one cluster per host, no scaling. Cluster creation is Azure CLI only, and the portal shows read-only cluster status. Kubernetes 1.35.5 is the default version, with patch-version upgrades handled through the CLI. Container networking runs on Cilium CNI, same as the Azure Linux path.
Billing is zero-rated during the preview: no charges for the AKS cluster resources themselves. Standard Azure charges still apply for the Arc-enabled machine and anything else you attach, like Monitor or Policy.
Why Microsoft is pushing this
The framing around the announcement is AI workloads on-premises. Running Kubernetes directly on physical hardware means no hypervisor tax, local access to your data, and your choice of GPUs, drivers, and runtimes. For teams doing inference or training at the edge with data that cannot leave the building, that combination is the whole pitch.
There is a competitive angle too. Teams that wanted Kubernetes on their own metal already had kubeadm, k3s, or OpenShift. AKS on bare metal offers something those do not: the same AKS management plane as the cloud, with Azure Arc gluing it together. Dropping the Azure Linux and validated-hardware requirements removes the biggest friction the original preview had. Ubuntu is what most shops run, and Microsoft knows it.
How the two flavors compare
Microsoft publishes a side-by-side comparison, and the differences are mostly about management surface rather than capability. The Azure Linux flavor runs on validated Azure Local hardware, supports cluster creation through the portal, ARM templates, or Bicep, and gets full portal experiences for status and patch upgrades. The Ubuntu flavor is CLI-only for creation and patching, with a read-only portal. Both run single-node clusters with Cilium, both are East US only, and both get zero-rated billing during the preview.
The capability you give up with Ubuntu is convenience. The capability you gain is hardware freedom. If your servers already run Ubuntu 24.04 and meet the requirements, you can have a cluster running today on hardware you were going to depreciate otherwise. That trade is obvious for lab and evaluation work, and Microsoft is betting that enough teams will accept CLI-driven management in exchange for it that the gap closes by GA.
Keep it out of production for now
Microsoft is explicit that preview clusters may not support in-place upgrade to the generally available version. You might have to delete and recreate your cluster when GA arrives, so plan for backup and redeploy. Single-node clusters with no scaling and best-effort support make this an evaluation tool, not a production target.
If you want to poke at it, the path is: check your host against the requirements, connect it to Arc in East US, register the required resource providers, then create the cluster with the Azure CLI. It is the cheapest trial run you will get, since the cluster resources bill at zero during the preview.