FAQ#
Questions experts ask about the design.
Why not start from Ubuntu and strip it down?
The result is still a general-purpose distribution with parts removed: its package manager, update model and assumptions remain. k8or-os starts from the kernel and adds only what a Kubernetes node needs, so there is nothing to remove, patch or audit that was never required.
Why not simply fork Talos?
A fork gets a working node fastest, but it is Talos renamed and tied to Talos's choices, including a musl base that needs a glibc extension for NVIDIA. k8or-os uses Talos as a reference design and builds its own base with GPU and fabric designed in from the start.
Why is a capability a layer and not a package?
A package is how software is built; a layer is how it is delivered. Inside the k8or pipeline, everything is built as packages from source. What ships is a signed layer, because the node has no package manager. Layers keep the OS immutable, are verified by dm-verity on every read, declare the exact kernel and base they need, and are removed by deleting one file.
Why is the AI-Ops engine not a DaemonSet?
A DaemonSet dies with the node's Kubernetes stack, exactly when triage is needed most. Underlay is started by k8or-init on the host, so it still reads the kernel log and DCGM data when kubelet is down, containerd is wedged or the driver is hung. Fleet-wide reasoning and dashboards still run as ordinary cluster workloads.
Is a Helm chart different from a DaemonSet here?
No. A Helm chart is a packaging format for Kubernetes objects; once installed, it creates DaemonSets and Deployments that run under kubelet and share their fate. The real distinction is host versus Kubernetes, not chart versus DaemonSet.
Does k8or-os run inside containers?
No. k8or-os runs on the node. Containers bring their own userland and share the node's kernel. A minimal k8or base container image is planned as a separate deliverable from the same pipeline.
Does the GPU layer include CUDA?
No, only the driver stack: kernel modules, libcuda, NVML, nvidia-smi, Fabric Manager, the container toolkit and DCGM. The CUDA toolkit, cuDNN and NCCL ship in each workload's image, and the container toolkit injects the host's driver libraries when the pod starts.
Why two flavors instead of two operating systems?
Two independent operating systems mean every patch is done twice and the two drift apart. k8or-os builds one base and adds the GPU layer only on GPU nodes, from the same pipeline and kernel, with one version number.
Does any operating system do AI layering today?
Not in this form, as of November 2026. Talos and Flatcar support layers but ship no AI layer. RHEL 10 includes the Lightspeed command-line assistant, an installed package that helps an administrator at the terminal through a hosted model, without node-level, GPU- or fabric-aware triage.
What happens when a node kernel-panics?
Nothing inside any OS can help at that point. The cluster-level side of AI-Ops polls the server's BMC out of band over Redfish or IPMI and correlates that with the last telemetry the node sent.