// a node OS for Kubernetes AI infrastructure
An OS Forged for Kubernetes AI Nodes#
k8or-os is a lean, immutable node operating system built from the Linux kernel up, not a general-purpose distribution stripped down. One signed base image runs every node; signed layers add the RDMA fabric, the GPU driver stack and an AI-Ops triage engine only where they belong.
mars@k8or-node-07:~$ k8or status --layers LAYER IMAGE STATE VERIFIED kernel k8or-6.12.x · UKI booted secure boot ✓ tpm measured ✓ base k8or-base-1.0.erofs mounted dm-verity ✓ fabric k8or-fabric-1.0.erofs mounted dm-verity ✓ gpu k8or-gpu-1.0-580.erofs mounted dm-verity ✓ kernel match ✓ aiops k8or-aiops-1.0.erofs running underlay · advisory kubelet v1.3x Ready started after GPU stack mars@k8or-node-07:~$ █
illustrative = the target design. k8or-os is in the architecture phase; this is what a GPU node will report once the layers exist.
Project status
Architecture: defined. Phase 0 (toolchain and kernel): next. GPU validation on H100 / RoCE v2: planned for phase 6. Each page's badge shows its own status.
Why another operating system#
A GPU node needs a few dozen components. A general-purpose distribution ships thousands. Converting Ubuntu, RHEL or Rocky into a Kubernetes node means removing, patching and auditing software that kubelet never touches. These problems are the everyday cost of running AI infrastructure on general-purpose operating systems:
Packages installed and upgraded node by node diverge over months. "Works on node 14, fails on node 15" becomes a debugging category of its own.
DKMS rebuilds GPU modules on the node after a kernel update. When that fails, a GPU server reboots without its GPUs.
Driver containers load the GPU stack after the node is up, so a node can report Ready before its GPUs are usable.
Everything a node does not need is still something to patch, audit and defend. A lean OS removes it instead of hardening it.
Six principles#
A Linux kernel and only the userland needed to run containerd and kubelet.
Nodes are never modified in place; they switch to a new, complete image.
Unified kernel image under Secure Boot, dm-verity on every image, TPM-measured boot.
Nodes are configured through a declarative k8or machine-config API only.
Capabilities arrive as signed, read-only layer images built in the same pipeline as the base.
Boot, mounts, drivers and kubelet behave identically every time. AI sits beside the core, never in its path.