Offering Kubernetes to GPU customers is not just a matter of installing a cluster: the operator must also provision machines, isolate customer control planes, and handle upgrades and node failures across a fleet. NVKE focuses on that management layer, while leaving the customer workload and the underlying bare-metal infrastructure in the operator’s hands.
What Sets It Apart
- Separate management from customer clusters: A management plane provisions workload clusters, each with its own control plane and etcd. This gives operators a repeatable service model rather than a single shared cluster.
- Builds on Kubernetes conventions: It uses upstream Kubernetes, kubeadm, and Cluster API, with NICo as the current infrastructure provider. Existing Kubernetes workloads and operational skills remain relevant.
- Treats node identity and artifact provenance as part of operations: TPM-backed node attestation and companion build tooling address trust from machine admission through deployed software, important when customers share a large GPU fleet.
Who It’s For
Great fit for AI factory operators and platform teams building managed Kubernetes for GPU customers, or for providers evaluating reusable pieces of its cluster-management stack. Look elsewhere if you need a production-ready platform today: NVKE is explicitly in preview, NICo is its only supported infrastructure provider, API authentication is off by default, and the management cluster has notable lifecycle and recovery gaps.
