nvke: Building a Kubernetes Factory on Bare-Metal GPU Infrastructure

This is a control plane for building a self-managed Kubernetes service on bare-metal infrastructure equipped with NVIDIA GPUs. Once the platform team installs NVKE, they can create and operate workload clusters for customers. Each workload cluster has an independent Kubernetes control plane, etcd, and node pools, and customers access them via the standard Kubernetes API and kubectl.

Instead of an existing Kubernetes distribution or a Cluster API variant, it adopts a 'factory' structure that produces clusters. It sets up a management cluster through a temporary seed cluster, and this control plane provisions actual machines via the NVIDIA Infra Controller (NICo) to create multiple isolated workload clusters. It manages the entire lifecycle, including node pool creation and scaling, Kubernetes upgrades, and node replacement.

From a security perspective, nodes obtain join credentials through SPIFFE/SPIRE authentication using TPM. etcd data encryption can be configured per cluster. Currently in the preview stage, API authentication is disabled and plaintext HTTP is used, so explicit authentication configuration and security settings are essential before deploying to production environments.

It is suitable for platform engineering teams that directly own GPU infrastructure and want to offer Kubernetes as a service in a multi-tenant environment. It can also be evaluated in a simulated environment without GPU hardware, but compatibility with NICo must be verified for actual installations. It is released under the Apache-2.0 license and written in Go.