dsx-ai-factory/nvke

What it is

NVKE is a management plane for operators who want to offer managed Kubernetes on their own NVIDIA GPU hardware. A platform team installs the management cluster; customers then get isolated workload clusters, each with its own Kubernetes control plane, etcd and node pools, accessible through the standard Kubernetes API and kubectl README lines 1–18.

It is explicitly not a Kubernetes distribution or a Cluster API flavour. A throwaway seed cluster stands up a management cluster, and the management plane on it provisions and lifecycle-manages workload clusters on demand README lines 65–87. Cluster creation, node-pool scaling, Kubernetes upgrades and node replacement are driven through the NVKE API or nvkectl, with machines provisioned via NVIDIA Infra Controller (NICo) README lines 40–64. Nodes attest through SPIFFE/SPIRE with their TPM before receiving join credentials; per-cluster etcd encryption is available.

Quick start

To evaluate without GPU hardware, follow the local quickstart. It runs on a Linux host or VM (amd64 or arm64) with at least 6 CPUs, 12 GiB free memory and 100 GiB free disk, and requires Docker and virtual TPM support. The flow ends with workload-cluster creation, and you confirm the cluster's nodes report Ready. On macOS, run the evaluation in a Colima VM using the local site guide.

To build the binaries from source, use Go 1.27.3 (pinned in go.work) and GNU Make 4 or later:

git clone https://github.com/dsx-ai-factory/nvke.git
cd nvke
make build

This writes nine binaries to bin/: agent, updater, controller, nvkectl, capd-reboot-controller, nvke-os-image-check, management-engine, tenant-api and operator-api. Building does not start a management plane or create a cluster README lines 88–110. For a real installation, start with the installation plan, which covers node preparation and management-cluster bootstrap.

Constraints

NVKE is a preview: APIs, configuration and behavior can change, and it is not recommended for production workloads Preview status and known limitations. API authentication is off and the API serves plain HTTP after installation, so operators must configure protected access before admitting callers API authentication guide. The management cluster does not upgrade, scale or replace itself, and there is no management-plane upgrade or backup/restore procedure. FDE node images are for evaluation only and provide no data-at-rest confidentiality.

NICo is the only compatible infrastructure provider at preview, so a real installation requires your own machines, networking, node images and registries plus configured NICo access; validated combinations and pinned dependencies are listed in the platform and version reference. The node agent runs on Linux. The companion project NVIDIA Kubernetes Recipes builds and signs the node images and components that NVKE deploys. Source is under Apache-2.0 LICENSE.

HotRepo's take

The project entered the Go daily list on 2026-11-05 at rank 11 with 3 new stars, while total stars were 73 and forks 65. That fork-to-star ratio is unusually high for a project this young, and the snapshot is a single day's record with no long-running star history, so it supports little more than the fact of a brief appearance on a language leaderboard. The scope itself is narrow and concrete: a self-hosted control plane for GPU Kubernetes on bare metal, with a candid preview-limitations document. Readers evaluating managed-GPU-cluster platforms should treat it as an early operator-level experiment, not a turnkey product.