On October 6, 2026, NVIDIA released version 1.0 of NVIDIA AI Cluster Runtime, or AICR, to address compatibility across GPU-accelerated Kubernetes clusters. An AI cluster depends on host kernels, GPU drivers, container runtimes, networking, storage, operators, schedulers, and workload frameworks. An upgrade to any one of those components can break a combination that previously worked.
AICR uses version-locked, validated recipes to record tested combinations and can render deployment bundles for Helm, Argo CD, Flux, or Helmfile. Each recipe carries signed validation evidence from the hardware where it was tested. Operators can therefore inspect not only what should be installed, but the tested hardware and platform, the checks that passed, and who signed the evidence.
Version 1.0 establishes a compatibility contract for the CLI, REST API, Go SDK, bundle layout, and artifact schemas. Its four core capabilities are deliberately distinct: Snapshot records observed cluster state; Recipe describes the version-locked target and validation phases; Bundle renders artifacts for a deployment tool; and Validation compares the recipe with the running state and can perform deployment, conformance, and performance checks before recording signed evidence.
The separation matters. A Snapshot is not a desired configuration, and a Recipe does not reconcile a cluster by itself. Helm, Argo CD, Flux, or Helmfile applies and reconciles the bundle. AICR then checks whether the running cluster matches the recipe. An operator might choose criteria such as EKS, GB300, Ubuntu, training, and Kubeflow, render the same recipe for a preferred GitOps tool, and validate the result against that recipe.
NVIDIA says AICR covers 11 Kubernetes services, 10 GPU accelerators, Ubuntu, COS, Oracle Linux, and Talos through a mixin, with integrations across Helm, Helmfile, Argo CD, Flux, Kubeflow, Slurm, Dynamo, and NIM. Pulumi Labs and Mirantis k0rdent demonstrate consuming the same configuration through different infrastructure-as-code and multi-cluster tools. The project has more than 100 contributors, with roughly half outside NVIDIA.
The practical value is not a guarantee that any GPU workload will succeed. It is turning compatibility knowledge, deployment artifacts, and validation evidence into a reusable public contract. Operators still need to check that their GPU model, network, operating system, Kubernetes version, workload, and performance thresholds are covered, then revalidate upgrades and production clusters instead of treating a successful installation as proof of readiness.
