NVIDIA has released four Kubernetes configuration recipes utilizing the AI Cluster Runtime (AICR) for Vera Rubin NVL72 systems targeting physical bare-metal RKE2 clusters, tokenpost.kr reported on November 3, 2026. The preview deployment covers general training, Kubeflow Trainer training, general inference, and Dynamo inference pathways.
AI Cluster Runtime Deployment and Verification Status
- Only the Kubeflow Trainer training pathway currently matches its deployment and verified evidence under strict three-step testing.
- General training, general inference, and Dynamo inference pathways rely on evidentiary data drawn from prior legacy recipes.
- The preview configuration requires Ubuntu 26.04 alongside a Vera-tailored 64K page kernel, resulting in physical host reboots during setup.
Bare-Metal Constraints and Kernel Requirements
The newly published recipes apply strictly to physical bare-metal RKE2 environments rather than managed cloud provider Kubernetes setups. Operating the infrastructure demands active intervention from human operators handling baseboard management controllers and node recoveries. Furthermore, host-level services like nvidia-imex must be explicitly disabled and blocked from restarting during deployment. Automated hardware discovery faces similar hurdles because AICR fails to automatically recognize the VR200 hardware. Operators must manually pass the --accelerator vr200 flag during recipe generation to prevent the system from recording the accelerator type as generic.
Version Divergence Across Training and Inference Pathways
Prerequisites shift significantly depending on whether clusters handle training workloads or production inference models. Training configurations demand Kubernetes version 1.34.1 or higher to function properly. Conversely, inference paths enforce a strict ceiling, requiring Kubernetes versions prior to 1.36.0 to avoid Gateway API configuration conflicts. These inference deployments also necessitate external load balancers running directly on the bare-metal hardware. NVIDIA maintains that this preview phase does not include comprehensive operational support or full lifecycle validation, separating it clearly from a production-ready software release.
| Configuration Pathway | Required Kubernetes Version | Evidentiary Status |
|---|---|---|
| Kubeflow Trainer Training | v1.34.1 or higher | Verified via current 3-step testing |
| General Training | v1.34.1 or higher | Relies on prior recipe evidence |
| General Inference | Under v1.36.0 | Relies on prior recipe evidence |
| Dynamo Inference | Under v1.36.0 | Relies on prior recipe evidence |
AICR Operational Mechanics and Storage Snapshots
Rather than provisioning new infrastructure or installing operating systems from scratch, AICR captures snapshots of existing cluster states, including the kernel and GPU statuses. The runtime uses these snapshots to generate version-locked configuration bundles that operators apply using tools like Helm, Argo CD, Flux, or Helmfile. NVIDIA originally introduced AICR v1.0 on October 6, establishing clear compatibility rules across command-line interfaces, REST APIs, and Go software development kits.
