Making Kubernetes AI-Ready for NVIDIA Vera Rubin on Day One

New GPU hardware usually means weeks of platform teams fighting mismatched drivers, kernels and operators. NVIDIA's AI Cluster Runtime, or AICR, packages a tested configuration into a reusable recipe, renders deployment bundles and validates the running cluster.

For Vera Rubin NVL72, the company says its training recipe passed deployment, conformance and a cross-node NCCL all-reduce performance check at 744.71 GB/s bus bandwidth over NVLink. It says a second cluster reproduced the setup the next day.

Teams moving from GB200 NVL72 get a map of what changes: a new driver branch, ConnectX-9 SuperNICs and BlueField-4 DPUs, a different device topology, and new MIG profiles.

We like the honesty in this one. NVIDIA labels the Vera Rubin recipes Preview, says plainly that inference serving performance isn't validated yet, and explains why: the vLLM images available don't yet include the compiled Rubin kernel path. That's more candor than most launch-adjacent posts offer.

The recipe idea is the real story. Encoding fussy details once, like the memory limit for one component that gets killed on multi-terabyte nodes, saves every platform team from rediscovering them. The detail that a validation run surfaced a crash in a low-level library, and the fix moved through the stack in about three days, is a good sign for the feedback loop.

The caution is lock-in. AICR stitches together a lot of NVIDIA components, from scheduling to health monitoring. The more convenient the blessed path gets, the harder it becomes to run anything else. That's fine for teams all-in on NVIDIA, but it narrows the field.

For most readers, the takeaway is simple: Rubin training clusters look close to plug-in ready, and anyone planning inference on Rubin should wait for proper runtime images.

Vera Rubin NVL72 is NVIDIA's third generation of its MGX rack-scale design, following GB200 NVL72. AICR is part of NVIDIA's DSX AI Factory Platform, and the recipes are published in the open-source NVIDIA/aicr repository at v1.0.0. NVIDIA says the first deployment hit a platform combination none of the components had been validated on together, including arm64, Ubuntu 26.04 with a 64k-page kernel, and RKE2, and still reached a healthy runtime within hours of cluster handoff.