1min.

1MIN · AI & AGENTS

GPU workload inventory separates declared assets from execution proof

NVIDIA’s AI Cluster Runtime now pairs reproducible GPU-platform recipes with k8s-aibom, which inventories supported AI workloads as Kubernetes resources. Records identify models, runtimes and images, with confidence levels and evidence locations. A crucial boundary remains: an inventory record can be ready while its workload’s pod is still pending, so inventory does not prove that inference actually ran.

NVIDIA Technical Blog · Nov 11, 2026

Why it matters

Deployment intent, observed configuration and successful execution are different kinds of evidence.

More context

The documented example uses AICR 1.2.0, k8s-aibom 1.6.2 and GKE 1.32+ with H100 nodes. Try an opted-in namespace and inspect each attribute’s evidence. Namespace reporting is opt-in, but the controller still reads workload and pod specifications cluster-wide.

Keep in mind

Available open-source integration with bounded workload recognition. Default observations are unsigned, and reported resource overhead came from CPU-only nodes with pending workload pods, not a loaded inference cluster.

Original read

More engineering ideas, one minute at a time.

Explore more stories