NVIDIA Adds Runtime AI Inventory to Its GPU Platform Tooling

AICR and k8s-aibom cover different layers

NVIDIA AI Cluster Runtime, or AICR, is an open source modular component of the NVIDIA DSX AI Factory Platform. It creates version-locked recipes and reproducible deployment bundles for GPU-accelerated Kubernetes. The recipes describe a combination of hardware, cloud, operating system and intent, including component versions, configuration values, constraints, deployment order and validation expectations.

The newly qualified AIBOM component is based on k8s-aibom, an open source project created by Google. It watches Kubernetes workload and pod APIs and produces CycloneDX 1.6 AIBOM records for recognized AI workloads. AICR validates the infrastructure configuration, while k8s-aibom records what is actually running on that infrastructure.

NVIDIA says the distinction addresses a gap between platform configuration and application workloads. Teams can update workloads independently of a platform recipe, while mutable image tags such as :latest can result in different code running under identical tags.

Supported recipes and workload inventory

Two stock recipes select k8s-aibom by default: h100-gke-cos-inference and h100-gke-cos-inference-dynamo. Other Google Kubernetes Engine recipes can enable the component during generation with the --runtime-inventory flag, or disable it with --runtime-inventory disabled.

Coverage is limited to supported workload kinds carrying a recognized AI signal. It includes native Dynamo graphs, NIMService, LeaderWorkerSet, DisaggregatedSet and CronJob resources. Tracked child workloads are combined into their owner’s AIBOM.

According to NVIDIA, the inventory can help teams find unregistered or shadow AI workloads in selected namespaces, identify models and inference runtimes, track external AI APIs and vector databases, detect changes in AI dependencies, and support security reviews, incident response and compliance audits.

Validated deployment on GKE

The examples in the announcement use AICR v1.2.0, which pins k8s-aibom 1.6.2. NVIDIA describes a workflow that generates a recipe, renders it into a deployment bundle, installs it and validates the resulting cluster. The bundle contains component directories ordered by the dependency graph, along with values, checksums and installation scripts. AICR supports Helmfile, Flux and Argo CD through its deployer option.

The documented example starts with a Google Kubernetes Engine cluster running Kubernetes 1.32 or later, NVIDIA H100 GPU nodes, Container-Optimized OS, and GKE-managed drivers and device plugins. Qualification binds the upstream source, chart, CRDs and digest-pinned controller image as one tested set. NVIDIA says it also verifies provenance against Google’s expected build identity, reviews rendered RBAC and security defaults, and tests installation, upgrade, health and uninstall on the deployment paths it evaluated.

Runtime inventory is opt-in at the namespace level. A namespace label controls which namespaces emit AIBOM records, while the controller continues to cache workload and pod information through its informers.

What the integration means for AI platform teams

The pairing gives platform teams two connected forms of evidence: AICR records qualified deployment intent and evaluates it against a cluster, while k8s-aibom observes supported workloads and records the assets it finds, including confidence information for attributes. NVIDIA says neither view is sufficient for AI governance on its own.

For users running AI workloads on Kubernetes, the integration is intended to make the relationship between a validated GPU foundation and the applications operating on it more visible. Teams can use AICR to reproduce and check the platform configuration, then use the runtime inventory to track supported AI workloads within opted-in namespaces.