NVIDIA has added an inventory of declared AI workload details to its GPU cluster software, giving operators a way to review configured components without proving that a job completed successfully.
The update connects NVIDIA’s AI Cluster Runtime (AICR), which packages tested platform configurations, with k8s-aibom, a Google-developed tool that produces CycloneDX 1.6 AI bills of materials (AIBOMs). NVIDIA described the pairing in a Nov. 11 blog post. AIBOM entries draw on workload specifications and available pod-status information. They may be generated before a pod starts and do not establish that execution succeeded.
The h100-gke-cos-inference and h100-gke-cos-inference-dynamo recipes enable the inventory by default for H100 inference on Google Kubernetes Engine. Other GKE recipes can activate it with the --runtime-inventory option.
Administrators select namespaces for AIBOM creation, while the controller can read workload specifications across the cluster. NVIDIA’s component catalog lists supported objects including Dynamo graphs, NIMService, LeaderWorkerSet, DisaggregatedSet and CronJob.
NVIDIA says teams can use the inventory to identify AI services absent from internal records and monitor dependency changes, giving security teams information for reviews and incident response. The k8s-aibom project documentation describes the records as evidence for compliance work; the inventory does not certify compliance.
NVIDIA cited upstream measurements of k8s-aibom version 1.6.1 on GKE with Dynamo. Across a 15-minute steady-state test, the controller contacted the Kubernetes API server about four times per minute, used 2.2 millicores of CPU on average and had 28 to 34 MiB in working-set memory. The test used CPU-only nodes while GPU workload pods were pending, so the figures measure controller overhead rather than inference performance.
NVIDIA has also published B200 inference configurations for four Korean open models. The two tools do not yet compare an AICR recipe’s planned setup with k8s-aibom’s collected entries; NVIDIA described that connection as a later step.
