NVIDIA Pinned 16 Components to a Recipe, in a Bite

NVIDIA now ships the stack underneath an inference model as a versioned recipe. The number that matters is 41.

• AI Cluster Runtime snapshots a cluster, generates a recipe, renders Helm values and manifests, deploys them, then validates the result before the application reaches the node pool.
• A fully specialized recipe carries up to 268 configuration values across 16 components, including the GPU Operator, the DRA driver, KAI Scheduler, Grove and NVSentinel.
• Switching intent from training to inference swaps 5 components and changes 41 values from the same base.
• Dynamo 1.4 drops NATS as the default, because frontends and workers now discover each other through Kubernetes resources.
• Two routing paths ship, the Dynamo frontend router and the Gateway API Inference Extension with Dynamo’s endpoint picker, and both need live KV-cache events to route well.

A green validation proves the runtime matches a tested version. It says nothing about your tokens per second. Read the full analysis.