Nvidia has launched NVIDIA AI Cluster Runtime (AICR) version 1.0, a tool designed to simplify the deployment of the Dynamo inference framework on Kubernetes clusters. The runtime uses version-locked recipes to ensure that all GPU drivers and software components are compatible, reducing configuration errors during deployment.
The AICR workflow consists of five stages: snapshotting the cluster state, generating recipes, rendering bundles, deploying, and validating the configuration. Nvidia provides cryptographically signed validation evidence to confirm that each specific configuration has been tested and verified for performance.
Documentation released in late October 2026 covers deployments on Amazon Elastic Kubernetes Service (EKS) using H100 and GB200-class GPUs. For EKS setups with H100 GPUs, the company utilizes version 1.4.2 of the Dynamo Platform Helm chart.
The framework requires Kubernetes version 1.34 or later for teams intending to utilize Dynamic Resource Allocation (DRA), a feature that enables flexible distribution of GPU hardware. The system is designed for distributed inference, supporting multi-node orchestration and KV-cache-aware routing.
