Nvidia's AI Cluster Runtime packages Dynamo into version-locked Kubernetes recipes
In brief
- AICR bundles Nvidia's Dynamo and its Kubernetes dependencies into tested, version-locked recipes.
- AICR version 1.0, its first stable release, arrived at the start of October 2026.
- Dynamo deployment docs cover Amazon EKS on H100 and GB200-class GPUs.
- Kubernetes 1.34 or later is required for Dynamic Resource Allocation, according to the documentation.
What AICR actually does
Dynamo is Nvidia's open-source framework for distributed inference. It's built for disaggregated serving, where the work of answering a request gets split up and coordinated across GPU workers (instead of one machine handling every step). It also offers KV-cache-aware routing and multi-node orchestration.
Getting it running isn't simple. GPU-accelerated Kubernetes clusters have a lot of moving parts, and Crypto Briefing noted that a single mismatched version can stall an entire deployment.
AICR's answer is version-locked recipes, which pin every component to a specific release that Nvidia has confirmed works with the other components. The runtime splits the job into five workflows: snapshot, recipe generation, bundle rendering, deployment and validation. The snapshot step records a cluster's current state before anything changes.
Nvidia says AICR supports multiple cloud services and accelerators, and that it's backed by signed validation evidence. That evidence is a cryptographically signed record showing a given configuration was tested.
The fine print on EKS and DRA
Per Nvidia's documentation as described by Crypto Briefing, the Dynamo guidance covers Amazon's Elastic Kubernetes Service (EKS) running on H100 and GB200-class GPUs. In the EKS examples built around H100s, Nvidia used version 1.4.2 of the Dynamo Platform Helm chart.
There's a catch for teams on older infrastructure.
Dynamic Resource Allocation (DRA) is a Kubernetes feature for allocating specialized hardware such as GPUs more flexibly, and teams that want to use it need Kubernetes 1.34 or later. The October guides also stressed integrating the NVIDIA GPU Operator and the Dynamo Platform as part of a Kubernetes deployment.
Why the audit trail matters
The snapshot and validation steps don't just bookend a deployment. In Crypto Briefing's analysis, they give teams an audit trail that can help with debugging (and possibly with internal compliance reviews). That's the outlet's read, not a claim Nvidia has made about the product.
Frequently asked questions
What is NVIDIA AI Cluster Runtime (AICR)?
AICR is a tool that bundles Nvidia's Dynamo inference framework and its Kubernetes dependencies into tested recipes teams can deploy and then verify. Its recipes are version-locked, so each component is pinned to a release Nvidia has confirmed works with the others.
What workflows does AICR use to deploy Dynamo?
AICR splits the job into five workflows: snapshot, recipe generation, bundle rendering, deployment and validation. The snapshot step records a cluster's current state before any changes are made.
What Kubernetes version does Dynamic Resource Allocation require?
Teams that want to use Dynamic Resource Allocation (DRA) need Kubernetes version 1.34 or later. DRA is a Kubernetes feature for allocating specialized hardware such as GPUs more flexibly.
