Nvidia's version-locked recipes aim to shorten the path from a model configuration to a working inference endpoint
Nvidia wants to take the guesswork out of running its Dynamo inference framework on Kubernetes. Its tool for the job is NVIDIA AI Cluster Runtime, or AICR, which bundles Dynamo and its Kubernetes dependencies into tested recipes that teams can deploy and then verify.
What AICR actually does
AICR applies the idea of version-locked recipes to GPU clusters, meaning every component is pinned to a specific release that Nvidia has confirmed works with the others.
That matters because GPU-accelerated Kubernetes clusters have a lot of moving parts. Drivers need to match the hardware, resources need to be allocated correctly, and a single mismatched version can stall an entire deployment.
The runtime organizes the job into a sequence of workflows: snapshot, recipe generation, bundle rendering, deployment, and validation. The snapshot step captures the current state of a cluster, so teams know exactly what they started with before changing anything.
Recipe generation and bundle rendering then turn that information into a deployable package. Validation closes the loop by checking that the result works as intended.
Nvidia says AICR supports multiple cloud services and accelerators, backed by signed validation evidence. Put plainly, teams get a cryptographically signed record showing a given configuration was tested, not just a promise that it probably works.
The fine print on versions and hardware
AICR hit version 1.0 in early October 2026, marking its first stable release. Documentation for deploying Dynamo through the runtime followed around October 30 to 31, 2026.
That documentation includes guidance for Amazon’s Elastic Kubernetes Service, known as EKS, running on H100 and GB200-class GPUs. In the EKS examples built around H100 GPUs, Nvidia uses version 1.4.2 of the Dynamo Platform Helm chart.
There is one notable requirement. Teams that want to use Dynamic Resource Allocation, or DRA, need Kubernetes version 1.34 or later.
DRA is a Kubernetes feature for handing out specialized hardware like GPUs more flexibly. If your cluster is running an older Kubernetes release, that feature stays off the table until you upgrade.
Recent guides from October 2026 also walk through quickstarts for various models. They stress the importance of integrating the NVIDIA GPU Operator and the Dynamo Platform as part of a Kubernetes deployment.
Why Dynamo needs the help
Dynamo is Nvidia’s open-source framework for distributed inference. Inference is the part of AI where a trained model actually answers questions, generates text, or produces images for users.
The framework is built for disaggregated serving. Instead of one machine doing every step of answering a request, the work gets split up and coordinated across GPU workers.
Dynamo also offers KV-cache-aware routing. The KV cache is essentially a model’s short-term memory of text it has already processed, and routing requests with that memory in mind can avoid redundant work.
Add in multi-node orchestration, and you have a system that can spread a single serving operation across many machines.
What this means for teams running AI in production
The snapshot and validation steps target configuration drift directly. Capturing cluster state before deployment and checking it afterward gives teams an audit trail, which is useful for debugging and potentially for internal compliance reviews.
The signed validation evidence pushes in the same direction. Organizations that need to prove a configuration was tested, rather than just assert it, get documentation they can point to.
Support for EKS with H100 and GB200-class GPUs means teams on Amazon’s cloud have a documented route. The broader claim of support across multiple cloud services suggests Nvidia wants AICR to work wherever its GPUs are rented, not just in one provider’s ecosystem.
A few things are worth watching from here. The first is how quickly the Kubernetes 1.34 requirement for DRA becomes a practical hurdle, since many production clusters lag behind the latest releases.
The second is breadth. AICR’s usefulness scales with how many cloud, accelerator, and model combinations Nvidia validates, so the pace of new recipes after the 1.0 release will say a lot about its staying power.
