NVIDIA announced a set of open source contributions at KubeCon + CloudNativeCon, as described on the NVIDIA Blog. The headline item is OpenShell, a runtime for securing and governing autonomous AI agents, which NVIDIA is handing to the Cloud Native Computing Foundation (CNCF) as a Sandbox project.
What OpenShell does
OpenShell applies deterministic controls to what an agent does. Teams can enforce policies when an agent calls tools, reaches inference services or runs workloads. Its architecture is extensible, so an organization can plug in its own compute, inference providers, credential storage and monitoring, and apply the same governance across different deployment environments and existing enterprise systems.
Under CNCF Sandbox status, the project will be governed by the organizations building it into their products. NVIDIA says hundreds of organizations already use OpenShell, and that it has drawn tens of thousands of contributions in under a year since launch.
Agent Substrate and scaling
NVIDIA is also contributing to Agent Substrate, a Google-originated framework for running long-running agent workloads on Kubernetes that is already in the CNCF Sandbox. It keeps an agent's environment between tasks and frees compute while the agent is idle, so teams do not need dedicated compute for every session. NVIDIA is making it work with its Dynamo inference software and KAI Scheduler for shared cluster resources. NVIDIA and Google are also working on an integration of OpenShell and Agent Substrate for governing agents across clouds and on-premises systems. The source describes that as ongoing collaboration, not a shipped product.
New Kubernetes components for GPU infrastructure
Three additions extend the NVIDIA DSX AI Factory Platform:
- NVIDIA Kubernetes Engine (NVKE) and Kubernetes Recipes (NVKR), in public preview, are building blocks for offering managed Kubernetes on GPU infrastructure. They automate cluster lifecycle management and supply signed, version-pinned images. Mirantis, Netris, OpenNebula Systems, Rafay, Spectro Cloud, vCluster, Firmus and SK Telecom have aligned on plans to bring them into their offerings.
- NVIDIA Snapshot restores checkpointed inference replicas in seconds without reloading the model or warming up, which speeds startup and recovery.
NVIDIA cites two design partners using HGX B300 systems. Crusoe tested GLM-5.3 and cut the time to readiness from roughly eight minutes to 31 seconds. Saturn Cloud reported vLLM startup falling from between two and 8.5 minutes to between six and 34 seconds in scale-from-zero tests. These are vendor-published results from partner workloads, not independent benchmarks.
What is not stated
The source gives no dates for OpenShell reaching any particular maturity level, no pricing, and no detail on how the integration plans of the listed partners will reach customers.
Top comments (0)