NVIDIA Donates OpenShell to CNCF and Unveils Open-Source Kubernetes Infrastructure for AI Factories

At KubeCon + CloudNativeCon, NVIDIA announced a major expansion of its open-source contributions aimed at securing autonomous AI agents and scaling enterprise AI infrastructure on Kubernetes.

As AI factories increasingly rely on Kubernetes to orchestrate massive GPU clusters, managing agent security, workload lifecycle, and model serving times has become a critical operational bottleneck. To address these demands, the chipmaker is handing governance of key agent tooling over to the community while introducing open-source software to accelerate GPU infrastructure deployment.

Securing AI Agents with OpenShell and CNCF

NVIDIA is contributing NVIDIA OpenShell to the Cloud Native Computing Foundation (CNCF), a nonprofit under the Linux Foundation. Accepted as a CNCF Sandbox project, OpenShell provides an open-source runtime designed to secure, monitor, and govern autonomous agents.

OpenShell allows enterprise teams to enforce deterministic controls over agent behavior as they access inference endpoints, execute workloads, and utilize third-party tools. Its modular architecture connects with an organization’s compute backends, credential vaults, monitoring suites, and inference providers. According to NVIDIA, hundreds of organizations have adopted OpenShell, with the project amassing tens of thousands of contributions within a year of its release.

Jim Zemlin, CEO of the Linux Foundation, highlighted that securing autonomous agents is becoming a defining challenge for cloud infrastructure, noting that moving OpenShell into CNCF ensures a shared community foundation to shape the future of agent security standards.

Collaboration with Google on Agent Substrate

In addition to OpenShell, NVIDIA is actively contributing to Agent Substrate, a Google-originated open-source framework accepted into the CNCF Sandbox. Designed for running long-running agent workloads at scale, Agent Substrate allows systems to suspend and resume agent execution environments across shared pools, freeing up GPU compute during idle periods without terminating ongoing sessions.

NVIDIA is aligning Agent Substrate with its Dynamo inference platform and the KAI Scheduler for shared cluster resource management. Furthermore, NVIDIA and Google are collaborating directly to integrate OpenShell with Agent Substrate, allowing developers to enforce security policies and manage agent execution seamlessly across hybrid and multi-cloud environments.

Accelerating GPU AI Factories with NVKE and Snapshot

To support organizations building managed AI factory infrastructure, NVIDIA revealed three new open-source components extending the NVIDIA DSX AI Factory Platform:

• NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR): Now in public preview, these projects offer automated cluster lifecycle management and signed, version-pinned images for managing GPU infrastructure on Kubernetes. Ecosystem partners including Mirantis, Netris, OpenNebula Systems, Rafay, Spectro Cloud, vCluster, Firmus, and SK Telecom are integrating NVKE and NVKR into their infrastructure stacks.
• NVIDIA Snapshot: An open-source, modular component built to tackle the latency of scaling inference clusters from zero. Rather than forcing nodes to reload model weights and undergo lengthy warm-up cycles, Snapshot restores checkpointed replicas in seconds.

Early evaluations highlight significant speed improvements on NVIDIA HGX B300 systems. Upstream contributor Crusoe tested the GLM-5.3 model on Crusoe Cloud, cutting readiness times from approximately eight minutes down to 31 seconds. Similarly, token platform Saturn Cloud benchmarked scale-from-zero inference using vLLM, lowering startup times from 2–8.5 minutes to between 6 and 34 seconds.