NVIDIA is donating its OpenShell runtime to the Cloud Native Computing Foundation (CNCF) and contributing to the Agent Substrate framework to standardize AI agent security and execution.
These initiatives, announced at KubeCon + CloudNativeCon, provide the infrastructure needed to deploy, govern, and scale autonomous AI workloads across enterprise Kubernetes clusters.
NVIDIA is addressing this by donating OpenShell to the CNCF as a sandbox project. According to the official company blog, OpenShell acts as a deterministic control plane, allowing organizations to enforce granular policies on how agents access inference services, execute tools, and run workloads.
The project’s architecture is intentionally extensible, meaning it supports integration with existing enterprise credential storage and monitoring stacks. This provides the security guardrails necessary for enterprise environments. Jim Zemlin, CEO of the Linux Foundation, noted that the contribution provides a “shared foundation to build on and an opportunity to shape how agent security evolves.”
While OpenShell handles the “who” and “what” of agent security, Agent Substrate—a Google-led CNCF project—manages the “how” of execution at scale. NVIDIA is contributing to this framework to ensure it works with its own ecosystem, specifically the Dynamo inference engine and KAI Scheduler for resource allocation.
The primary technical benefit here is efficiency. Agent Substrate allows systems to preserve agent environments between tasks while releasing compute resources during idle periods. By enabling the suspension and resumption of these environments across shared clusters, organizations can scale their agentic workloads without the overhead of maintaining dedicated compute for every active session.
NVIDIA and Google are currently collaborating to unify these tools, ensuring that agent governance and execution remain consistent whether running in the cloud or on-premises.
Beyond agent-specific tools, NVIDIA is pushing further into the Kubernetes management layer with three new components for its DSX AI Factory Platform. These tools help teams deploy, manage and scale GPU infrastructure on Kubernetes for AI factories.
- NVIDIA Kubernetes Engine (NVKE) & Recipes (NVKR): These provide automated cluster lifecycle management with signed, version-pinned images. They are currently in public preview with backing from ecosystem partners including Mirantis, Netris, and SK Telecom.
- NVIDIA Snapshot: This component targets the latency bottleneck of model loading and workload warmup. It restores checkpointed inference replicas in seconds, significantly reducing recovery times.
The following performance data, reported by NVIDIA, demonstrates the impact of the Snapshot component on startup and recovery times for AI inference workloads:
| Use Case | Original Startup Time | Snapshot Startup Time |
|---|---|---|
| GLM-5.3 (Crusoe Cloud) | ~8 minutes | 31 seconds |
| vLLM Inference (Saturn Cloud) | 2–8.5 minutes | 6–34 seconds |
By moving these components into the CNCF, NVIDIA is deepening its commitment to open source. Tradersunion.com reports that the move is an effort to make Kubernetes the leading foundation for AI factories, where performance and scalability are handled by standardized, open-source building blocks. The alignment of third-party infrastructure providers like Rafay and Spectro Cloud suggests a concerted effort to bake these capabilities directly into managed service offerings.
For developers, the integration of these tools—specifically the ability to move from minutes to seconds in workload readiness—means that elasticity is no longer just a theoretical feature of cloud-native AI.
It is becoming a functional reality, provided they adopt the emerging standards set forth by these new CNCF contributions. As of this week, these projects are being integrated into the broader cloud-native ecosystem, with further updates expected as they mature within the CNCF sandbox.
