NVIDIA Strengthens Its Open Source Commitment by Donating OpenShell and Launching New AI Infrastructure
NVIDIA backs OpenShell, Agent Substrate, and Kubernetes tools for secure, scalable AI agents and faster GPU inference.
Key Takeaways
- NVIDIA contributed OpenShell to the Cloud Native Computing Foundation Sandbox, making it an open-source runtime for securing and governing autonomous AI agents with rule-based controls.
- NVIDIA is supporting Agent Substrate, a Google-led Kubernetes framework that runs long-lived agent workloads while freeing idle compute through suspend-and-resume scheduling.
- NVIDIA added Kubernetes Engine, Kubernetes Recipes, and Snapshot to its DSX AI Factory Platform to help teams deploy, manage, and scale GPU infrastructure on Kubernetes.
- The announcements aim to strengthen Kubernetes as a foundation for secure, efficient, and scalable AI infrastructure, including faster inference startup and elastic workload management.
Meta description: NVIDIA contributes OpenShell and supports Agent Substrate at KubeCon, adding Kubernetes tools for secure AI agents and GPU infrastructure scale.
Tags: Kubernetes, NVIDIA OpenShell, Cloud Native Computing Foundation, AI agents, GPU infrastructure
Featured image: Abstract friendly illustration of Kubernetes clusters, GPU racks, and secure agent shields, with no people, logos, or text.
At KubeCon + CloudNativeCon, the flagship conference for Kubernetes and cloud-native communities, NVIDIA announced that it is deepening its open source commitment by contributing NVIDIA OpenShell to the Cloud Native Computing Foundation, supporting Google-led Agent Substrate, and adding new components to the NVIDIA DSX AI Factory Platform. This is for infrastructure teams building AI systems on Kubernetes; they get open tools for agent governance, scaled agent execution, and faster GPU inference startup.
The announcement matters because Kubernetes is an open source system used to run and coordinate application workloads across many machines. It now serves as a foundation for AI factories, and demand on it is rising as teams move to agents that use tools, access inference services, and execute workloads. NVIDIA framed the work as a way to make Kubernetes a stronger base for scaled AI workloads while giving developers open tools to build safer and more efficient infrastructure.
OpenShell moves agent security into open source
NVIDIA OpenShell is a runtime built to secure and govern autonomous agents. It is being contributed to the Cloud Native Computing Foundation, a Linux Foundation nonprofit that hosts KubeCon, and will enter the Cloud Native Computing Foundation Sandbox. Under the Sandbox model, the groups incorporating it into products will help steer it, while security experts in the community can review and improve it.
The runtime applies deterministic, rule-based controls around agent behavior. Policy can cover tool use, inference endpoint access, and workload execution. Because the architecture is extensible, teams can attach their own compute resources, inference providers, credential stores, and monitoring services. The goal is consistent governance across deployment models while keeping the runtime compatible with existing enterprise systems.
According to NVIDIA, hundreds of organizations have already adopted OpenShell, and the project accumulated tens of thousands of contributions in under one year after launch. Moving it into the Cloud Native Computing Foundation Sandbox is intended to speed development by letting the community inspect and improve the safeguards.
Jim Zemlin, CEO of the Linux Foundation, described the community's new infrastructure challenge as providing agents with safe places to run. He said the shared foundation created by the OpenShell contribution lets developers shape how agent security evolves and supports broader collaboration on infrastructure that next-generation cloud-native applications will depend on.
Agent Substrate helps run agents without idle compute
Security is only one side of the agent problem. Teams also need to run many agents for long periods without paying to keep dedicated compute active between tasks. NVIDIA is also supporting Agent Substrate, a framework started by Google for running long-lived agent workloads on Kubernetes. It has been accepted into the Cloud Native Computing Foundation Sandbox. NVIDIA is connecting that framework with NVIDIA Dynamo for model inference and KAI Scheduler for shared cluster resource allocation.
The framework keeps an agent's environment available between tasks, then frees the compute used by that environment when the agent is idle. Instead of assigning permanent capacity to each session, it suspends and resumes environments across a shared pool. That approach lets organizations scale agent workloads on Kubernetes while tying together execution, inference serving, and cluster resource scheduling.
NVIDIA and Google are also collaborating to connect OpenShell and Agent Substrate. The collaboration aims to let enterprises secure and govern agents across clouds and on-premises systems.
DSX tools target managed GPU Kubernetes and fast inference startup
Alongside the agent projects, NVIDIA added NVIDIA Kubernetes Engine, NVIDIA Kubernetes Recipes, and NVIDIA Snapshot to the NVIDIA DSX AI Factory Platform. The components are designed to help teams deploy, manage, and scale GPU infrastructure on Kubernetes for AI factories. The software stack spans GPU infrastructure, Kubernetes management, workload orchestration, and model serving. Together, the projects extend the platform to make Kubernetes easier to run and inference faster to scale.
NVIDIA Kubernetes Engine and NVIDIA Kubernetes Recipes are available in public preview. They are open source building blocks for managed Kubernetes services on GPU infrastructure. The tools automate cluster lifecycle tasks and supply signed, version-pinned images. The goal is to let AI factories provide managed Kubernetes services while retaining control over infrastructure, software, and customer experience.
The two tools launch with ecosystem support that extends their reach across AI infrastructure platforms and cloud services. Mirantis, Netris, OpenNebula Systems, Rafay, Spectro Cloud, and vCluster, together with Firmus and SK Telecom, have aligned with NVIDIA on integration plans to bring NVIDIA Kubernetes Engine and NVIDIA Kubernetes Recipes capabilities into their offerings.
NVIDIA Snapshot targets the delay that occurs when extra inference capacity is needed. Model loading and workload warmup can slow scaling and recovery. The open source, modular component restores checkpointed replicas in seconds without reloading the model or warming it up, helping inference and agentic workloads on Kubernetes start faster, recover faster, and scale elastically.
Crusoe and Saturn Cloud, both design partners, tested Snapshot on their own workloads with NVIDIA HGX B300 systems. Crusoe, which contributes upstream to Snapshot, ran GLM-5.3 on Crusoe Cloud and shortened the time for the workload to become ready from roughly eight minutes to 31 seconds. Saturn Cloud, described as a token factory platform, benchmarked scale-from-zero inference with Snapshot and cut vLLM startup times from between two and 8.5 minutes to six and 34 seconds.
What this changes for AI infrastructure teams
For readers evaluating agent platforms, the takeaway is that NVIDIA is positioning Kubernetes as the common ground for three needs: governance for autonomous agents, efficient long-running execution, and faster GPU-backed inference scaling. OpenShell answers the question of how to constrain agent actions. Agent Substrate answers how to run many agents without dedicated idle compute. NVIDIA Kubernetes Engine, NVIDIA Kubernetes Recipes, and Snapshot answer how to operate managed GPU clusters and reduce startup delays.
The next developments to watch are whether Cloud Native Computing Foundation Sandbox governance helps OpenShell and Agent Substrate gain wider adoption, and whether NVIDIA DSX integrations make managed Kubernetes on GPU infrastructure easier for AI factories to deliver. NVIDIA invited readers to learn more by joining it at KubeCon.
Comments (0)