NVIDIA Promotes Open Source Kubernetes for AI

NVIDIA aligns contributions and open-source tools to make Kubernetes a foundation for AI in production. With agent security through OpenShell, large-scale execution via Agent Substrate, and reported startup gains with Snapshot, the ecosystem is expanding around the DSX AI Factory Platform.

Partner tests report inference startups in seconds

Two design partners, Crusoe and Saturn Cloud, have evaluated NVIDIA Snapshot on NVIDIA HGX B300 systems. Crusoe, an upstream contributor to the project, tested GLM-5.3 on Crusoe Cloud and reports a reduction in preparation time from about eight minutes to 31 seconds. Saturn Cloud, positioned as a token manufacturing platform, assessed Snapshot for cold inference and indicates vLLM startups decreasing from two and 8.5 minutes to six and 34 seconds. Snapshot is an open-source, modular component capable of restoring saved replicas in seconds, without model reloading or preparation phase. This enables faster startups, recoveries, and elasticities for inference loads and agents on Kubernetes.

OpenShell places agent security under CNCF governance

OpenShell is now hosted by the Cloud Native Computing Foundation at the Sandbox level, with governance announced by the organizations integrating it into their products and expected reinforcement from security professionals. Jim Zemlin, CEO of the Linux Foundation, describes providing safe environments for agents as a new infrastructure challenge and sees OpenShell as a shared foundation to evolve agent security through broader collaboration. OpenShell allows control over AI agent actions and applies deterministic controls during the use of tools, access to inference services, or execution of loads. Its extensible architecture permits the integration of compute providers, inference services, credential storage, and monitoring tools, while facilitating consistent governance across environments and integration with enterprise systems. The project already claims hundreds of user organizations and tens of thousands of contributions in less than a year.

NVKE and NVKR in public preview for managed Kubernetes services

NVIDIA has made NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR) available in public preview. These open-source components aim to provide managed Kubernetes services on GPU infrastructure, automate the cluster lifecycle, and deliver signed and versioned images. They extend the DSX AI Factory Platform, whose stack covers GPU infrastructure, Kubernetes management, load orchestration, and model serving, to facilitate Kubernetes execution and accelerate inference. The launch is accompanied by ecosystem support: Mirantis, Netris, OpenNebula Systems, Rafay, Spectro Cloud, vCluster, Firmus, and SK Telecom have aligned plans to integrate NVKE and NVKR capabilities into their offerings. The goal is to help teams deploy, manage, and scale GPU infrastructures on Kubernetes as part of AI factories.

Agent Substrate and integrations for agent execution at scale

Agent Substrate, an open-source framework originally from Google accepted into the CNCF Sandbox, is receiving contributions from NVIDIA, which is working on its compatibility with NVIDIA Dynamo for inference and KAI Scheduler for resource allocation on shared clusters. The aim is to bring together agent execution, inference serving, and resource scheduling on Kubernetes. Agent Substrate maintains agent environments between tasks while freeing resources when they are idle, and allows for suspending and resuming these environments in a shared pool to avoid dedicating resources to each session. NVIDIA and Google are collaborating to integrate OpenShell and Agent Substrate to secure and govern agents across different clouds and on-premises deployments. AI factories are described as needing to offer managed Kubernetes services while retaining control over infrastructure, software, and customer experience. Kubernetes is presented as having become the foundation of AI, with increasing demands, and NVIDIA emphasizes the deepening of its open-source commitment at KubeCon + CloudNativeCon.