NVIDIA Expands Open Source Commitment With OpenShell Donation and New AI Infrastructure

At KubeCon + CloudNativeCon, NVIDIA said it is contributing OpenShell, its open source runtime for securing and governing autonomous agents, to the Cloud Native Computing Foundation as a Sandbox project. It is also contributing to Agent Substrate, a Google-led framework for running long-lived agent workloads on Kubernetes, and it announced three new open source pieces of its DSX AI Factory Platform: NVIDIA Kubernetes Engine, NVIDIA Kubernetes Recipes and NVIDIA Snapshot.

The headline for most teams is governance. Agents that call tools, hit inference services and run workloads need guardrails that sit outside the model. Putting that layer under community governance, instead of a single vendor's roadmap, changes who gets a say in how agent security evolves.

We think the OpenShell donation is a genuinely useful move, and we want to give it credit before we get skeptical. The approach NVIDIA describes is deterministic control over agent activity: policies enforced as agents use tools, reach inference and execute work. That is the correct instinct. You do not want the thing deciding whether an agent can touch production to be another probabilistic model. Rules that behave the same way every time are auditable, and auditability is what security teams will demand before they let agents loose.

NVIDIA says hundreds of organizations already use OpenShell and that it drew tens of thousands of contributions in under a year. If that holds, moving it to CNCF is less about charity and more about making it the default. That's where our eyes-open part kicks in. A neutral home is better than a proprietary one, but the integrations NVIDIA highlights tie Agent Substrate to its own Dynamo inference stack and KAI Scheduler. Open governance and a hardware vendor's gravity can coexist; we're watching whether OpenShell stays as useful on non-NVIDIA infrastructure as it is on NVIDIA's.

Agent Substrate is the more interesting engineering idea. It keeps an agent's environment alive between tasks while releasing compute when the agent is idle, suspending and resuming across a shared pool. That is how agent fleets become affordable. It is also how they become long-lived, stateful and harder to reason about. An agent that can be paused and resumed for weeks carries credentials, context and permissions with it. The OpenShell and Agent Substrate integration NVIDIA and Google say they are building is the piece that has to get this right, because a suspended agent with stale permissions is a quiet liability.

On the factory side, Snapshot's numbers are the eye-catcher, and they come from design partners, not independent testing. Crusoe says a GLM-5.3 workload went from roughly eight minutes to ready down to 31 seconds; Saturn Cloud reports vLLM startup falling from a two-to-8.5-minute range to between six and 34 seconds. Faster cold starts mean less idle GPU held in reserve, which is good for cost and for energy. We'd still like to see results on more varied hardware before treating it as typical.

Where this leads if it scales: a cloud-native stack where agents are first-class workloads, governed by shared policy engines and scheduled like any other job. That is better than every company inventing its own sandbox. The downside is concentration, with the tools that decide what agents may do coming increasingly from the company that sells the chips they run on.

Kubernetes is the open source system that orchestrates applications across clusters, and CNCF, part of the Linux Foundation, hosts it and KubeCon. Sandbox is CNCF's entry stage for early projects. NVIDIA Kubernetes Engine and Kubernetes Recipes are in public preview and aim to help providers offer managed Kubernetes on GPU infrastructure, with integration plans announced by firms including Mirantis, Rafay, Spectro Cloud, vCluster and SK Telecom.