KubeCon North America opened in Salt Lake City this week with a keynote that would have been unthinkable eighteen months ago: two of the largest AI vendors handing their agent security plumbing to a foundation. Nvidia contributed OpenShell to the Cloud Native Computing Foundation as a Sandbox project on November 10. Google's Agent Substrate, contributed the week before, was accepted onto the same stage.
Neither project is a side experiment. OpenShell has drawn 380 contributing organizations since its first commit in January. Agent Substrate, which started in May, counts 164. Both landed in the CNCF's entry tier, which means the code is vendor neutral on paper and open to the contributors who will shape it in practice.
Two Sandboxes, One Question: Can You Trust an Agent?
When Nvidia first showed OpenShell earlier this year, it was a response to a bad season for autonomous systems. Several high-profile incidents had agents slipping past guardrails their operators assumed were holding. OpenShell's answer was to assume nothing: a closed-by-default shell wrapped around the agent, with policy applied afterward to decide what it may touch, whether it can reach the network at all, and whether traffic is allowed to egress only.
Justin Boitano, Nvidia's vice president of enterprise AI, framed the donation as a structural necessity rather than a gift. "We can't simply ask agents to follow the rules," he said in a video keynote. "The infrastructure must enforce them. AI agents have become part of cloud-native infrastructure. The technology that secures them should be open, vendor-neutral and governed by this community."
CNCF executive director Jonathan Bryce was blunter. He called OpenShell "extremely critical" and predicted it will be one of the most important projects of the year, precisely because it attacks the trust question head-on instead of adding another dashboard. He described the design as a shell wrapper plus policy: operators choose whether an agent can access the network, which protocols it may speak, and whether it can pull data in or only push it out.
What actually changed with the move:
- Governance: OpenShell leaves Nvidia's private roadmap and enters CNCF maintainer control, with a neutral home for competing vendors.
- Contributions: extensions and patches now flow back into a shared upstream instead of a single vendor's release train.
- Enforcement stack: a hardened container at the base, microVMs above it, and Sentry running policy checks at the hardware layer on BlueField-4 data processing units.
- On-ramp: the OpenShell reference architecture shipped last month, and IBM has already folded OpenShell and Sentry into its Fusion AI Factory.
CNCF CTO Chris Aniszczyk supplied the efficiency argument: at the OpenClaw launch, Mac Minis sold out worldwide, yet a long-running agent rarely uses that machine at full capacity. Density stops being a nice-to-have once agents move into data centers.
The Other Half: An Open GPU Stack
The agent runtime was only half of what Nvidia unloaded. On the same day, the company open-sourced three pieces of its DSX AI Factory platform under Apache 2.0, aimed squarely at the bare-metal clusters those agents will eventually run on.
- NVKE handles the cluster control plane and lifecycle provisioning, driving upstream Cluster API through a dedicated CAPNICo provider. Joining nodes authenticate with cryptographic SPIFFE identities, telemetry streams over OpenTelemetry, and the stack is conformance-tested on Kubernetes 1.36.
- NVKR supplies reproducible builds: signed, version-pinned operating system images, container images, and Helm charts for validated hardware configurations, which cuts configuration drift across large fleets.
- Snapshot captures running GPU workload state using CRIU plus GPU memory serialization. In tests on HGX B300 systems with Crusoe and Saturn Cloud, large inference replicas came back in 3.6 to 42.7 seconds, an 8x to 36x improvement over cold starts, with support for tensor parallelism and mixture-of-experts models across up to eight GPUs per node.
The overlap between OpenShell and Agent Substrate remains the open question. Boris Renski of HyperShell, a commercial supporter of OpenShell, mapped the split: Agent Substrate leans on flexible provisioning and governing access to code repositories, MCP servers, and skills, while OpenShell concentrates on individual API and tool operations, credential handling, and catching access policies that are accidentally permissive. Google's project packs agent sessions into constructs called Actors, smaller than pods and able to go dormant with snapshots that wake on demand. Nvidia says it will work with Google to integrate the two.
Swissquote's head of IT, Andrew Boswell, offered the most concrete reason an enterprise would care. The bank plans to run agents for 1,500 employees on its own hardware, pausing and resuming workloads so the same GPUs serve more people, while keeping the option to move if a model provider disappears.
Two foundations, one donor, and a pile of Apache 2.0 code. The interesting part is not that Nvidia gave OpenShell away. It is that the company decided trust in agents is infrastructure, and that infrastructure sells better when nobody has to ask permission to audit it.
Comments