NVIDIA Expands Open Source Commitment With OpenShell Donation and New AI

NVIDIA expanded its cloud-native AI software lineup at KubeCon + CloudNativeCon by donating its OpenShell agent runtime to the Cloud Native Computing Foundation as a CNCF Sandbox project. According to NVIDIA Blog's report, the company is also contributing to the Google-led Agent Substrate project and introducing three new open source components for its DSX AI Factory Platform. These releases provide standardized operational tooling to deploy, govern, and scale accelerated infrastructure across enterprise Kubernetes clusters.

This report details the architectural components, governance frameworks, and performance benchmarks that platform teams need to evaluate these open source additions. The coverage examines runtime policy controls, multi-cloud agent scheduling, cluster lifecycle recipes, and cold-start reduction metrics on production hardware. Platform architects operating large-scale GPU environments can use these concrete implementation details to plan agent deployment pipelines and shared cluster capacity.

01Open source runtime donation to CNCF

The contribution of OpenShell transfers ownership of the agent security runtime directly to the CNCF Sandbox under the Linux Foundation. Hundreds of organizations have adopted OpenShell, contributing tens of thousands of code updates within a year of its launch. Moving the codebase into the foundation allows vendors and researchers to inspect code paths, propose patches, and strengthen safeguards through neutral community governance.

OpenShell introduces deterministic controls that govern autonomous agent activities across corporate systems. Platform teams can define explicit operational boundaries when agents invoke specialized tools, communicate with inference services, or execute workloads. Because the architecture remains modular, administrators can connect proprietary compute pools, designated inference providers, internal credential vaults, and telemetry pipelines without modifying underlying agent logic.

Jim Zemlin, chief executive officer of the Linux Foundation, stated that the cloud-native community faces a new infrastructure challenge in providing secure operating environments for agents. He noted that donating OpenShell gives developers a shared foundation to build on and shape how agent security evolves. This community-driven approach establishes uniform security practices across heterogeneous enterprise environments.

02Agent Substrate integration with NVIDIA Dynamo

Deploying persistent autonomous workloads requires specialized lifecycle orchestration alongside policy enforcement. NVIDIA is contributing to Agent Substrate, an open source framework originated by Google that recently entered the CNCF Sandbox. The software orchestrates long-running agent tasks on Kubernetes clusters without stranding expensive accelerator resources during periods of inactivity.

Agent Substrate preserves task state across execution steps while actively releasing compute when agents sit idle. The framework suspends and restores environments across a shared pool of agents, preventing teams from dedicating compute to intermittent background sessions. This shared pool mechanism improves overall hardware utilization across multi-tenant GPU installations.

NVIDIA is engineering compatibility between Agent Substrate, NVIDIA Dynamo for model inference, and KAI Scheduler for cluster resource allocation. In parallel, NVIDIA and Google are collaborating on unified integration between OpenShell and Agent Substrate. This engineering effort allows enterprise teams to enforce governance policies and scale execution environments consistently across on-premises clusters and commercial clouds.

03Managed GPU infrastructure with public previews

Modern AI factories require managed Kubernetes environments that give operators control over underlying hardware, software configurations, and user experiences. To meet these requirements, NVIDIA extended its DSX AI Factory Platform software stack with two modular projects currently available in public preview. These components bridge the gap between raw accelerated compute and automated cluster orchestration.

NVIDIA Kubernetes Engine, known as NVKE, delivers standardized building blocks for deploying and managing production Kubernetes services on GPU infrastructure. The software automates complex cluster lifecycle tasks, eliminating manual administrative procedures during maintenance and expansion. Operators receive signed, version-pinned operating images to maintain configuration integrity across massive distributed computing installations.

NVIDIA Kubernetes Recipes, designated as NVKR, accompanies the engine to supply proven operational templates for cluster orchestration. These blueprints codify recommended practices for running AI workloads, streamlining provisioning steps from bare metal up to model serving tiers. Together, NVKE and NVKR allow operators to deliver managed services without ceding architectural sovereignty.

04Cloud ecosystem alignment across infrastructure providers

Broad adoption of AI factory components depends on immediate compatibility with established platform tooling and cloud vendors. A coalition of enterprise virtualization, container management, and cloud networking vendors has aligned with NVIDIA to integrate NVKE and NVKR directly into their software distributions. This partner ecosystem guarantees that organizations can deploy standard infrastructure recipes across existing infrastructure stacks.

Enterprise virtualization and orchestration providers Mirantis, OpenNebula Systems, and Rafay have announced concrete integration plans for the preview technologies. Container and networking specialists Netris, Spectro Cloud, and vCluster are similarly embedding the cluster building blocks into their platforms. These coordinated integrations allow enterprise customers to provision accelerated Kubernetes clusters within their preferred management consoles.

Global telecommunications and data center operators are also preparing production rollouts for these AI factory building blocks. Infrastructure provider Firmus and telecommunications operator SK Telecom have joined NVIDIA to adopt the platform tools within their AI cloud services. Their participation confirms that the new software addresses the scale requirements of international telecommunications networks and specialized hosting centers.

05Rapid inference restoration with NVIDIA Snapshot

Bringing additional inference replicas online quickly remains a critical bottleneck for bursty agentic workloads on Kubernetes. Standard scaling operations require loading deep neural network weights into memory and warming up execution graphs, delaying autoscaling responses. NVIDIA Snapshot resolves this latency hurdle as an open source component that restores checkpointed replicas in seconds without reloading weights or re-executing warmup routines.

Design partner Crusoe tested Snapshot on NVIDIA HGX B300 systems within Crusoe Cloud to measure performance gains on large models. Crusoe evaluated GLM-5.3 inference workloads and observed that total startup latency dropped from roughly eight minutes down to 31 seconds. As an upstream contributor, Crusoe verified that avoiding model reloads enables responsive elastic scaling during severe demand spikes.

Token factory platform Saturn Cloud benchmarked Snapshot across scale-from-zero inference scenarios on HGX B300 hardware running the vLLM serving engine. Saturn Cloud recorded startup times dropping from between two and 8.5 minutes down to between six and 34 seconds. These measurable reductions demonstrate that instant replica restoration enables cost-effective scale-from-zero architectures for high-capacity Kubernetes deployments.