The infrastructure required to run artificial intelligence in production is becoming a major focus of the cloud native ecosystem. The Cloud Native Computing Foundation (CNCF) is expanding its portfolio to address that shift, adding projects focused on AI agent security, Kubernetes-native execution and distributed inference. Its latest ecosystem update highlights new Sandbox projects including OpenShell and Agent Substrate, alongside advances in Kubeflow, llm-d and Kubernetes. The foundation says its global cloud native developer community has reached 21.2 million, reflecting the growing overlap between AI infrastructure and open-source cloud technologies.
CNCF Builds Open Infrastructure for AI Agents and Inference
The infrastructure behind artificial intelligence is undergoing a significant transition. As organisations move from experimenting with generative AI to operating models and autonomous agents in production, they need platforms that can coordinate specialised hardware, distribute inference workloads, isolate untrusted code and enforce security policies beyond the model itself.
The Cloud Native Computing Foundation (CNCF) is expanding its open-source ecosystem to address these requirements. Its latest update, announced at KubeCon + CloudNativeCon North America 2026, highlights new and advancing projects for AI training, distributed inference and agent execution, including OpenShell, Agent Substrate, OpenRL and llm-d. The foundation says its portfolio now spans 231 projects, supported by a global cloud native developer community of 21.2 million people.
The developments point to a broader change in cloud native computing: Kubernetes and related open-source technologies are increasingly being adapted to run the full AI lifecycle, from model development and post-training to inference and autonomous task execution.
OpenShell Brings Agent Security Into CNCF
One of the most prominent additions is OpenShell, an open-source runtime originally spearheaded by NVIDIA and now accepted as a CNCF Sandbox project.
OpenShell addresses a security problem that becomes more pronounced as AI agents gain the ability to execute code, access files, connect to external services and interact with enterprise systems. An agent may need those capabilities to complete a task, but granting unrestricted access to credentials, networks and sensitive data creates a wider attack surface.
Rather than relying exclusively on instructions embedded in an agent’s prompt, OpenShell places policy enforcement in the infrastructure surrounding the agent. It is designed to govern how agents execute, what resources they can access and where inference requests are routed. NVIDIA’s technical description outlines a gateway, supervisor and sandbox architecture, with policy controls and audit capabilities intended to operate outside the agent itself.
The CNCF contribution moves the project into a vendor-neutral community environment where developers, platform engineers and security practitioners can participate in its direction. That matters as agent runtimes become part of production infrastructure rather than remaining isolated developer tools.
OpenShell is not the only project addressing agent execution. Agent Substrate, contributed by Google, provides a zero-trust runtime for executing AI agents on Kubernetes while isolating untrusted code. The two projects approach the problem differently: OpenShell emphasises policy enforcement and controlled access, while Agent Substrate focuses on efficiently managing isolated agent execution environments.
Kubernetes Expands Across the AI Lifecycle
The new project activity reflects a growing effort to make Kubernetes useful for more than deploying conventional containerised applications.
Agent Substrate and gVisor address workload isolation. gVisor uses a user-space kernel to provide an additional boundary between workloads and the host operating system, which is relevant when AI agents generate or execute code that cannot be fully trusted.
KubeRay brings Ray-based distributed computing workloads into Kubernetes environments. Ray is used to coordinate distributed applications, including AI training and inference workloads. Its Kubernetes integration gives platform teams a way to manage those workloads through familiar cloud native infrastructure.
OpenRL adds another component: Kubernetes-native orchestration for reinforcement learning. Post-training techniques such as reinforcement learning can require large numbers of parallel environments and substantial compute resources. OpenRL is intended to help coordinate those workloads on Kubernetes.
Together, these projects cover different layers of the AI execution stack: isolation, distributed processing, post-training orchestration and runtime governance.
Distributed Inference and Accelerator Management Advance
CNCF’s update also highlights progress in projects designed to make AI inference more scalable.
The llm-d project, which joined CNCF earlier in 2026, reached its 1.0 release. It is an open framework for distributed large language model inference, addressing the challenges of serving models across multiple computing resources. Its goal is to make distributed inference a first-class cloud native workload rather than requiring every AI team to build its own orchestration layer. CNCF+1
Kubeflow’s graduation is another important milestone. The project provides Kubernetes-native capabilities for data processing, model training, fine-tuning and model serving. Graduation recognises its technical maturity and established role in production AI and machine learning operations. CNCF previously reported that Kubeflow’s Python packages had accumulated nearly 260 million downloads, reflecting broad developer adoption.
Meanwhile, HAMi advanced to incubation, supporting the sharing and scheduling of heterogeneous accelerators. This is relevant to infrastructure operators managing environments that combine different GPU and other specialised computing resources.
Kubernetes itself is also progressing in workload-aware scheduling and Dynamic Resource Allocation. These capabilities help coordinate workloads that need access to specialised hardware, an increasingly important requirement as AI systems compete for accelerators, memory and other infrastructure resources.
The Kubernetes AI Conformance Program has reached 40 certified platforms, an effort intended to improve consistency across AI infrastructure implementations.
Open Governance Becomes a Strategic Advantage
The significance of these developments extends beyond individual software releases. AI infrastructure is becoming a strategic layer in the technology stack, and organisations increasingly want flexibility across models, accelerators, cloud providers and deployment environments.
Open-source governance can help reduce dependence on a single vendor’s implementation. It also gives competing technology providers a shared environment in which to contribute components, establish interfaces and improve interoperability.
CNCF’s membership expansion reflects that commercial interest. New Platinum members include Adyen, Broadcom, CoreWeave, CVS, HAProxy, NVIDIA, OVHcloud and Vultr. The group spans semiconductor technology, cloud infrastructure, networking and enterprise users, illustrating how broadly the AI infrastructure market now reaches.
The foundation also reports 82,000 attendees at cloud native and CNCF events across more than 50 countries during the past year. Its community added 58 Ambassadors and 28 Cloud Native Community Groups, while contributors published 40 case studies and 12 reference architectures.
These figures are CNCF-reported measures of ecosystem activity, rather than independent assessments of software adoption or production performance.
Skills and Standards Follow the Infrastructure Shift
The foundation is also preparing to expand its professional certification programme as AI workloads become more common in Kubernetes environments.
CNCF plans to introduce two AI-focused credentials in 2027: Kubernetes AI Associate (KAIA) and Certified Kubernetes Inference Engineer (CKIE). The certifications are intended to address skills related to AI infrastructure and inference operations. CNCF
The move reflects a growing need for engineers who understand both cloud native platforms and the operational requirements of AI systems. Deploying models at scale involves more than model development: teams must manage accelerators, networking, workload isolation, inference performance and security.
The larger message from CNCF’s latest update is that production AI is becoming an infrastructure discipline. As organisations deploy agents that can act on systems and data, they need controls that are enforced at the runtime level, along with open platforms capable of coordinating models and specialised hardware.
By bringing more of these capabilities into a shared open-source ecosystem, CNCF is seeking to shape how AI is trained, served and operated across clouds and infrastructure providers. Whether the projects become widely adopted production standards will depend on their technical maturity and implementation, but the direction is clear: cloud native infrastructure is being redesigned for an AI workload era.
4. Market Landscape
The cloud native ecosystem is moving beyond container orchestration toward infrastructure designed specifically for AI workloads. The shift is driven by three interconnected requirements:
- AI agent security: Agents need access to tools, files, credentials and networks, creating demand for runtime isolation and enforceable permissions.
- Distributed AI operations: Training, post-training and inference increasingly require coordinated access to GPUs, specialised accelerators and distributed computing resources.
- Open infrastructure and portability: Enterprises want flexibility across models, hardware and cloud providers without becoming dependent on a single proprietary AI stack.
CNCF’s project portfolio reflects this convergence. OpenShell and Agent Substrate address agent execution and isolation, OpenRL supports reinforcement learning workloads, while llm-d and Kubeflow contribute to distributed inference and the broader AI lifecycle.
The competitive landscape includes hyperscalers, chipmakers, cloud infrastructure providers and enterprise software companies. NVIDIA and Google are contributing important building blocks, while organisations such as CoreWeave and Broadcom represent the growing infrastructure investment behind AI workloads.
The strategic question is whether open-source projects can establish interoperable foundations that enterprises can deploy consistently across different infrastructure environments. CNCF’s vendor-neutral governance model gives these efforts a shared development environment, but adoption will depend on technical maturity, integration quality and production reliability.
Top Insights
- CNCF’s portfolio now spans 231 projects, with new contributions targeting AI agent security, distributed inference and Kubernetes-native AI operations.
- NVIDIA’s OpenShell introduces policy-enforced runtime controls designed to govern agent access to files, networks, credentials and model endpoints.
- Agent Substrate, gVisor and OpenRL expand open-source capabilities for isolated agent execution, workload security and reinforcement learning.
- The llm-d 1.0 release and Kubeflow graduation mark progress in distributed inference and production AI lifecycle management.
- CNCF plans two AI-focused certifications for 2027 as demand grows for engineers skilled in Kubernetes-based AI infrastructure.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
