AI
Why Standard Kubernetes Conformance Falls Short for AI
AI workloads bring requirements that standard Kubernetes conformance never anticipated. They depend on GPUs, specialized networking, synchronously scaled workloads, and specific security controls. Standard Kubernetes conformance only covers stable API endpoints. It says nothing about GPU drivers, accelerator types, or inference serving. The Kubernetes AI Conformance working group formed in 2025, sponsored by SIG Architecture, to close that gap. Requirements split into two tiers: “must” for mandatory platform behavior and “should” for community-recommended direction.
What the Program Certifies and What It Does Not
A platform must pass standard Kubernetes conformance before pursuing AI conformance. The program enforces capabilities on platforms, not practices on users. Users remain free to run only training workloads, only inference workloads, or both. The platform must support all those cases. The goal is the same write-once, run-anywhere guarantee Kubernetes conformance established, extended to AI and ML workloads. Conformance checklists exist for Kubernetes versions 1.33, 1.34, and 1.35, with each checklist requiring vendors to supply evidence per requirement.
From Working Group to Permanent Sub-Project
Working groups cannot own code. The program transitioned into an official SIG Architecture sub-project to give it a permanent home and allow it to hold a test repository. The repo contains design documents, requirements, discussion threads, and a Go-based automated test framework with no extra wrappers. Platforms specify their accelerator type as a flag when running the test suite, so a vendor using a non-standard accelerator can still pass. Future work may delegate DRA driver conformance to sub-areas, letting AI conformance simply require a conformant DRA driver.
The Vendor-Neutral Governance Model
Biweekly meetings run with 50 to 60 attendees. Competitors, including Google, Red Hat, Kubermatic, and Microsoft, sit alongside Nvidia, which ships no Kubernetes platform but still contributes requirements. The CNCF governing board protects the conformance trademark. The CNCF project team reviews every submitted distribution. The TOC holds technical oversight over all CNCF projects. The design intent is that a user can move between hyperscaler, on-premises, and edge environments and know the platform meets the same baseline. On-prem and edge coverage was the most debated topic in working group sessions.
Roadmap: Inferencing in 1.36 and Agents in 1.37
Kubernetes 1.36 introduces new requirements focused on inferencing. Janet Kuo live-demoed those requirements on the KubeCon keynote stage on Tuesday. Kubernetes 1.37 may add requirements around AI agents, though the scope is still open for community input. Requirement lifecycle follows a process called KCAR, modeled after the Kubernetes Enhancement Proposal process. Each KCAR document captures background, graduation criteria, testing plans, and implementation notes. The GitHub project board replaced Google Docs for tracking requirement status across community meetings.
Q&A
Is there value in joining the meetings as a newcomer with no current AI deployments? The maintainers welcome any input, including early-stage use cases the group may not have considered, and many attendees join simply to learn what the program covers. ▶ 17:45
Notable Quotes
we only put uh the limitation or enforcement on the platforms so they should follow the standard to provide the capabilities or APIs in the platform so that it can cover most of the uh AI workflow use cases and users are free to choose however they want to use those clusters. Janet Kuo · ▶ 02:59
basically we have like meetings of 50 to 60 people to give input from different sit and it’s not only vendors we also have folks from Nvidia in there they have no Kubernetes platform but they still give the input to define those standards Mario Fahlandt · ▶ 12:44
you’re not locked in into a specific vendor and the community works together and as you have seen like technically all of us are competitors and we basically work together to get all of the input to um have like a standard that defines the industry Mario Fahlandt · ▶ 12:12
Key Takeaways
- Kubernetes AI conformance sits on top of standard Kubernetes conformance, requiring platforms to pass both.
- Conformance checklists already cover Kubernetes 1.33 through 1.35, with inferencing requirements landing in 1.36.
- The program is now a permanent SIG Architecture sub-project and owns an automated Go test suite.
About the Speaker(s)
Janet Kuo is a Staff Software Engineer at Google Cloud and a foundational contributor to Kubernetes, having helped shape the project since before its 1.0 launch in 2015. She serves as a Kubernetes maintainer and lead of SIG Apps.
Yuan Tang is a Senior Principal Software Engineer at Red Hat AI. He holds leadership positions across Argo, Kubeflow, KServe, Kubernetes, and CNCF, and has authored multiple technical books and popular open source projects.