KubeCon and CloudNativeCon North America runs from 9 to 12 November in Salt Lake City. The schedule has the usual shape, with one change that is worth more attention than the rest of the programme put together. There is now a dedicated track for AI inference and agentic workloads.
Conference tracks sound like administrative detail. They are not. A track is where a foundation decides to put its stage time, its reviewers and its reputation, and the decision is made months ahead by people reading what the community actually submitted. A new track means the talks were already being written.
What the new track covers
The AI Inference and Agentic track takes in model serving, GPU scheduling, agent workflows and observability for AI systems running in production. The named projects include vLLM, KServe, Ray and OpenTelemetry. Read that list again and notice what it is: a serving layer, a scheduler, a distributed runtime and an instrumentation standard. That is not a track about machine learning. It is a track about operations, aimed at the people who will be paged when a model stops responding.
The foundation is citing a figure of 66 percent of organisations running generative AI workloads doing so on Kubernetes. Treat the precise number with the usual caution owed to any vendor survey, but the direction is not controversial. The question of where inference runs has largely been settled, and the interesting problems have moved to how you schedule expensive hardware, how you serve a model without wasting it, and how you tell whether any of it is working.
The other two pillars
Platform engineering remains a main track, covering internal developer platforms, self service and the operational practice of running software delivery at scale. After several years of debate about whether platform engineering was a real discipline or a rebrand, the argument has gone quiet, which is usually what happens when something becomes ordinary.
Security covers supply chain, identity, runtime protection and vulnerability management, with Cilium and eBPF prominent. The pattern here has been consistent for a while: the interesting security work keeps moving down the stack towards the kernel, because that is the one place where an attacker cannot easily lie about what a workload is doing.
The Monday before
The co-located events on Monday 9 November are frequently better value than the main conference, because they are smaller, the audience is self selected and the hallway conversations are with people who have the same problem as you. This year includes Cloud Native AI and Inference Day, ArgoCon, BackstageCon, CiliumCon and WasmCon.
If you only have budget for part of the week, the co-located day covering your actual problem is usually the better purchase. Main stage keynotes get recorded. The side rooms do not, in any meaningful sense, because what you went for was the conversation afterwards.
How to plan it
Pick one problem you are currently stuck on and build the week around it, rather than sampling across every track. People who try to see everything come home with notes they never read.
Go to the maintainer sessions for projects you already run in production. The track talks explain what a project does. The maintainer sessions tell you where it is going and what is about to be deprecated, which is the part that affects your next twelve months.
If your organisation is anywhere near putting models into production, the inference track is the one to prioritise, because the practice here is still being formed and the gap between the teams doing it well and everyone else is currently enormous.
If you cannot attend, the main stage recordings usually appear within a few weeks. Watch the maintainer and track talks over the keynotes. The keynotes are for the industry. The track talks are for you.
What it signals
Three years ago the Kubernetes conversation was about whether you needed it at all, and the honest answer for many teams was no. Two years ago it was about the platform teams built on top. This year the foundation has given its own stage to the question of how to run inference well, which tells you the centre of gravity has moved again.
For teams hiring against this, our September hiring report covers which of these skills employers are actually asking for, and our AI infrastructure guide covers the serving layer the new track is built around.
Sources: the CNCF schedule announcement for KubeCon and CloudNativeCon North America 2026, and CNCF event listings. Checked on 3 October 2026.
