KubeCon North America 2026 Guide: AI Track, Co-located Days

When I opened the Monday grid for KubeCon North America 2026, the first thing I did was count: thirteen sessions starting at 10:40 in the part of the schedule I could read, all in different rooms. The problem with this conference is rarely a shortage of talks. It is that the pass, the flight and the Monday room all have to be decided before you can see which talks collide.

This guide is built from the CNCF announcement, the CNCF week-in-Salt-Lake-City post, the Platform Engineering Day post and the Linux Foundation event, track and schedule pages, all read on 17 October 2026. I am not attending on anyone's behalf. Where a page stated nothing, such as ticket prices, I say so instead of guessing, and every session time is a snapshot that can change.

What the official pages confirm, and what they leave out

The basics agree across every page I read. The conference runs 9 to 12 November 2026 in Salt Lake City, and the Linux Foundation pages place it at the Salt Palace Convention Center. Monday 9 November is pre-event programming, meaning CNCF-hosted co-located events and Project Lightning Talks. Tuesday to Thursday are keynotes, breakouts, the Solutions Showcase, Maintainer Tracks and ContribFests. The CNCF week post says there is no virtual component.

What the pages do not give me: ticket prices, a total session count, and anything about seat capacity. The announcement said standard registration closes on 2 September 2026, so I could not verify what a late registration costs or whether passes remain, and you should check the register page before planning around it. One small inconsistency is worth knowing: the announcement URL carries 10 August while its dateline reads 7 August, so I describe it only as an August announcement. Recordings of breakouts and keynotes go to the CNCF YouTube channel afterwards, within two weeks according to the published schedule.

The new AI Inference + Agentic track is the reason to read the schedule

The tracks page describes the track as exploring the intersection of generative AI and cloud native infrastructure. It lists five areas, and the announcement adds the project names it expects to see: vLLM, KServe, Ray and OpenTelemetry.

• Inference, serving and running models on Kubernetes
• Agentic systems built on open standards such as MCP, Agent Skills and A2A
• AI gateways
• GPU cluster optimisation, GPU slicing and dynamic scaling
• Energy-efficient inference

The announcement frames the track with two figures from the CNCF Annual Cloud Native Survey: 82% of container users run Kubernetes in production, and 66% of organisations using generative AI workloads rely on it for them. CNCF executive director Jonathan Bryce is quoted as saying Kubernetes was not built for AI and that the real engineering challenge is now running models in production. Read those numbers as the organiser's motivation rather than as a benchmark of your own platform.

Two Tuesday 11:30 sessions were visible in the part of the schedule I read: Ask Your Platform (Beginner, Grand Ballroom H) and Your Agent Needs a Platform (Advanced, room 255 BC, Soham Ghadge). The announcement also features Kubernetes Solutions for Agent-Shaped Problems by Tim Hockin and Dmitry Berkovich of Google, but I could not confirm its day, time or room from the schedule text I read, so look it up before you plan around it.

Monday: eleven co-located events, one room each

The co-located overview lists eleven CNCF-hosted events, and the Monday schedule gives each a room. The table pairs each event with one session I read in the grid, so you can judge the register of the day, plus my own one-line test for whether it is yours. The test is my judgement, not the organiser's. Platform Engineering Day is the one with its own CNCF post: a full day, a single morning track that splits into two after lunch, co-chaired by Paula Kennedy, Stacey Potter and Ramiro Berrelleza, aimed at platform engineers, product managers and technical leaders, with no prior platform knowledge required.

A limit on this table: the schedule page is very long and I read only its first part, so the afternoons of most events, and every session after the entries shown, are missing. OpenTofu Day and Kubernetes on Edge Day appear only from 1:20 PM in what I read, and the morning rooms are listed with Welcome and Opening Remarks at 9:00. Treat the table as a map of rooms, not a complete agenda.

The 10:40 problem: thirteen rooms, one body

Every co-located event opens at 9:00, and from 10:40 the grid fans out into parallel talks. In the part I read, thirteen sessions start at 10:40, in thirteen different rooms. A pass entitles you to any co-located event, so you can move between rooms, but you can only be in one. The script below turns your own priorities into a plan, using start times and rooms copied from the schedule, and builds a recording queue from what you skip.

// agenda-clashes.mjs  -  run with: node agenda-clashes.mjs
// Start times and rooms are copied from the CNCF-hosted co-located schedule
// (Monday 9 November 2026, MST). Weights are MY priorities: change them.
// Walking time between rooms is not modelled; the schedule gives no distances.
const picks = [
  { at: "10:40", room: "254", weight: 3, title: "From Routing Weights to Tail Latency: Simulating Kubernetes-Native LLM Inference Without GPUs" },
  { at: "10:40", room: "Grand Ballroom BDF", weight: 2, title: "Don't Lose Your Agent's Mind: Platform Patterns for Stateful Agents" },
  { at: "10:40", room: "Grand Ballroom ACE", weight: 2, title: "MCP Without Polling: Building Reactive Agents with Continuous Queries" },
  { at: "10:40", room: "250 A-C", weight: 1, title: "Designing an Internal SLO Platform" },
  { at: "11:15", room: "254", weight: 3, title: "Kueue, DRA, and WAS: The Kubernetes Scheduling Triangle" },
  { at: "11:15", room: "251 D-F", weight: 3, title: "llm-d Observability: Observing LLM Inference at Scale on Kubernetes" },
  { at: "11:15", room: "Room 151", weight: 2, title: "Securing the Repo Server: A Practical Guide to Argo CD's New Native mTLS" },
  { at: "11:50", room: "254", weight: 3, title: "What the Scheduler Doesn't Know: Why Your Inference Optimizations Aren't Working" },
  { at: "11:50", room: "250 A-C", weight: 2, title: "Inference Is a Black Box: How We Built GPU-Aware Observability with Fluent Bit and OpenTelemetry" },
];

const slots = new Map();
for (const p of picks) slots.set(p.at, [...(slots.get(p.at) ?? []), p]);

const watchLater = [];
for (const [at, group] of [...slots].sort()) {
  // Stable sort: on a tie the pick listed first wins, so order is a tiebreak.
  const [keep, ...lose] = [...group].sort((a, b) => b.weight - a.weight);
  console.log(at + "  GO  " + keep.room + "  " + keep.title);
  for (const l of lose) {
    console.log("       skip " + l.room + "  " + l.title);
    watchLater.push(l.title);
  }
}
// Co-located sessions are not live-streamed but are recorded to the CNCF
// YouTube channel, so the skipped list is your post-event queue, not a loss.
console.log("\nRecording queue: " + watchLater.length + " sessions");

I keep this deliberately dull. The weights are mine, ties go to whichever pick you listed first, and walking time is not modelled because the schedule publishes no distances. The useful part is the last line: co-located sessions are not live-streamed but all of them are recorded to the CNCF YouTube channel, so a skipped session is a deferred one. If you would rather stay in one room all morning, that is a defensible choice, and it is the one I would make for a room with a coherent morning such as the AI + Inference Day.

Tuesday keynotes: read the sponsor labels

All Tuesday keynotes run back to back from 9:00 to 10:45 in Hall DE on Level 1, in slots of five to ten minutes. The table shows the ones relevant to platform and AI work. The schedule marks some as Sponsored, and I keep that label because a sponsored five minutes is a different kind of evidence from a project update.

Two slots at 10:03 and 10:38 read Keynote to be Announced, and the Monday sponsored keynotes in the co-located rooms are listed only as TBA with the sponsor's name. After the keynotes, the same day has Maintainer Tracks covering autoscaling, device management, security, authentication and authorisation, Helm, Flux, Kyverno and KubeVirt, and ContribFests for Harbor, OpenTelemetry and Podman, some explicitly welcoming first-time contributors. If you have never touched an upstream project, a ContribFest is the cheapest way to meet maintainers.

Where this plan goes wrong

The first trap is the pass. The KubeCon + CloudNativeCon Only Pass does not include the CNCF-hosted co-located events, and the co-located overview states there is no single-day, co-located-only pass. If Monday is the only thing you want, you still buy the All-Access Pass. Accepted co-located speakers are given one automatically.

The second is assuming a ticket buys the talks. Breakout and keynote recordings are published after the event and the co-located sessions are recorded too, so what a ticket buys that a recording cannot is the hallway track, the maintainers and the Solutions Showcase. If a talk is the only reason you are travelling, the recording is the cheaper way to get it, and I would be wary of any plan that fails that test.

The third failure is mine to own: I judge sessions by title, abstract and track, and a title can promise more than a session delivers. I have not seen any of these talks, so the table and the clash script rank relevance to your stack, not quality. That is also why the recording queue matters more than it looks.

Pick rooms from your own cluster inventory

The cheapest filter is what you already run. The commands below look for the CRDs behind the Monday rooms, check whether your API server serves the Dynamic Resource Allocation group that the scheduling talk is about, and count the inference engines named in the new track. A room whose project you do not run is a room you can skip without guilt.

# Which projects with a Monday room do you actually run? Each hit below maps
# to a room: argoproj.io -> ArgoCon, fluxcd.io -> FluxCon, cilium.io -> CiliumCon.
kubectl get crd -o name | grep -E 'argoproj.io|fluxcd.io|cilium.io|kueue.x-k8s.io|opentelemetry.io'

# Dynamic Resource Allocation sessions only matter if this group is served.
# Empty output means your API server does not expose it, so skip those talks
# until you plan an upgrade.
kubectl api-resources --api-group=resource.k8s.io

# Do you already serve models? Count running images for the engines the new
# track names (vLLM, KServe, Ray) plus llm-d, which has its own Monday talk.
kubectl get pods -A -o jsonpath='{range .items[*]}{.spec.containers[*].image}{"\n"}{end}' \
  | tr ' ' '\n' | grep -iE 'vllm|kserve|ray|llm-d' | sort | uniq -c

Read the output as a prioritisation, not a verdict. Zero hits for argoproj.io but a strong Flux footprint means FluxCon, not ArgoCon. An empty resource.k8s.io listing means the DRA talks are about your next upgrade rather than this quarter. Hits in the third command mean Cloud Native AI + Inference Day and the Observability Day GPU session are probably your highest-value hours of the week.

What to do on Monday 19 October

Today is Saturday 17 October, so the next working Monday is 19 October, about three weeks before the event. Five steps turn this guide into a decision.

1. Decide whether Monday matters to you. If any co-located room in the table is yours, buy the All-Access Pass; if none is, the Only Pass covers Tuesday to Thursday and the lightning talks.
2. Run the inventory commands against your clusters and circle the rooms that match what you run.
3. Fill the picks array in the clash script with your own weights, run it, and treat the skip list as your recording queue.
4. Build your agenda in the official schedule tool, which the CNCF week post says offers a personal agenda builder, and open it again in the final week to catch changes.
5. Put a reminder two weeks after the event for the CNCF YouTube recordings, and write down which three talks you would act on in your own platform if they were good.

If you cannot attend, steps 2, 3 and 5 still work: the inventory and the clash script produce a watch list, and the recordings are public.

Choose the conference by the cluster in front of you, not by the schedule grid. Eleven Monday rooms and a new AI Inference + Agentic track are an invitation to over-plan, and the sources confirm only the structure, the rooms and the titles. Decide the pass first, filter rooms by what you run, expect to miss most sessions, and treat the recordings as part of the plan rather than the fallback.