KubeCon + CloudNativeCon North America 2026 ran from 9 to 12 November in Salt Lake City. As usual, Microsoft hosted its own co-located event on day zero: Azure Day with Kubernetes, a free full-day session on Monday 9 November with a keynote, deep dives, hands-on labs and an AMA with the AKS product team. A day later, Brendan Burns published Microsoft’s KubeCon announcement post.
This post collects the Microsoft announcements from that day and the accompanying blog in one place. I’ve limited it to Microsoft products and the open-source projects Microsoft is driving, and for each item I’ve added a short explanation of what it is and why it matters. It is based on Microsoft’s own announcement post and on public write-ups of the day; the sources are listed at the bottom.
A note on status: Microsoft’s blog does not label every item as GA or preview. Where a status below comes from slides shown on the day rather than from Microsoft’s written announcement, treat it as “as presented” and check the AKS release notes before you plan around it.
The short version
| Area | Generally available | Public preview | Coming / in development |
|---|---|---|---|
| Security | Azure Container Linux (IPE in audit mode), identity bindings, Pod Sandboxing with NAP | A/B updates for Azure Container Linux, KMS platform-managed keys by default | |
| AI infrastructure | Managed GPU node pools, confidential GPU, managed DRANET, Azure Container Storage v2.2.1 | Managed DRA for GPUs, multi-NIC pods, GPU memory profiling, AGC AI gateway | |
| Agents | Agent Sandbox as an AKS extension, pod snapshot and restore | ||
| Beyond Azure regions | AKS on bare metal, flex nodes, AI Investigator | ||
| Day-2 operations | Fleet drift detection, takeover and eviction; NAP image pins, capacity reservation groups, Machine API | Upgrade health gates, PDB Manager, annotation-based placement, NAP capacity buffers | Cross-cluster networking GA |
| Cost | Scheduling profiles, AKS Automatic managed system node pools | Managed Kueue extension | AKS Cost Export |
Dates to put in your calendar
Two announcements change default behaviour, so they deserve attention before the shiny things.
KMS with platform-managed keys becomes the default. New clusters on Kubernetes 1.37 get envelope encryption of etcd secrets out of the box, with keys that AKS manages and rotates. You no longer need to set up Key Vault to get KMS encryption. Existing clusters follow from February 2027, when they upgrade to 1.38. If you run customer-managed keys today for compliance reasons, this is the moment to check how the new default interacts with your setup.
Platform-initiated upgrades for unsupported clusters. From 1 June 2027, AKS will upgrade clusters that run an unsupported Kubernetes version once a 60-day grace period after end of support has passed. Microsoft’s documentation is blunt about it: such an upgrade does not follow your maintenance window, does not stop when deprecated APIs are still in use, and can override a PodDisruptionBudget that blocks a node drain. The change was already listed in the AKS release notes of 25 October; the day was a reminder rather than the first announcement. In other words: if you let a cluster fall out of support, Microsoft will eventually upgrade it for you, and not necessarily at a moment or in a way you would have chosen. Staying inside the support window was always good practice; it now has a deadline attached.
Security and the node OS
Azure Container Linux. This is Microsoft’s immutable container host, derived from Flatcar: a read-only OS with SELinux enforcing. It has been GA on AKS since June. What is new is Integrity Policy Enforcement (IPE), a Linux security module that went upstream in kernel 6.12. IPE checks that executables come from a signed dm-verity volume. On AKS it is GA in audit mode, which means untrusted or tampered binaries are logged but not blocked yet. Node auto-provisioning (NAP) can now create Azure Container Linux nodes, so the hardened host also applies to capacity that is created on demand.
A/B updates for Azure Container Linux are announced as coming soon. The new OS version is written to a second slot while workloads keep running, a reboot switches to it, and a failed update rolls back to the last known-good slot. This is the same model that Flatcar and Talos users already know.
Identity bindings are GA. With workload identity you previously needed a federated credential per cluster, and a managed identity is limited to 20 of those. Identity bindings let several clusters share one managed identity through a single federated credential. For anyone running the same workload across many clusters, that removes an awkward limit.
Pod Sandboxing with NAP is GA. Pod Sandboxing puts a VM boundary around each pod. Until now you had to build dedicated node pools for it; NAP can now provision them for you. Microsoft already listed this as GA in September, so this was a recap rather than news.
GPUs and AI infrastructure
Most of the AI news is about one upstream API: Dynamic Resource Allocation (DRA). The classic device plugin model lets a pod ask for “2 GPUs” and nothing more. With DRA a workload describes what it needs, such as a specific model or amount of memory, and the scheduler finds a matching device.
- Managed DRA for GPU node pools (public preview). AKS installs and runs the DRA driver for you.
- Managed DRANET (GA). DRANET is the DRA network driver from kubernetes-sigs. AKS runs it on RDMA-capable node pools, which matters for multi-node training and HPC jobs. A slide on the day claimed up to 4.8x RDMA bandwidth.
- Multi-NIC pods (public preview). Built on DRANET, this attaches dedicated network interfaces on separate subnets directly to a pod.
- Managed GPU node pools (GA). AKS manages the NVIDIA driver, the device plugin and the DCGM exporter, and GPU metrics land in Managed Prometheus and Grafana by default.
- Confidential GPU (GA). AMD SEV-SNP confidential VMs combined with NVIDIA H100 GPUs, for workloads where the data or the model must stay protected in use. Also GA since September.
- GPU memory profiling with Inspektor Gadget (public preview). Flame graphs for GPU memory, useful for finding leaks before a job dies with an out-of-memory error.
Around the GPUs themselves, a few storage and delivery pieces were shown:
- Azure Container Storage v2.2.1 (GA) for local NVMe.
- Gantry, an open-source project for peer-to-peer sharing of image layers between nodes. The test figures presented were striking: 99.2% less registry traffic and 24.5% faster pod starts when rolling a 40 GiB image out to 1,000 nodes. With model images this size, pulling from the registry is often the slowest part of scaling out.
- KAITO updates (open source): streaming models from Blob storage, MIG support, inference autoscaling based on KEDA, a memory estimator, and prefill/decode disaggregation.
- Application Gateway for Containers AI gateway (public preview, since June). One gateway for web applications, model endpoints and MCP servers, including routing that is aware of the KV cache on the model servers. If you read my posts on leaving ingress-nginx, this is the same AGC, now extended towards inference traffic.
Running agents on AKS
Agents were the other big theme. Two open-source building blocks were demonstrated on AKS:
- Agent Sandbox, a Kubernetes SIG Apps project that gives each agent an isolated, stateful pod, claimed from a warm pool.
- Agent Substrate, which packs many agents onto a small pool of warm worker pods. Each agent runs in its own microVM, and idle agents are snapshotted to disk and Blob storage. This project was started by Google and donated to the CNCF; Microsoft is contributing pluggable snapshot storage, Kubernetes identity support and work on the microVM runtime and worker lifecycle.
The density demo had more than 22,000 agents on 65 worker nodes, with only a few dozen active at any time and the rest suspended. The quoted figures were 550 agent activations per second and a median resume time of 160 ms.
None of this is a managed AKS feature yet. The roadmap, without dates, lists Agent Sandbox as an AKS extension, pod snapshot and restore, and an open-source distributed agent runtime on top of Agent Substrate.
AKS outside the Azure region
- AKS on bare metal (public preview) runs managed Kubernetes without a hypervisor, aimed at on-premises, edge and sovereign environments, with central visibility across the fleet. The preview itself dates from June; the day included physical devices running it.
- Flex nodes for AKS (public preview, since September) let you attach CPU or GPU capacity from another Azure region, your own datacenter or another environment to an existing AKS cluster. The typical example is bursting into a region where you still have GPU quota.
- AI Investigator for AKS (public preview) is an AI troubleshooting assistant aimed at disconnected and sovereign environments, where a cloud-hosted copilot is not an option.
For readers following the digital sovereignty discussion, this group is worth watching: it is Microsoft’s answer to “we want the AKS operating model, but not everything in an Azure region”.
Related, on the tooling side: Azure Copilot and Azure SRE Agent are GA, and Microsoft published open-source AKS skills for GitHub Copilot and for coding agents such as Claude Code and Codex.
Upgrades and multi-cluster operations
Upgrades. Microsoft grouped the upgrade work around two causes of failed upgrades:
- Readiness: upgrade insights checks a cluster before a Kubernetes or node image upgrade starts and flags restrictive PodDisruptionBudgets, deprecated APIs and quota or subnet limits, per cluster and across a fleet.
- Health during the upgrade: upgrade health gates (preview) watch node, cluster and workload health, can include signals from your own monitoring, and stop the rollout when things get worse.
- Also in preview: PDB Manager, shared maintenance windows, max unavailable with fallback, and a node disruption policy.
Node auto-provisioning. Node image pins, capacity reservation group support and Machine API integration are GA. Capacity buffers, which keep spare room for spiky and latency-sensitive workloads, are in preview.
Azure Kubernetes Fleet Manager, which is built on the CNCF Sandbox project KubeFleet, got three GA features:
- Drift detection reports resources on member clusters that no longer match what was placed.
- Takeover lets Fleet adopt workloads that already run on a cluster when it joins, under rules the platform team defines.
- Eviction removes resources from a member cluster without logging in to it, with a disruption budget that limits how many placements go offline at once.
New in preview is annotation-based placement: instead of writing a separate placement object, you add a kubefleet.dev/cluster-selectors annotation to a resource on the hub. For simple cases, such as “all staging clusters”, that is a lot less YAML. Cross-cluster networking was presented as reaching GA shortly.
It also helps to know how Fleet is packaged. Without a hub cluster it is free and gives you multi-cluster update orchestration for up to 1,000 member clusters. With a hub, at roughly the cost of a one-node AKS cluster, you get managed fleet namespaces, resource placement, Arc-enabled cluster support and the Flux GitOps extension (preview).
Cost and efficiency
- AKS Cost Export (coming soon) adds an AKS dataset to the Cost Management Exports API: daily cost per namespace and Azure asset, including idle cost, exported to Blob storage as CSV or Parquet. That is the raw material for proper showback without running a separate cost tool.
- Configurable scheduling profiles (GA) let you bin-pack pods instead of spreading them, so the autoscaler can remove empty nodes.
- Managed Kueue extension (public preview). AKS runs Kueue with default ResourceFlavors and priority classes; you define queues, quota and preemption.
- AKS Automatic managed system node pools (GA). Azure hosts, patches and repairs the system node pool, with no compute charge for it. This has been GA since Build in May.
- AMD v7 VM sizes, with Dadsv7 quoted at 1.8x the performance per dollar of Dadsv5.
Two upstream features were highlighted alongside these: HPA scale to zero (beta in Kubernetes 1.37) and in-place pod resize in VPA (alpha in VPA 1.7.0), which adjusts running pods without restarting them.
Developer experience
A sneak peek, still in development: an AKS desktop developer cluster plugin that runs a local single-node cluster in WSL with Cilium, intended for developing against open-weight models locally.
Open source
From Microsoft’s announcement post, the upstream work:
- Hyperlight (CNCF Sandbox), Microsoft’s lightweight hypervisor-based isolation, now integrates with Unikraft, which adds support for .NET, Python, Node.js and PowerShell workloads, and has extended support in the urunc runtime.
- AI Runway was accepted as a CNCF Sandbox project. It is a Kubernetes-native API for deploying and managing inference across providers and inference engines.
- Unbounded is in the queue for the CNCF Sandbox. It extends any conformant Kubernetes control plane with worker nodes in other clouds, on-premises and at the edge, with provisioning and cross-site networking built in.
- Sessions and Suite, two agent primitives: the first gives an agent run a durable lifecycle that can be suspended, resumed, replayed or branched; the second deals with portable packaging of agents.
- Lampyr, an early-stage project to bring this kind of isolation to serverless and lightweight Kubernetes workloads.
What I take from it
Three things stand out to me.
First, the defaults are moving. Platform-managed KMS keys and platform-initiated upgrades both mean that AKS will do more on your behalf whether you ask for it or not. That is good for the average cluster, and a reason to review your own if you have deliberately chosen something else.
Second, the AI story is becoming less about special AKS features and more about upstream APIs that AKS operates for you: DRA, DRANET, Kueue, Agent Sandbox. That makes the skills portable, which I appreciate.
Third, a lot of this is still preview. Bare metal, flex nodes, upgrade health gates and annotation-based placement are all worth testing, but I would not build a production design on them yet.
Sources
- What’s new with Microsoft in open-source and Kubernetes at KubeCon North America 2026 (Microsoft Open Source Blog, Brendan Burns)
- Azure Day with Kubernetes at KubeCon NA 2026: My AKS Highlights (Pixel Robots), a detailed first-hand account of the day and the source for the slide-level details above
- Microsoft Adds Managed DRA, Fleet Placement Controls and Upgrade Health Gates to AKS (Cloud Native Now)
- Platform-initiated upgrades in AKS (Microsoft Learn)
- What is Azure Container Linux for AKS? (Microsoft Learn)
- Run and scale AI applications on AKS (Microsoft Tech Community)
- Announcing the public preview of flex nodes for AKS (AKS Engineering Blog)
- Sandboxing Untrusted AI Agents at Scale (Google Open Source Blog, background on Agent Substrate)
- KubeCon North America 2026 co-located events schedule (Linux Foundation)