The short answer: someone who upgrades them on your schedule, before the provider does it on its own. All three major clouds now move clusters that fall too far behind. Amazon EKS bills six times the standard cluster fee for a version past standard support, then upgrades the control plane when extended support ends, at a time it does not announce. Google Kubernetes Engine upgrades clusters that run past the end of support. On 25 October 2026 Microsoft added Azure Kubernetes Service: from 1 June 2027 it will upgrade unsupported clusters outside your maintenance window, without stopping for APIs your workloads still use. On 10 November, at KubeCon in Salt Lake City, the Cloud Native Computing Foundation showed how much more those clusters now carry, from virtual machines to AI models. Keeping them current is a recurring operations job, and it needs an owner.
The dates that matter
| Date | Provider | What happens |
|---|---|---|
| 25 October 2026 | Microsoft | AKS release notes announce automatic upgrades of unsupported clusters from 1 June 2027 |
| November 2026 | Microsoft | Community support for Kubernetes 1.34 and long-term support for 1.31 end on AKS |
| 10 November 2026 | CNCF | KubeVirt and Kubeflow graduate; 40 platforms are certified under the Kubernetes AI Conformance Program |
| 26 November 2026 | AWS | Extended support for 1.31 ends on EKS, and remaining control planes are upgraded automatically |
| 2 December 2026 | AWS | 1.34 leaves standard support on EKS, and extended support billing begins |
| 25 January 2027 | 1.34 reaches the end of standard support on GKE, by Google’s estimate | |
| 1 June 2027 | Microsoft | AKS begins upgrading clusters that remain on unsupported versions |
How each cloud treats a cluster that falls behind
Amazon EKS
Each Kubernetes version gets 14 months of standard support on EKS, followed by 12 months of extended support that is switched on by default. Standard support costs USD 0.10 per cluster hour and extended support USD 0.60, so a cluster that drifts into extended support costs about USD 438 a month in fees instead of about USD 73, before a single node runs. When extended support ends, EKS upgrades the control plane to the next version at a time it does not specify, with no notice beforehand, and that upgrade cannot be rolled back. Managed and self-managed node groups stay on the old version, and add-ons must be updated by hand.
Azure Kubernetes Service
AKS gives each version a year of community support. Long-term support adds a second year, but it is an opt-in that requires the Premium tier. After a version reaches end of life, a cluster has a 60-day grace period, and AKS will then upgrade it itself, starting on 1 June 2027 for clusters that are already out of support. That upgrade ignores the maintenance window and the auto-upgrade channel, does not stop when workloads use APIs the new version removes, can override a Pod Disruption Budget that blocks a node drain, and cannot be cancelled or rolled back. A cluster several versions behind is moved one minor version at a time, and there is no opt-out.
Google Kubernetes Engine
GKE upgrades clusters that run past the end of support for their minor version. Its Extended channel lets a cluster stay on a version for up to 24 months, with pay-per-use charges once that version enters extended support, and when extended support ends Google upgrades every cluster still on it, whatever blocks it. Google describes its release dates as best-effort estimates.
Why the stakes rose at KubeCon
The CNCF’s announcements of 10 November show Kubernetes taking on work that used to run elsewhere. KubeVirt, which runs virtual machines on Kubernetes alongside containers, graduated, the foundation’s highest maturity level. It counts 2,088 contributors from 520 organisations, and the CNCF cites a 2025 survey in which 26 percent of Kubernetes adopters used it in production. Kubeflow, which runs machine learning workflows, graduated the same day, and the Kubernetes AI Conformance Program reached 40 certified platforms. A company that moves its virtual machines off VMware, or deploys its first AI models on Kubernetes, puts more of the business on clusters that receive a new minor version on average every four months.
Few companies have a team dedicated to that. In the CNCF and SlashData Technology Radar of March 2026, 28 percent of organisations reported a dedicated platform engineering team. The tooling is widespread all the same: the foundation’s State of Cloud Native Development report for the third quarter, released on 10 November, finds 93 percent of backend developers working at organisations that use DevOps standardisation or an internal developer platform, up from 77 percent a year earlier.
Six things to put in place before the next deadline
- List every cluster and its support date. Record the provider, region, Kubernetes version and the day it leaves standard support, including test clusters nobody owns any more.
- Check the policy each cluster is on. EKS enrols clusters in extended support by default. On AKS, a cluster with no auto-upgrade channel, or only the patch channel, drifts out of support. GKE clusters follow their release channel. Paying for extended support should be a decision, not an accident.
- Remove APIs before the provider removes them. Each minor version can drop APIs that manifests, Helm charts and controllers still call. Scan for them before every upgrade, because a forced upgrade on AKS will not stop for them.
- Make sure an upgrade can finish. Test that Pod Disruption Budgets let nodes drain, and confirm quota, surge capacity and subnet addresses for the replacement nodes.
- Upgrade in your own window. Upgrade staging first, then production, inside an agreed maintenance window. EKS lets an in-place upgrade be rolled back within seven days, an option a forced upgrade never offers.
- Keep the version in code. Pin cluster versions in Terraform or OpenTofu, so every upgrade is reviewed, repeatable across environments and recorded.
Where SoftwareBiz fits. Our managed cloud and DevOps service runs Kubernetes on EKS, AKS and GKE: cluster setup, upgrades, scaling and cost control, with environments defined in Terraform or OpenTofu, monitoring and alerting, and patching in agreed maintenance windows. Upgrades are tested in staging and scheduled ahead of each support date, so no cluster reaches a forced upgrade or an extended support bill by accident. Your cloud accounts stay in your name, with role-based access you can review or revoke at any time. Where virtual machines are leaving VMware, our cloud migration and modernization team plans the move, and our FinOps and cost optimization team finds extended support fees and idle clusters on the invoice.
What to do this month
Run the inventory first. EKS clusters still on 1.31 have until 26 November before their control planes are upgraded for them, and clusters on 1.34 start paying extended support on 2 December. AKS clusters on 1.34 leave community support this month, and Microsoft’s calendar lists them for a platform-initiated upgrade from June 2027. Extended support fees are one line in a cloud budget that is already rising in 2027, and our guide to protecting your 2027 cloud budget covers the rest.
