VM API and KubeVM Service, from the vCenter side

Broadcom's VM API is now a CNCF Sandbox project and KubeVM Service is its vSphere provider on VCF. The VMs stay native ESX VMs. What changes is who asks for them and how, and that is a governance job for the infrastructure team.

On 9 November, at KubeCon North America, Broadcom announced that VM API had been accepted into the CNCF Sandbox. Most of what I read that week was aimed at platform engineers. I'm more interested in what it means for the people running the vCenter those VMs end up in.

What it is

Broadcom's press release puts it plainly: VM API "enables virtual machines to be provisioned and managed through Kubernetes APIs" across different infrastructure providers. In practice that means you declare a VM the way Kubernetes declares everything else, as a YAML manifest against a set of custom resources, and let a controller make it real.

Broadcom's engineering blog calls it "a standardized, provider-neutral API contract for virtual workloads". Behind the contract sit providers, plugins that turn the request into a VM on a particular platform. If that sounds like Cluster API, that's deliberate. Broadcom says it's modelled on it, and that it builds on VM Operator, which has been inside vSphere for more than six years.

The same blog lists providers for vSphere (KubeVM Service), Proxmox and KubeVirt on premises, and AWS and Google Cloud in public cloud, and says the list isn't complete.

Sandbox is the CNCF's entry stage. It means the project has a neutral home and is open to outside contributors. It doesn't mean it's finished, and it doesn't mean every provider is equally mature.

It came alongside two other announcements worth knowing about: Broadcom is co-chairing a new Workload Conformance Working Group in the Kubernetes community, and more than 20 partner solutions were newly validated for vSphere Kubernetes Service.

Where VCF comes in

KubeVM Service is the commercial vSphere provider. On VCF, a VM requested through VM API becomes a native ESX VM. That's the bit that matters to an infrastructure team:

  • The VM is still visible in vCenter and VCF Operations.
  • vSphere HA, DRS and vMotion still apply. Broadcom's phrase is that hardware failures are abstracted away from the Kubernetes layer.
  • Your existing backup tooling still sees it.
  • It can be switched on selectively in an existing VCF deployment.

The VM itself is business as usual. How it gets requested is what's different.

What that changes for the VI team

If developers start requesting VMs from a Git repository rather than a ticket, several things that used to happen quietly at the service desk need a home somewhere else.

  • Quota and placement. Who decides how much a team can request, and on which clusters? That needs to exist before the first manifest, not after the first capacity alert.
  • Naming, tags and IP addressing. If your backup policy, monitoring or chargeback keys off a tag, the manifest has to carry it. If IPs come from an IPAM request today, decide how that works when nobody raises a ticket.
  • Backup membership. Visible to backup tooling isn't the same as protected by it. Check whether new VMs land in a protected group automatically or whether someone has to add them. If someone has to add them by hand, assume some will get missed.
  • Templates and images. Somebody still owns the golden image, its patch level and the hardening, and that doesn't change because the request came from Git.
  • Change records. The Git history becomes the change record. Decide whether that satisfies your change process or whether it needs to feed something else.
  • Decommissioning. Deleting a manifest deletes a VM. Make sure that's what everyone expects, and that anything with data on it has a retention step first.

We had most of these conversations when self-service portals came in. VM API just makes it likely you'll be having them again.

What I'd do now

  1. Write down how a VM gets requested, named, tagged, addressed, backed up and retired today, step by step. That's the list a manifest-driven process has to reproduce.
  2. Agree a tag schema with the backup, monitoring and finance owners before anybody pilots this.
  3. Pilot it in a non-production namespace with a hard quota, and watch what the requests actually look like.
  4. Decide who approves changes to the manifests themselves, because that's where control sits now.

What isn't settled yet

Broadcom's press release doesn't mention KubeVM Service at all, and the engineering blog doesn't give version numbers or availability dates for it. Neither says whether it applies to vSphere Foundation or standalone vSphere. The blog ties it to VCF. It also says features and outcomes "will depend on the underlying provider", which is a polite way of saying a manifest that works on one provider won't necessarily behave the same on another.

If you run VVF or vSphere and this interests you, that's the first question to put to your account team.

Related: The VKS policy bundle, on admission control, which is where most of these governance rules end up being enforced.

Sources

Checked on 15 November 2026. Views are my own.

More on VCF

This is one of the VCF pieces here. The rest are grouped by problem rather than by date on the VCF index, which is the fastest way to find the one you actually need.

The free planning tools run in your browser, with nothing sent anywhere. They are all on the tools page, from Estate Readiness Triage to the Restore Evidence Check.

Read next

Write down what you could not find

A design document asserts things its format never earned. The fix is structural: a numbered section for what you could not establish, and attribution by source type in the body rather than only in the references.