Virtual machines are not legacy – design your platform for both
Every few years someone declares the virtual machine finished. New research from the Linux Foundation and CNCF suggests otherwise: in a survey of 492 cloud native practitioners, virtual machines made up an average of 53% of production applications, and 71% of those who answered expected their VM share to stay in the same range over the next year.
That matches what many platform teams already know. The interesting question is not whether VMs survive, but how you design a platform that treats them as a first-class workload instead of an embarrassment.
What the survey says – and what it doesn't
The headline figures, from Virtual Machines, Agentic AI and the Mix of Workloads (September 2026):
- VMs are an average of 53% of production applications; 64% in organisations with more than 10,000 employees and 45% in those with fewer than 500.
- 61% describe their VM and container teams as collaborative or fully integrated; 14% say they work in silos.
- Among organisations growing their VM footprint, 47% cite security and isolation and 34% cite AI and agentic workloads.
- Only 8% of respondents who run VMs manage them on Kubernetes with KubeVirt.
Two caveats before anyone builds a strategy on this. The research was sponsored by CNCF and Broadcom, which sells a VM platform, and the foreword is written by a Broadcom engineering director. The respondents are also cloud native practitioners answering for their own organisations, not a random sample of the market. Neither makes the data wrong, but both are reasons to read it as a signal rather than a benchmark.
Why VMs stay
The reasons are mostly unglamorous. Licensed software is often supported only on a VM. Some systems need a specific kernel, a Windows Server or a vendor appliance. Rewriting a stable application just to containerise it rarely pays back – the business sees cost and risk, not value.
The newer reason is isolation. Containers share the host kernel. When a workload runs code nobody reviewed – an AI agent executing tools, a customer's plugin, a build job – a hardware-virtualised boundary is a stronger fence. That is why the survey's planned growth is in MicroVM sandboxing (20% today, 30% planned) and confidential computing (12% to 26%), while containers inside VMs stay roughly flat at around 40%. These are stated plans, not deployments, but the direction is clear.
Three ways to run both – and their costs
Separate platforms, one way of working. VMs on a hypervisor or an IaaS cloud, containers on Kubernetes, both described in Terraform or OpenTofu and changed only through pull requests. This is the most common pattern and often the right one. The risk is two teams, two backup regimes and two security baselines drifting apart – which is exactly what the 14% "silos" figure warns about.
VMs as Kubernetes resources. KubeVirt lets a VM be declared, scheduled and deployed like a pod, managed by the same GitOps tooling. That is attractive when VMs are few and container skills are strong. But it is not free: live migration needs shared storage that supports it, VM networking on Kubernetes adds complexity, and backup and restore of VM disks must be designed and tested, not assumed. The 8% adoption figure suggests most teams have not yet found that trade-off worthwhile.
Stronger sandboxes for untrusted code. MicroVM runtimes such as Kata Containers give a pod its own lightweight VM. They fit agent sandboxes and multi-tenant build workloads well, at the price of some start-up overhead and node requirements to plan for.
Our default recommendation for most organisations: keep VMs and containers on separate runtimes, but on one delivery path. One Git repository structure, one review process, one set of policies for identity, network and backup – regardless of whether the unit of deployment is a VM image or a container image. Move individual VMs into Kubernetes only when there is a concrete reason, such as tight coupling to containerised services.
The sovereignty angle
For organisations moving towards European infrastructure, this is good news. OpenStack-based European providers offer virtual machines and managed Kubernetes on the same underlying platform, in the same jurisdiction. You do not need to finish containerising everything before you can leave a non-EU cloud. Plan the move per workload: what can be rebuilt as containers, what moves as a VM, and what should simply be retired.
Questions to ask about your own platform
- Do your VMs and your clusters go through the same review and deployment path, or does one of them still change by hand?
- Which workloads run code you did not write or review – and what isolation do they have today?
- If you had to restore a VM-based system and a Kubernetes-based system tomorrow, would both restores have been tested?
If any answer is "not sure", that is where the platform work starts – not with the next migration.
Sources
- CNCF blog: New research: Virtual machines remain central to cloud native infrastructure as AI workloads evolve (9 November 2026)
- Linux Foundation Research: Virtual Machines, Agentic AI and the Mix of Workloads
- KubeVirt project: kubevirt.io
- Kata Containers project: katacontainers.io