KubeVirt’s Next Decade: Hardware Access, Confidential VMs and Pluggable Hypervisors

KubeVirt’s next phase looks less like a project focused on placing existing virtual machines in Kubernetes and more like an effort to make Kubernetes a common control plane for specialized hardware, confidential workloads and multiple virtualization technologies. That’s a meaningful shift in scope, while the original job—running VMs alongside containers—remains the foundation.
Red Hat started KubeVirt in 2016 to let organizations operate legacy VM workloads on the same platform as modern applications. Since the project reached v1.0, other systems have built on its APIs in different ways. Open Sovereign AI Cloud (OSAC) uses KubeVirt for VM, cluster and bare-metal services. Cluster API Provider KubeVirt (CAPK) treats KubeVirt VMs as Kubernetes nodes. Flight Control runs virt-launcher at the edge without deploying the full KubeVirt control plane.
Those examples suggest that KubeVirt’s value increasingly depends on how well it serves as infrastructure for other tools, not only as a way to run a VM. The project is a CNCF incubating project, with contributors from Nvidia, Microsoft, AMD, Intel, SUSE and other organizations. Its community is working toward CNCF graduation, while release trackers for v1.8, v1.9 and v1.10 show a growing set of proposed work. The queue grew from a small set of items in v1.6 to more than 100 by v1.10; v1.9 included dozens of new proposals.
Growth makes project process part of the technical story. KubeVirt Enhancement Proposals (VEPs) give contributors a shared design and review path before code lands. Special interest groups (SIGs) take ownership of areas, helping the project weigh new capabilities against the long-term stability expected from its core.
Hardware allocation moves closer to Kubernetes
One of the most practical changes is how VMs request devices. Kubernetes Dynamic Resource Allocation (DRA) already supports GPU claims for VMs, with the feature at beta in v1.9. The same model is being extended to generic host devices, network interfaces such as SR-IOV, and CPUs. The intended result is a VM that can request a coordinated set of resources—say, a GPU, a network device and CPU capacity—through Kubernetes allocation mechanisms.
There’s also work on creating those claims from the VM specification and vendor policy. That could spare users from manually defining ResourceClaim objects for each workload. The architectural direction is consistent: when Kubernetes already provides a mechanism, KubeVirt aims to use it instead of maintaining a virtualization-specific path. Past examples cited by contributors include Swap, ImageVolumes and volume populators.
Hardware support is expanding alongside allocation. KubeVirt contributors, including Nvidia engineers, have worked to enable the project on newer Nvidia platforms such as Grace Hopper, Grace Blackwell and Vera Rubin. The effort began in the Accelerated Infrastructure Enablement working group and is moving upstream. For operators building GPU-heavy infrastructure, the important question is whether these capabilities mature into support that can be managed through the same workflows as other Kubernetes resources.
Confidential computing needs a full trust path
Confidential VMs are another priority, particularly for sensitive applications and AI workloads running on shared infrastructure. Hardware-backed trusted execution environments (TEEs) can protect workloads from privileged host software within the platform’s threat model. Remote attestation adds a check before the workload owner releases secrets: it provides evidence about the environment where the VM will run.
KubeVirt is broadening support for confidential VMs and exploring connections to attestation and secret-release services, including Trustee from the Confidential Containers project. Fan Zhang, co-lead of KubeVirt’s Accelerated Infrastructure Enablement and Confidential Compute working groups, argues that AI workloads need protection across CPUs, accelerators and the links between them. In Zhang’s view, workload owners should be able to verify that environment through attestation before releasing sensitive data or model keys. Kubernetes workflows could then bring together confidential VMs, protected accelerators and verification.
Support varies by platform and maturity. In v1.9, the feature gate for AMD SEV and SEV-SNP advanced to beta; Intel Trust Domain Extensions (TDX) remained alpha; and IBM Secure Execution reached general availability. Arm Confidential Compute Architecture (CCA) support is still being developed. These distinctions matter: the broader roadmap spans several hardware families, but capabilities are not yet at the same stage across them.
Contributors are also considering more consistent discovery of TEE capabilities across cluster nodes, so Kubernetes can place confidential VMs on suitable machines. Other proposals address initialization data for attestation, evidence handling and connections to external verification and secret-release services such as Trustee. Capability discovery can help determine where a VM is eligible to run; attestation helps an owner decide whether to trust the environment it receives.
Extending those protections to accelerator assignment and communication between CPUs and accelerators will take coordination across hardware, firmware, hypervisors, guest software and attestation services. These proposals are at different points in discussion and implementation, and their arrival in releases will depend on community review, integration and testing.
Extensions can keep the core stable
KubeVirt’s core has a deliberately high bar for new features. Once a capability becomes part of the central project, maintainers take on the responsibility to support it over time and protect clusters that already depend on KubeVirt. That approach can make core development slower, but it also gives users a clearer expectation of long-term behavior.
The project is developing a way to add complete features outside the core while retaining KubeVirt API guarantees. Structured plugins, which entered alpha in v1.9, are cluster-scoped objects that can add domain hooks, node hooks and admission behavior. Domain changes can use a CEL expression or, when external state is needed, a sidecar. Cluster administrators can inspect and control the plugins.
This offers a route for capabilities that are useful to particular users or vendors but aren’t ready to become universal KubeVirt features. The design could expand customization while leaving maintainers responsible for a smaller core surface.
Alternative hypervisors are on the roadmap
Today, KubeVirt relies on libvirt, QEMU and KVM by default. That foundation remains central, but the project has begun work to separate KubeVirt from any single virtualization stack. Version 1.8 introduced an in-tree Hypervisor Abstraction Layer and Microsoft’s MSHV backend as alpha. The longer-term objective is to let other virtual machine monitors, including Cloud Hypervisor, OpenVMM and Firecracker, operate as external backends without requiring a KubeVirt fork.
Microsoft engineer Harshit Gupta describes the motivation as extending Kubernetes orchestration, storage and networking to VMs running on alternative stacks. Gupta’s proposal builds on existing work to support other hypervisors and aims to decouple KubeVirt from its underlying virtualization layer. The project’s KubeVirt Enhancement Proposal #82 outlines the main pluggability effort.
That separation requires changes to how KubeVirt selects and manages a VM monitor and hypervisor, and implementation work is underway. If it succeeds, plugin-based backends could give users more choice and make it easier to test and maintain the different virtualization options. The useful measure will be whether teams can adopt those stacks without taking on the cost of maintaining their own fork.
From VM compatibility to a platform for new workloads
KubeVirt began with a clear migration use case: move existing VM workloads onto Kubernetes and manage them alongside containers. Its next decade will be shaped by a broader set of demands, including AI inference, edge systems, multi-tenant infrastructure and emerging processor architectures. Device claims, confidential execution, structured extensions and pluggable hypervisors point toward a platform that can serve those needs while retaining Kubernetes as its operating model.
The proposals are not guarantees of future release dates. Their practical value will depend on how contributors turn them into stable, interoperable features. For platform teams, the signal to watch is whether the expanding ecosystem can make hardware and virtualization choices easier to manage through Kubernetes without making the core harder to operate.
KubeCon + CloudNativeCon North America is scheduled for November 9–12 in Salt Lake City. Event details.
The post The next 10 years of KubeVirt appeared first on The New Stack.
===