Skip to main content

Changelog

Follow new updates and improvements to vCluster.

Platform v5.0 and vCluster v0.38 enable Private Customer Clouds

Give every customer a private, self-service cloud on hardware you already own. Platform v5.0 brings tenant isolation, customer administration, and infrastructure access into one model. Paired with vCluster v0.38, it connects the boundaries you set for each organization to the clusters its teams create and run.

Why v5.x

AI cloud providers want to build services on top of the capacity they already sell. That requires more than provisioning a cluster. Each customer needs a place to manage its people, machines, networks, and operating system choices within limits the provider controls.

Platform v5.0 introduces a new Tenant boundary that connects those pieces. vMetal turns the rack into inventory. The tenant turns that inventory into a customer's cloud. vCluster and Stacks turn that cloud into services, including Kubernetes, Slurm, Ray, inference, and more.

For AI cloud providers, that means organizing your offering around each customer. For enterprise platform teams, it means giving business units more autonomy within centrally defined policies. Platform admins decide what infrastructure is available. Tenant admins manage their organization's users, projects, and access to assigned infrastructure. Project admins manage access and settings within a project. Project users create clusters within those limits.

Platform v5.0 and vCluster v0.38 are limited availability releases, recommended only for new deployments. Work with vCluster Labs to evaluate the new capabilities.

Give each customer a tenant boundary

Platform v5.0 introduces the Tenant resource above projects. Each tenant represents a customer, business unit, or other organization, with its own users, projects, and infrastructure allowances.

That gives you a clear place to delegate. Tenant admins can see the machines and networks available to their organization and manage its teams and projects. Project admins organize work within a project, while platform admins retain control of the resources each tenant can use. Customize the UI per tenant with its own branding and settings.

Diagram titled "How Platform organizes tenants". There are five layers of roles listed along the left side. System admin installs platform, platform admin creates tenants, tenant admin manages infrastructre and projects, project admin manages project access, and a project user creates clusters. The diagram shows a large box for the vCluster Platform. Within that box are two smaller boxes side by side, one for Tenant acme and one for Tenant globex. Each tenant has assigned infrastructure for machines and networks. Each tenant contains two projects. One project of each tenant has two clusters and the other project has a single cluster.
Give each customer its own organizational boundary. Tenant admins manage the organization's infrastructure and projects. Project admins manage individual projects. Project users create clusters.

Give each customer its own organizational boundary. Tenant admins manage the organization's infrastructure and projects. Project admins manage individual projects. Project users create clusters.

You can curate the cluster templates, Stacks, node types, images, and integrations available to each tenant. Teams get a self-service catalog shaped around their needs, and administrators keep the policy decisions that govern it.

Explore tenant administration in Platform v5.0.

Bring NICo capacity into the same workflows

With Platform v5.0, vMetal adds NVIDIA Infrastructure Controller (NICo) as a provisioning driver alongside Metal3. Offer NICo-backed capacity through administrator-defined node types, with machines reserved for a tenant or allocated on demand.

Tenants, network environments, and Platform identity connect that capacity to the same administration model. NICo remains Alpha and may include breaking changes in future releases.

Turn infrastructure into a customer offering

Different customers need different capacity. Reserve machines for a tenant's contract, or offer on-demand allocation from a common capacity pool. vMetal, the infrastructure orchestrator beneath vCluster, brings those choices into one machine inventory across Metal3 and NICo. It orchestrates the provisioning drivers, while Platform connects that capacity to tenant access and self-service.

Start, stop, and reboot bare metal machines from Platform, as you already could with KubeVirt VMs.

Networking and operating system images become part of the offering. Administrators choose the network environments a tenant can use, including networking backed by Netris or NICo. Metal3-backed infrastructure can also use Aviz-backed network environments. Tenants can also bring their own OS images when that capability is enabled, alongside images the provider makes available.

Separate the service from its infrastructure

Node pools are a new Platform resource attached to a cluster. The operator's cluster template defines the managed service. The tenant's node pools define the compute and network choices that support it, using infrastructure the tenant owns or has been assigned.

Set a fixed node count or let a dynamic pool respond to workload demand. The service definition and its capacity can be managed separately, following the node-pool model familiar from hyperscaler services.

Connect your existing inventory with NetBox

The NetBox integration imports selected device records into Platform's machine inventory for Metal3 provisioning. That connects the hardware records your team maintains in NetBox to the capacity you offer through Platform.

See Platform node pools and machine allocation for the workflows and requirements.

Connect capacity to AI workloads

Infrastructure is the foundation of the service. The applications running on it are what teams consume.

Build on Stacks such as NVIDIA Dynamo to give teams a repeatable AI-serving environment. Stacks bring related applications together as one dependency-aware deployment, so teams can start from a defined application offering.

Together, tenant access, infrastructure allocation, and Stacks connect the path from available capacity to an environment teams can use.

Explore Stacks.

Run the clusters behind the offering

vCluster v0.38 supports the new Platform-managed node pools attached to a cluster. It also improves how node claims respond to configuration changes and missing machines, and includes local development, networking, and recovery improvements.

Some supporting improvements also shipped in v0.37 patch releases. What's new in vCluster v0.38 explains the capabilities and requirements. The release notes identify the exact changes and patch-release overlap.

Plan your evaluation

If you are expanding customer self-service, delegating administration across business units, or bringing more infrastructure into a common offering, evaluate how Platform v5.0's tenant and capacity model fits your plans. The goal is a hyperscaler-quality customer experience, with self-service clusters and tenant isolation on infrastructure you control.

Platform v5.0 and vCluster v0.38 are limited availability releases, recommended only for new deployments. If you need the new capabilities on an existing installation, contact vCluster Labs to discuss your use case and plan the upgrade.

Use the Platform v5.0 release notes and vCluster v0.38 release notes to review how the new releases differ from v4.12 and v0.37, including removed features, breaking changes, and known issues.

Coming from Platform v4.12? Platform v5.0 removes shared secrets, cost control, the project-level Argo CD integration, and the bundled infrastructure apps, and deprecates project namespaces (spaces). Apps and the example Helm, Bash, and Manifests apps remain available. See the removed features and breaking changes in the release notes.


To our v4.x customers

There's no rush to move to 5.x. Platform v5.0 and vCluster v0.38 are recommended only for new deployments. For upgrade guidance, contact vCluster Labs. Platform v4.12 and vCluster v0.37 form the LTS release set and are supported as a pair. vMetal in Platform v4.12 follows the same lifecycle.

LTS provides 12 months of active support, followed by six months of critical security updates. During active support, you can request backports of fixes, including support for newer Kubernetes versions. Upgrading either product past its LTS version takes that combination off LTS.

See the LTS policy and support dates, and work with vCluster Labs when you are ready to evaluate v5.x.

PlatformvCluster

vCluster Platform v4.12 + vCluster v0.37 - Launch and scale complete tenant environments

vCluster Platform 4.12 and vCluster 0.37 move beyond tenant clusters to the complete environments around them. Define reusable environments with Stacks, run them on broader infrastructure, protect them with stronger recovery, and operate them through GitOps and fleet observability.


Deploy complete application environments with Stacks

With Stacks, platform teams can turn a collection of application components into a reusable, self-service building block. Define an environment once, expose the parameters its users should control, and deploy it consistently to a tenant cluster or control plane cluster. Teams can install a Stack from those available to them, add it during tenant-cluster creation, or declare it through vcluster.yaml so its lifecycle follows the cluster.

Stacks coordinate more than installation. Applications without dependencies can deploy concurrently, while dependent tasks wait for health gates and can consume values captured from earlier tasks. Each application remains a long-lived installation with its own status and logs, while the Stack provides an aggregate view and keeps ownership and cleanup tied to the deployment. Teams retain one durable place to understand and manage the complete environment through day-2 operations.

The underlying StackTemplate and StackInstance custom resources define the reusable offering and each deployment. They can be made of any combination of AppInstance or ArgoCDApplication. This provides the managed application foundation, replacing the older Task and HelmRelease installation path.

Deploy Run:ai with a certified Stack

Platform 4.12 includes certified NVIDIA Run:ai Stacks, giving platform teams a ready-to-use path for deploying Run:ai as a complete tenant environment. The bundled templates coordinate registry setup, ingress, TLS, NVIDIA GPU components, the Run:ai control plane, tenant registration, and cluster agents in reusable, dependency-aware workflows. Teams can deploy these components consistently while retaining the status, logs, lifecycle management, and cleanup provided by Stacks.

For hard multitenancy at the Run:ai control-plane layer, the dedicated model gives each tenant an independent Run:ai control plane and GPU Operator lifecycle. For trusted soft-multitenant environments, the central model installs the shared host foundation once and reuses its Run:ai control plane and GPU Operator while keeping registration and runtime resources separate for each tenant. Both bundled tenant templates use shared control plane cluster nodes by default; node labels control scheduling, not runtime isolation. Add private nodes when node-level tenant isolation is required.


Argo CD and Fleet Observability now included in the Free tier

The rollout of Stacks expands Platform’s Argo CD-based application workflows. Alongside that work, the Argo CD integration and Fleet Observability, previously starting at the Dev tier, are now included in the Free tier. Anyone running vCluster Platform can register tenant clusters and control plane clusters with Argo CD and monitor them through the bundled Fleet Observability stack without a paid license. See the Argo CD integration and Fleet Observability docs for setup steps.


Protect environments with stronger snapshot recovery

Complete environments need dependable recovery as well as repeatable deployment. Before a restore begins, vCluster verifies that it can access the requested snapshot. Expired or invalid S3 credentials now fail before the control plane is restarted, so an unsuccessful recovery attempt does not create unnecessary downtime.

Teams can also better protect snapshots at rest. S3-backed snapshots support server-side encryption, including customer-managed KMS keys, while object-store credentials are stored in a Kubernetes Secret instead of being embedded in the control-plane Pod specification.

Scheduled snapshot coverage now includes Azure Blob Storage alongside existing S3-compatible paths. Platform can also configure and manage automatic snapshots for standalone tenant clusters, bringing scheduled protection into the same workflow used for other Platform-managed clusters. vCluster can convert embedded-etcd snapshots when restoring into another supported backing store, including deployed etcd or a supported external database target. Together, these changes make snapshots a more dependable and flexible recovery mechanism across supported datastores and storage providers.

Breaking Changes

App deployment APIs have changed

Platform removes the Task API, HelmRelease API, the legacy virtual cluster API group, and the related chartinfos, clusters/charts, and projects/charts discovery endpoints as part of the move to long-lived AppInstance resources. Automation that installs apps through Tasks or HelmReleases, or discovers charts through those endpoints, must migrate to AppInstance and the new app APIs.

Standalone snapshot credentials must be stored in the Platform project namespace

For standalone tenant clusters, snapshots.auto.storage.<provider>.credential.secretNamespace is no longer supported. Platform always resolves the Secret named by secretName from the same project namespace as the VirtualClusterInstance. Before upgrading, copy or recreate affected S3, OCI, or Azure credential Secrets in that namespace, then remove secretNamespace from the virtual cluster configuration or template. Pod-based tenant clusters are unaffected.

Before upgrading a standalone tenant cluster:

snapshots:
  auto:
    storage:
      type: s3 
      s3: 
        credential: 
          secretName: aws-cred 
          secretNamespace: backup-secrets 

For standalone tenant clusters in Platform 4.12:

snapshots:
  auto:
    storage:
      type: s3 
      s3: 
        credential: 
          secretName: aws-cred 

The aws-cred Secret must exist in the Platform project namespace containing the VirtualClusterInstance. Apply the same change to OCI or Azure credentials that set secretNamespace.


Kubernetes Version

This release uses Kubernetes 1.36 by default (v1.36.0).


Additional improvements

  • Restore CSI VolumeSnapshot synchronization. vCluster again syncs VolumeSnapshot and VolumeSnapshotContent resources to the host cluster and VolumeSnapshotClass resources into the tenant cluster. This corrects an unintentional removal in v0.36 that affected customers using CSI snapshot resources. The separate removal of snapshot-managed volume backup and restore is unchanged.

  • Run private-node workloads on more infrastructure. Platform adds KubeVirt CSI for stateful workloads, GPU and PCI passthrough, Netris-backed networking for Metal3, and Kube-OVN load balancing and pod routing.

  • Let Argo CD-managed tenant clusters sleep. Argo CD polling no longer resets the inactivity timer for onboarded tenant clusters, allowing sleep mode to suspend clusters that would otherwise appear continuously active.

  • Set properties per node pool. Static and dynamic node pools can override provider-level properties for the NodeClaims they create, giving platform teams more control over different pools without duplicating the provider configuration.

  • Meter GPU usage across more accelerator vendors. Platform now counts AMD, Intel—including Gaudi—AWS Neuron, and Google TPU accelerators alongside NVIDIA and reports usage by vendor and resource type. Existing environments using non-NVIDIA accelerators may report a higher billable GPU total after upgrading.

vCluster

vCluster Platform v4.11 + vCluster v0.36 - Operating tenant clusters at scale

Release Date: July 21, 2026

vCluster Platform v4.11 and vCluster v0.36 focus on day-2 operations for tenant clusters: built-in monitoring for platform admins, consistent configuration for private-node fleets, and clearer recovery guidance for backing-store incidents.


Fleet Observability

On shared infrastructure, platform admins need one fleet-wide view, while each tenant needs a safe, isolated slice of the same data. Building this by hand often means per-tenant Prometheus deployments and brittle access rules. vCluster Platform v4.11 introduces Fleet Observability, giving platform admins one fleet view from which they can drill into tenant and control plane clusters that need attention. This initial release focuses on metrics. Logs and traces are not part of the Fleet Observability pipeline.

When admins deploy the optional bundled Grafana application template and configure the Observability Connector, Platform embeds fleet, cluster, and GPU dashboards. The connector can use the bundled Prometheus backend or another metrics backend that accepts OTLP metrics and provides a Prometheus-compatible HTTP query API.

OpenTelemetry collectors write metrics to a single backend through an authenticated gateway that stamps trusted Platform labels. Dashboard and API requests pass through the vCluster-provided query proxy that scopes PromQL to the control plane clusters or tenant clusters the authenticated identity can read. Together, authenticated labels and query rewriting enforce tenant-isolated access over a shared metrics backend. Platform dashboard views are available to platform admins, while direct API clients can use scoped metrics-reader access keys.

Ready-to-deploy metrics stack

For teams that want a ready starting point, Platform includes Argo CD application templates for the metrics stack. The bundled Prometheus template accepts OTLP metrics. The Grafana template integrates with Platform SSO and provides fleet, cluster, and GPU dashboards. OpenTelemetry collector templates collect workload, Kubernetes, node, and optional GPU metrics. Admins create applications from these templates for the backend, Grafana, and the clusters they want to monitor.

GPU health and operations

For AI and GPU fleets, the same metrics path surfaces GPU-specific signals alongside tenant and control plane cluster health. Fleet Observability includes a GPU dashboard and collector scrape paths for exporters such as DCGM.

Separate GPU application templates help admins deploy NVIDIA GPU Operator, NVSentinel, and NVIDIA Fleet Intelligence Agent. These applications operate independently of Fleet Observability, but their metrics can feed the same Platform-centered operating view.

As more teams and clusters share the platform, Fleet Observability gives admins a clearer starting point for everyday operations. Admins can see the fleet, find the cluster that needs attention, and follow its metrics without leaving Platform.

Group private nodes and keep them configured

vCluster Platform v4.11 introduces NodeProfile as a way to group private nodes and maintain their configurations in common. Platform admins define named profiles with the labels, annotations, and taints that should apply to any node referencing them, and the platform keeps those settings in sync continuously, not just at join time. Editing a profile pushes the change out to every node that already uses it, without rejoining or replacing the node. Profiles can also apply startup taints that hold a new node out of scheduling until setup finishes, then clear automatically once it's ready.

Profiles apply the same way across manual joins and auto-node pools, including both static and dynamic pools. A node gets the right configuration no matter how it was provisioned, without teams hand-rolling their own join scripts or one-off per-pool settings. Application teams can target the right capacity with familiar Kubernetes scheduling controls, while platform teams manage the configuration from a single place.

Project admins can curate which profiles a project is allowed to use and set a project-level default for nodes that don't select one explicitly. A tenant cluster's own default (privateNodes.defaultProfile) always takes precedence over the project default, so platform teams get a safe baseline per project, while application teams can still opt into specialized profiles where they're allowed. Because the assignment is enforced through a protected node label that only the platform can set, a node can't grant itself a different profile after it joins.

Node profiles also make private-node operations easier to reason about as fleets grow. A node no longer joins as an undifferentiated machine that needs follow-up patching. Instead, it joins with its labels, annotations, and taints already applied, and stays in sync with its profile for as long as it's part of the tenant cluster.


Breaking Change

  • CSI volume snapshot backup and restore has been removed. vCluster v0.35 deprecated CSI volume snapshot backup/restore for persistent volumes. In vCluster v0.36 and Platform v4.11, the volume snapshot implementation, CSI snapshot controller deployment, VolumeSnapshot syncers, and related API fields have been removed. Existing sync.toHost.volumeSnapshots, sync.toHost.volumeSnapshotContents, and sync.fromHost.volumeSnapshotClasses configuration is retained only as a no-op so older configs can still parse. Customers that previously relied on backup (--include-volumes) or restore (--restore-volumes) should move persistent data protection to application-level or storage-provider-native workflows, such as Velero-based backups, and use vCluster snapshots for vCluster state.

  • ArgoCD Application names now include a Platform instance hash. To avoid collisions when two or more Platform instances deploy same-named tenant clusters with the same applications into a shared ArgoCD, the computed ArgoCD Application name now includes the Platform instance ID as a hash suffix. Platform migrates existing Applications automatically without deleting managed workloads. Teams using the Argo CD integration should confirm their Applications reconcile cleanly after upgrading.


Deprecation

  • Moving tenant clusters between projects by API is deprecated. The API for moving virtual clusters between projects is being deprecated. Customers should plan to create tenant clusters directly in the intended project and avoid building new automation around cross-project moves.

  • Legacy token forwarding boolean is deprecated. Platform now uses an explicit token forwarding mode for tenant clusters. Existing forwardToken configuration remains compatible, but new configuration should use the forwardTokenMode enum instead of the legacy boolean.


Kubernetes Version

This release uses Kubernetes 1.36 by default (v1.36.0).


Additional improvements

  • Tenant cluster troubleshooting gets more complete. Platform includes updates to the tenant cluster status experience, host-cluster drilldowns, vCluster logs in Platform, VPN diagnostics, and Gateway API inspection paths so admins have more context when they investigate a cluster.

  • Template-backed tenant cluster creation is easier to review. The Platform UI now shows template summaries and template parameters during creation, making template-backed clusters easier to understand before users create them.

  • Standalone clusters can set an explicit advertise address. vCluster now supports controlPlane.standalone.advertiseAddress, which is useful when the address advertised to peers should differ from the address vCluster would auto-detect.

  • Task execution moves in-process (runner removed). Platform no longer uses a standalone runner. App and template tasks now run in-process in the Control Plane Cluster (bounded by LOFT_MAX_CONCURRENT_TASKS). Task logs are stored per-task in Kubernetes secrets rather than streamed through a runner. There is no user-facing API change, but self-hosted setups that provisioned dedicated runner peers (for example, Tailscale runner peers, runner access keys) no longer need them. Review any automation that assumed a separate runner.

  • Node join tokens now last 24 hours. Bootstrap tokens used by autoscaling worker-node user data and tenant cluster join scripts now expire after 24 hours instead of 1 hour. This makes provisioning and join flows more resilient in the case of token expiry.


For a list of additional fixes and smaller changes, refer to the vCluster release notes, the vCluster Pro release notes, and the vCluster Platform release notes. For detailed documentation and migration guides, visit vcluster.com/docs.

PlatformvCluster

Earlier updates