platform-eng
Kubernetes Hardware Isolation for AI Workloads
Dedicate physical GPU servers to Kubernetes tenants. vMetal provisions bare metal, and vCluster uses those servers as Private Nodes with separate tenant control planes.
Trusted by the fastest-growing AI cloud providers
Problem
Hardware Isolation Requires Dedicated Physical Workers
Namespace boundaries do not dedicate physical worker capacity to a tenant cluster.
Namespaces Share Cluster Components
Namespace-only designs share the API server and can also share worker nodes between tenants.
Separate Management Stacks Add Work
Building and operating a separate Kubernetes management stack for every tenant increases infrastructure work.
Shared Worker Pools Create Contention
Shared worker nodes expose workloads to cross-tenant contention for CPU, GPU, memory, network, and storage resources.
Solution
Kubernetes Hardware Isolation With Private Nodes
vMetal provisions dedicated physical servers and vCluster attaches them as tenant Private Nodes. Separate tenant control planes reduce control-plane server duplication; network, storage and runtime controls complete the deployment design. Lintasarta runs 170+ tenant clusters in production.
Built for Kubernetes Hardware Isolation
Kubernetes hardware isolation combines dedicated physical Private Nodes with a separate tenant control plane and optional runtime and network controls.
Hardware Isolation
Dedicated Physical Private Nodes
Use Private Nodes backed by dedicated physical servers when tenants require exclusive worker hardware. vMetal provisions the bare metal machines, while each tenant configures its own CNI and storage. VM-backed workers need separate physical-host allocation guarantees.
- Dedicated worker nodes per tenant
- Tenant-owned CNI and storage configuration
- Dedicated bare metal for physical exclusivity
Workload Security
Runtime Isolation With vNode
vNode uses Linux user namespaces and seccomp to restrict workload privileges and system calls. It adds runtime hardening for physical worker isolation without a guest kernel or hypervisor, complementing dedicated workers rather than allocating GPUs.
- Linux user namespaces and seccomp
- No guest kernel or hypervisor
- Helps limit workload breakout
Network Isolation
Optional Hardware Network Isolation
Tenant-owned CNI configuration and supported Netris integrations provide separate network controls for physical worker isolation. Ethernet segmentation can use VLANs, VRFs and ACLs; InfiniBand isolation uses fabric partitions rather than Ethernet constructs.
- Tenant-owned worker network configuration
- Netris integration for supported fabrics
- Explicit segmentation and access policies
Control Plane
Separate Control Plane Per Tenant
Every tenant cluster receives its own virtualized API server and RBAC boundary on the control plane cluster.
- Independent tenant API and data store
- Tenant-scoped scheduling on Private Nodes
- No dedicated control-plane servers per tenant
Compliance
Air-Gapped Platform Deployment
vCluster Platform supports air-gapped deployments and FIPS features on supported plans for controlled infrastructure environments.
- Air-gapped deployment supported
- FIPS features on supported plans
- Regulated deployment patterns
Why vCluster
This isn’t a side project. Behind every vCluster deployment is 5+ years of deep K8s engineering, security hardening, and battle-tested infrastructure work at massive scale.
100K+
GPUs Powered
50+
GPU Clouds & F500s
<45
Days to Launch
30K
GitHub Stars
Get Started in 3 Steps
1
Schedule a Demo
Talk to our team about your stack
2
Deploy vCluster
Deploy vCluster on your infra in minutes
3
Onboard Your Tenants
Go live with a hyperscaler-grade tenant experience in days
FAQs
What does Kubernetes hardware isolation mean?
Private Nodes are worker nodes joined to one tenant cluster, with that tenant's own CNI and storage configuration. They can be physical machines or supported VMs. Use dedicated bare metal when the service requires exclusive physical servers. vMetal can provision the underlying machines, while platform policies define which capacity a tenant may request and how it is released.
How do Private Nodes provide hardware isolation?
vCluster gives each tenant an independent Kubernetes API, data store and RBAC scope. Private Nodes dedicate worker capacity to that tenant, including its own CNI and storage configuration. vMetal orchestrates the machines underneath. This separates tenant administration and worker placement; shared infrastructure, application permissions and network policies still require an explicit security design.
Does every tenant need separate physical control-plane servers?
Namespaces divide resources within a Kubernetes cluster, but share its API and cluster-wide administrative scope. RBAC, quotas and network policies remain useful controls. vCluster adds a separate tenant API and data store, with Private Nodes dedicating production workers. This reduces dependencies between tenant environments while platform administrators still manage the underlying infrastructure and its failure domains.
How does dedicated hardware affect GPU performance?
Dedicated Private Nodes remove cross-tenant workload placement from those workers. Bare metal avoids a hypervisor layer, but this is not a throughput or latency guarantee. GPU drivers, workload configuration, network fabrics and storage can still influence performance. vNode adds runtime hardening using the Linux kernel; benchmark the selected stack against the workload and service levels you intend to offer.
What deployment controls support regulated environments?
Supported plans include air-gapped deployment and FIPS features. Private Nodes, tenant-scoped access and optional network segmentation provide controls you can include in a regulated deployment. Select supported images and configurations, and validate the full environment against your requirements. These capabilities do not establish certification of the customer's service, application, hardware or operational procedures.
What evidence supports vCluster for GPU infrastructure at scale?
vCluster Labs software powers 100K+ GPUs and 1M+ CPUs, serving 50+ GPU clouds & Fortune 500s combined. In production, Lintasarta operates 170+ tenant clusters, while Boost Run completed its managed Kubernetes launch in under 45 days from the decision. These named deployments provide scale and launch examples for teams evaluating the platform for their own infrastructure.
Achieve Kubernetes Hardware Isolation Today
See how Private Nodes provide Kubernetes hardware isolation for production tenant clusters.
