ai-cloud
GPU Cluster Isolation for AI Clouds
Separate GPU tenants beyond namespace boundaries. vMetal supplies machines, and vCluster combines independent tenant control planes with dedicated Private Nodes and optional network isolation.
Trusted by the fastest-growing AI cloud providers
Problem
Why Generic Isolation Fails at Scale
GPU tenant isolation needs clear boundaries at the control plane, worker node, and network layers.
Namespace Isolation Is Too Weak
Namespace-only designs share control-plane and node-level components between tenants.
Separate Clusters Are Too Expensive
Running a separate management stack for every tenant increases infrastructure and operational work.
No Control Plane Boundaries
A shared API server and RBAC model cannot provide the same tenant control-plane boundary as a dedicated tenant cluster.
Solution
Layered Tenant Isolation From Control Plane to Kernel
vCluster separates tenant APIs and data stores, and Private Nodes dedicate workers. vMetal manages infrastructure; optional vNode runtime hardening and Netris network automation extend the deployment's tenant boundaries. Lintasarta runs 170+ tenant clusters in production.
Full-Stack GPU Cluster Tenant Isolation
GPU cluster isolation spans the tenant control plane, dedicated Private Nodes, optional runtime isolation, and optional hardware network isolation.
Worker Isolation
Private Nodes Per Tenant Cluster
Private Nodes dedicate worker capacity to one tenant cluster, with its own CNI and storage configuration. For GPU cluster boundaries, select physical servers or VMs according to the required infrastructure boundary.
- Dedicated worker nodes per tenant
- Tenant-owned CNI and storage configuration
- Physical servers or supported VMs
Control Plane
Virtualized Control Planes Per Tenant
Each tenant receives its own virtualized Kubernetes control plane, API server, and RBAC boundary.
- Independent tenant API and data store
- Tenant-scoped scheduling on Private Nodes
- No dedicated control-plane servers per tenant
Workload Security
Kernel-Native Workload Isolation
vNode uses Linux user namespaces and seccomp to restrict workload privileges and system calls. It adds runtime hardening for GPU cluster boundaries 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
Hardware-Enforced Network Boundaries
Tenant-owned CNI configuration and supported Netris integrations provide separate network controls for GPU cluster boundaries. 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
Compliance
Air-Gapped and FIPS Deployments
vCluster Platform supports air-gapped deployments and FIPS features on supported plans for regulated deployment patterns.
- Air-gapped deployment support
- FIPS features on supported plans
- Sovereign and regulated environments
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 makes vCluster's GPU cluster isolation stronger than Kubernetes namespaces?
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 vCluster Platform handle GPU workload isolation between tenants?
vNode uses Linux user namespaces and seccomp to restrict workload privileges and system calls, helping prevent container breakout. It uses the underlying Linux kernel rather than a separate guest kernel. Add it as runtime hardening alongside dedicated Private Nodes, patching and least-privilege policies. Validate kernel, container runtime, GPU access and workload compatibility; it is not a replacement for GPU allocation or application authorization.
Can vCluster support compliance requirements for tenant isolation on GPU infrastructure?
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 cluster types can run as isolated tenant environments on vCluster Platform?
Offer Kubernetes tenant environments and compatible workload stacks through templates you configure and support. Private Nodes dedicate production workers, while vMetal supplies underlying infrastructure. The native Certified Stacks catalog documents NVIDIA Run:AI, with release-specific prerequisites. Ray and inference workloads can run on Kubernetes, but this does not make every separately packaged cluster product generally available; packaged managed Slurm remains planned.
How quickly can an AI cloud provider deploy tenant isolation on GPU infrastructure with vCluster?
Boost Run went from decision to production launch in under 45 days for its managed Kubernetes service. That is a customer example rather than a promised implementation time. Your launch depends on machine inventory, driver configuration, networking, storage, templates and service integrations. vMetal and vCluster connect machine orchestration with tenant cluster delivery, reducing the infrastructure workflows you need to build yourself.
Does vCluster's isolation model require provisioning separate physical clusters per tenant?
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.
See GPU Cluster Isolation in Action
See how GPU clouds and enterprises implement dedicated tenant environments with vCluster.
