Offer tenant Kubernetes environments on DGX SuperPOD infrastructure. vMetal orchestrates machine provisioning, while vCluster separates tenant control planes and dedicates DGX workers through Private Nodes.
DGX hardware performance only pays off when the Kubernetes layer on top can serve many tenants cleanly.
In a namespace model, tenants can see platform internals, and every customer sits in the same blast radius.
A separate control-plane server stack per tenant consumes capacity and increases infrastructure operations alongside customer GPU workers.
AI teams have used AWS and GCP. They expect self-service Kubernetes and familiar tooling, or they go back to a hyperscaler.
vMetal orchestrates supported DGX infrastructure provisioning. vCluster supplies independent tenant control planes and dedicates DGX workers through Private Nodes; fabric partitioning and storage performance require their own policies. Lintasarta runs 170+ tenant clusters in production.
What it takes to turn DGX SuperPOD infrastructure into a managed Kubernetes service for AI customers.
Each tenant has an independent Kubernetes API, data store and RBAC boundary. Lightweight control-plane hosting reduces dedicated server requirements for DGX SuperPOD environments; Private Nodes keep production workers tenant-specific.
vMetal exposes one Machine API above bare metal and VM provisioning drivers. Registered inventory, images and credentials enable repeatable machine lifecycle workflows for DGX SuperPOD environments. Network automation uses supported integrations such as Netris.
Every tenant cluster is a CNCF-conformant control plane that exposes standard Kubernetes APIs. Tenants run the GPU Operator, training jobs and inference services without proprietary constraints.
Use supported workload templates for DGX SuperPOD environments, including the NVIDIA Run:AI partner integration. Configure GPU, network and storage prerequisites, and validate the selected template for Private Nodes before offering it to production customers.
Private Nodes dedicate worker capacity to one tenant cluster, with its own CNI and storage configuration. For DGX SuperPOD environments, select physical servers or VMs according to the required infrastructure boundary.
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.
Talk to our team about your stack
Deploy vCluster on your infra in minutes
Go live with a hyperscaler-grade tenant experience in days
vCluster Labs software powers 100K+ GPUs across a combined customer base of 50+ GPU clouds & Fortune 500s. Named examples include Boost Run and Lintasarta, which runs 170+ tenant clusters in production. These figures describe adoption and deployment scale; they are not a benchmark or a guarantee that a particular hardware, network or application configuration will meet your requirements.
Tenant control planes run as lightweight pods in a control plane cluster, and each tenant's workloads run on its own Private Nodes. If you want the control plane cluster itself on bare metal, vCluster Standalone runs as a single CNCF-certified binary with no k3s or kubeadm underneath.
At the control plane, every tenant has its own API server and data store. At the node level, Private Nodes give each customer dedicated DGX nodes, which is the recommended production model. Internal dev, test and CI clusters for trusted teams can run on shared nodes. Hardware network isolation such as VLAN, VXLAN or VRF comes through the Netris integration, and vNode can add a workload boundary using Linux user namespaces and seccomp.
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.
NVIDIA Run:AI is available through the native Certified Stacks integration, with release-specific prerequisites and templates. Its bundled templates do not automatically supply dedicated Private Nodes. Configure and validate the production worker model, GPU access, networking and storage before offering it to customers. Ray and Jupyter can use compatible Kubernetes deployment tooling and require their own supported configuration.
vCluster Platform coordinates tenant environments through a central UI, CLI and API. Operators use templates, access policies and quotas while each Kubernetes tenant retains its own API and dedicated production workers. vMetal manages supported machine provisioning below the fleet. Operational consistency still depends on template configuration, version management and the monitoring and recovery workflows you connect to the platform.
See how AI cloud providers run isolated tenant clusters on their DGX SuperPOD hardware.
