Guest: Simone Morellato (LinkedIn)
Company: vCluster Labs
Show Name: An Eye on AI
Topics: Kubernetes, Cloud Native

Most platforms scramble to retrofit for AI workloads. vCluster didn’t need to. Simone Morellato, Customer Success Lead at vCluster, explains how the platform’s foundational multi-tenancy design made GPU infrastructure management a natural evolution rather than an architectural overhaul.

The vCluster story reveals an important lesson about platform design: solve the right abstraction layer, and you’re ready for workloads that don’t exist yet.

Morellato traces vCluster’s origins to a developer problem. Central IT would provision one Kubernetes cluster, maybe two if you were lucky. Developers discovered vCluster worked differently. Deploy vCluster as an app, and suddenly you have your own cluster. Deploy ten of them. IT doesn’t know. They think they gave you one. You’re using ten.

That capability, sharing infrastructure while maintaining isolation, became the foundation for everything that followed.

When GPU workloads arrived, vCluster didn’t need redesign. The same multi-tenancy architecture that lets developers self-serve clusters now handles AI infrastructure demands. Morellato details the technical evolution. Private nodes and auto nodes eliminate manual cluster configuration. vCluster provisions nodes automatically. It creates VMs on platforms like NVIDIA DGX, solving a critical density problem.

The math matters. Even with 100 gigabytes of RAM, Kubernetes servers max out at 200 pods. If your applications aren’t large, you waste resources. Create VMs, and each one runs 200 pods. Multiply by the number of VMs, and suddenly you’re running thousands of pods efficiently.

Networking segregation operates at layer two and layer three. Different tenants don’t interfere with each other. Storage isolation prevents accidental data access across boundaries. These aren’t AI-specific features. They’re foundational infrastructure capabilities that AI workloads desperately need.

Security took center stage with vNode technology. Container escapes remain a persistent threat. Break out of a container, and you access the underlying node. vNode creates a layer around the pod that looks like a node. Escape your container, and you’re trapped inside the vNode. The real node remains unreachable.

Morellato shares a compelling validation story. VMware, a company known for virtualization technology, used vCluster internally. His team approached him about adopting vCluster because running multiple clusters was becoming too expensive. His reaction captures the moment perfectly. How is this going to work? We are VMware. Isn’t there another technology that can do the same?

The team’s response: it just works. VMware deployed vCluster for labs shared among hundreds of people. They’re still using it.

The pattern repeats across organizations. People discover vCluster, find uses the original designers never imagined, and deploy it to solve infrastructure problems traditional Kubernetes struggles with.

Looking forward, Morellato sees the next evolution clearly. AI teams currently have abundant funding. Efficiency isn’t their primary concern. That changes. Eventually, AI data centers will need to optimize. They’ll need to share more efficiently. They’ll need the same multi-tenancy capabilities that made vCluster valuable for traditional workloads.

The Kubernetes ecosystem continues optimizing. vCluster sits at the intersection of that optimization drive and AI infrastructure demands. The platform that made developer self-service possible now makes AI workload efficiency achievable.

The broader lesson extends beyond vCluster. Design platforms to solve abstraction and isolation problems well, and you create flexibility for future workloads. vCluster didn’t know GPU workloads were coming. It didn’t need to. The architecture was ready.