Skip to content

Virtualizing Large Scale GPU Cluster for Sovereign AI: Petasus AI Cloud Journey with Kubernetes#

KubeCon + CloudNativeCon Europe 2026 digest Program Adjacent pick 23–26 March 2026

Abstract Recording

Jian Li, SK Telecom

Play the recording: Virtualizing Large Scale GPU Cluster for Sovereign AI: Petasus AI Cloud Journey with Kubernetes

  • Why: Carries the number that matters for the virtualize-or-not argument on accelerated infrastructure: under 5% overhead versus bare metal, on a real 1000+ Blackwell GPU cluster.
  • What: How the Haein Supercluster — described as Korea's largest AI cluster and part of its Sovereign AI initiative — was virtualized on Kubernetes and KubeVirt.
  • How: The sub-5% figure is attributed to NVLink/NVSwitch and GPUDirect RDMA virtualization, explicitly contrasted with traditional PCIe- and TCP-based virtualization. Resources are partitioned via namespaces for on-demand tenant GPU clusters; automation cuts provisioning from days or weeks to roughly 10 minutes; DCGM and Prometheus provide observability.
  • Where applicable: Requires NVLink/NVSwitch-class hardware and GPUDirect RDMA — the performance claim depends entirely on those paths, and would not hold on PCIe/TCP. Multi-tenancy approach is broadly instructive.