This series follows one lab build from start to finish: adding a single GPU host to an existing VCF 9.1 management domain as its own cluster, then running a vSphere Supervisor on it as the base for VCF Private AI Services. It uses no new vCenter, no new NSX Manager and no NSX edges. Read the parts in order; each one picks up where the last stopped, including the failures and how they were fixed.

This is a lab and POC design. A one-host cluster has no redundancy, and the SDDC Manager properties in Part 1 are lab workarounds, not a supported production configuration.

The four parts

  1. Allowing single-host clusters in SDDC Manager: the feature flag and minimum cluster size, the service restart, and checking the flag through /v1/features.
  2. Building a one-host vSAN ESA cluster: commissioning the host, the cluster spec, and the FTT=0 default policy that fixes FAILED_TO_CREATE_FILE_INSIDE_DATASTORE.
  3. NSX VPC networking without edges: a distributed VLAN connection on the default transit gateway, an external IP block, and a two-node VNA cluster.
  4. Enabling a Supervisor with one control plane VM: the enable spec, the nested-snapshot and vLCM health-gate stalls, and a test pod that proves DNS and egress.

For SDDC Manager and fleet operations on the same release, see the VCF 9 Fleet Management series.