Hardware procurement runs on its own calendar, and it is not yours. Ours took months. The cluster did not wait: we pointed the whole provisioning pipeline at a pile of VMs and started making real mistakes early.

Not a simulation — the same path. Cluster API declares the cluster, CAPT claims a machine from the Tinkerbell inventory, Tinkerbell netboots it, images the disk and power-cycles it through a small BMC shim, and the Talos providers bootstrap the control plane. Nine nodes, three control-plane. From CAPI CR to ready cluster: 6-8 minutes, nobody touching a terminal. A bare-metal pipeline is testable long before it has anything to stand on.

We will cover why the Tinkerbell stack cannot be remote (L2 is L2), reaching an on-site API server from a cloud controller, Service LoadBalancer via Cilium BGP with no cloud LB, which parts of the state machine a VM cannot fake — and the CAPT Hardware-claim race we debugged, fixed and got merged upstream.