16.1 Deployment Models
The five-ish ways to deploy an AI training cluster — bare metal, VM with SR-IOV passthrough, containers on bare metal, Kubernetes, and cloud-managed. Pros, cons, and where you'll see each.
16.2 Host Networking
The host-side stack that gets RDMA into a Kubernetes pod — PF/VF, SR-IOV, Multus, GPU Operator — as a build-time map, with the deployment order and pointers to the deep mechanics in Phase 5.