Seven chapters of protocols, algorithms, and topology. All of it becomes worthless the moment a single photon fails to transit a one-metre copper cable.
Chapter 8 ended with NCCL translating gradient tensors into RDMA operations. Chapter 7 calculated exactly how many switches and links it takes to deliver non-blocking bandwidth. Chapter 5 tuned the congestion control that prevents buffer overflow. Every layer assumed the physical layer worked. This chapter explains what it takes to make that assumption true.
The physical layer in an AI cluster is not passive. It is an active, engineered, thermally significant system. The optics industry is moving faster than switch ASICs -- 400G is the H100 baseline, 800G is ramping now, 1.6T is arriving in 2026, and co-packaged optics will fundamentally change how AI switches are built and managed. If you are sizing, deploying, or troubleshooting an AI fabric, the physical layer is not infrastructure background noise. It is a primary constraint.