The AI infrastructure layer

A modern training cluster is thousands of accelerators that have to behave like one computer, and this layer is the wiring, buildings and materials that make that possible. Scale-up and scale-out fabrics, spine-leaf networks, the switch silicon that moves the bits and — above all — the optics carrying traffic between racks (transceivers stepping from 400G to 800G to 1.6T, plus co-packaged and linear optics and the silicon photonics behind them) are scaling faster than almost anything else in the build-out. Alongside the networking sit the AI factories themselves, the data-center construction that erects them, high-density rack systems, and the critical materials — like indium phosphide for optics — that gate what can be built. Even subsea bandwidth matters once clusters span regions. This layer is where raw compute becomes usable capacity, and where a single scarce component — a specific optic, a materials supplier — can throttle an entire deployment.