How AI Data Moves: From GPU Racks to Optical Networks | BEP Research

English automatic transcript

Recovered from the final Ainsley narration through automatic speech recognition and checked against the public YouTube transcript. Technical spellings are corrected. Wording and caption timings are approximate; this is not a producer-verbatim transcript.

Video: https://www.youtube.com/watch?v=yOfBEaNsUSc
Duration: 102.541667 seconds. Narration starts after the 2-second lead.

AI infrastructure is a system. Faster processors need bandwidth to memory, and to each other. Unlike many independent server tasks, distributed AI training can require tightly coordinated data exchange. Optics connects racks and sites, while memory bandwidth remains a separate constraint.

Follow one connection. Inside a GPU tray, electrical data reaches a transmitter, which encodes bits onto laser light. Fiber carries the signal to an optical circuit switch. Paired MEMS mirrors tilt together. One aims the beam toward an output mirror, and the other guides it into the destination fiber. The switch changes physical connections without decoding packets. At the destination, a receiver converts light back into electrical data.

Reconfiguration takes time. Traffic pauses while mirrors settle and the link recovers. Circuits change, rather than routes for individual packets.

Stronger lasers provide a power budget for optical losses and, in some architectures, splitting light among channels. Each receiver still needs enough signal.

Narrow linewidth means concentrated optical frequency, not a thinner physical beam. Requirements depend on the architecture.

Beyond the building, coherent transceivers carry data between sites, with optical amplification where needed. Distance still adds delay. These optical links can carry Ethernet traffic. Optics describes the physical transport, rather than replacing the network protocol.

Lumentum supplies distinct lasers, optical components, transceivers, and circuit switches across these layers. Competitors exist. The challenge is combining power, low noise, heat control, and reliability at manufacturing scale.

This is an illustrative system, not a specific deployment.