Who makes optical transceivers for AI data centers

Optical transceivers plug into the switch systems (3.03) and NICs to link them over the fiber and connectors in 3.11, carrying the scale-out traffic that copper cabling (3.09) can't reach as far or as fast. Inside each module sit the lasers (3.06), DSPs and drivers (3.07), and — as speeds rise — silicon photonics (3.08) that do the actual signal conversion; module vendors largely assemble and test these components rather than build them all in-house, which is why optics contract manufacturing (3.05) is its own layer. AI clusters consume far more optical ports per GPU than general-purpose data centers because of the scale-out bandwidth growth described in 3.02, and the industry's steady move from 400G to 800G to 1.6T tracks rising per-GPU network bandwidth requirements directly. An investor should watch module vendor concentration by speed grade, since a small number of merchant module makers plus a GPU vendor's own branded modules account for most volume at the newest speeds, and watch qualification cycles as each new speed grade requires re-qualifying lasers, DSPs, and photonics together.