Essay - 2026-06-01

AI's photonics bottleneck: why moving data is now harder than making chips

AI clusters are running out of room on copper. As back-end fabrics climb from 800G to 1.6T and 3.2T, the constraint stops being how many GPUs you can buy and becomes how cheaply you can move data between them. The case for optics as the next chokepoint, and a map of where it bites first.

AI's photonics bottleneck is the point where moving data between GPUs, racks, and sites becomes harder than adding more compute. As back-end fabrics climb from 800G to 1.6T and 3.2T, copper links hit reach, power, signal-integrity, and cabling limits.

Photonics solves part of the problem by moving data as light, but it relocates the constraint into lasers, silicon photonics wafers, photonic IC packaging, electro-optical test, fiber attach, blind-mate connectors, and rack-level serviceability.

Co-packaged optics shortens the high-speed electrical path by moving the optical engine next to the switch or accelerator chip. It does not replace pluggables soon; the realistic outcome is a hybrid stack where 800G and 1.6T modules carry volume and CPO takes the densest fabrics first.

NVIDIA Spectrum-X and Quantum-X Photonics, Broadcom Bailly, GlobalFoundries Fotonix, and multibillion-dollar NVIDIA deals with Coherent and Lumentum are the receipts that optics is becoming a core layer of the AI factory.