Photonics and silicon photonics

Core thesis

As AI clusters scale, the bandwidth between chips outruns what copper can carry; co-packaged optics (CPO) bonds the optical engine into the switch or accelerator package and removes the electrical-reach wall. That makes optical interconnect the third constraint on the AI build-out, after power and silicon. The non-obvious part is where the value lands. Nvidia has funded rival laser makers (Lumentum, Coherent) and a fibre maker (Corning) on non-exclusive terms - commoditising the inputs it can fund a second source in. The one layer it cannot fund a rival in is the leading-edge silicon-photonics integration process, TSMC's COUPE, the optical analogue of CoWoS. So the optics suppliers win volume and a competed-down margin; the scarcity premium sits with the process holder.

State of the art (2026)

Broadcom said on 8 October 2025 that it was shipping Tomahawk 6 Davisson, an Ethernet switch with co-packaged optics that carries 102.4 terabits per second and is built with TSMC's COUPE process. Marvell showed its own co-packaged-optics platform in 2026. Both are suppliers' own announcements. Wider use depends on customers qualifying the parts, on reliability in service and on the cost of the whole system. Pluggable optics still carry most ports. Optical computing remains at the demonstration stage.

The live leg is interconnect, not compute

Two legs, different clocks. Optical interconnect - co-packaged optics and optical I/O moving data between conventional silicon chips - is crossing to volume now because copper's electrical reach fails as switch-ASIC bandwidth doubles every 18-24 months. Optical computing - matrix multiply in the photonic domain - remains demonstration-stage. The capital, the deployment, and the value are all on the interconnect side; betting on photonic accelerators is betting on the wrong leg.

Commoditised inputs, one un-commoditised tier

Nvidia funded rival laser makers (Lumentum, Coherent) and fibre (Corning) on non-exclusive terms - the move that commoditises an input so no supplier wins HBM-style pricing power. The laser and fibre names (COHR, LITE, GLW) therefore get the volume of the build-out and a competed-down margin, not a scarcity premium. The one tier no one multi-sources is the silicon-photonics integration process - TSMC's COUPE, the optical analogue of CoWoS - and that is where any surviving rent sits, with the process holder, not with the buyer that rents it. The buyer captures value only where it in-houses the integration (Quantum-X, Spectrum-X on COUPE), a narrower and separate claim from multi-sourcing the inputs.