Why is indium phosphide the bottleneck behind co-packaged optics?
CPO moves optical conversion from a pluggable module onto the same substrate as the switch ASIC. The constraint is not the engineering. It is that silicon cannot generate light, and the material that does is in deep shortage.
Co-packaged optics is the next networking jump in the data center. Its supply ceiling is set by indium phosphide laser capacity, which sits with a handful of vertically integrated vendors and their substrate suppliers.
Silicon cannot make light
Inside a chip, you can move electrons across copper or aluminium traces at very high speeds. To move signal between racks at the bandwidth AI training requires, those electrons have to become photons. Silicon is a poor light emitter, so every laser in the data center sits on a different material.
The material of choice is indium phosphide. It has the right band gap to produce light at the wavelengths optical fibres carry well. Every long-reach link in an AI cluster begins and ends at an indium phosphide laser.
CPO changes the laser physics required
A pluggable transceiver lives outside the switch and can dissipate its own heat. A co-packaged optics module sits on the same substrate as the compute or switch die. The lasers have to be higher power, narrower in linewidth, and quieter to compensate for the shorter integration path.
In practice that means each laser die has to be physically larger. So the same wafer of indium phosphide produces fewer lasers per dollar. CPO is not adding new customers to the laser supply. It is enlarging every customer.
Source: Irrational Analysis interview, Chris Barber, May 2026
EML lost to SiPho one cycle early
Historically, each new transceiver generation started on EML, a monolithic indium phosphide chip that holds both the laser and the modulator. After a year or two of the generation, the industry would move to silicon photonics, which packages a smaller continuous-wave laser inside a silicon photonics chip to save money.
For 1.6T transceivers, silicon photonics took majority share within six months. The reason is supply: EML lasers were already short, and the vendors that produce them reallocated capacity to the higher-margin continuous-wave lasers that CPO demands. The EML side starved.
Source: Irrational Analysis interview, Chris Barber, May 2026
The supply graph
- Lumentum and Coherent. The two main vendors of indium phosphide laser modules. Both sold into 2027 and 2028 capacity at recent earnings.
- Sumitomo and AXT. Indium phosphide substrate vendors. Substrate quality bounds laser yield.
- IQE. UK epitaxy specialist that grows the layered indium phosphide structures lasers are printed onto.
- Aixtron. German equipment vendor that makes the deposition tools the substrate and epitaxy houses buy.
- Tower Semiconductor. Silicon photonics foundry that is taking the SiPho work the indium phosphide shortage created. Sold out into 2027 and 2028 per Irrational Analysis.
- Semtech. Analog amplifier and equaliser supplier whose parts sit inside active copper cables, transceivers, linear pluggable optics, and CPO modules. Near-duopoly with MACOM.
- Broadcom. Operates a large indium phosphide laser division on top of its switch ASIC business.
Source: Irrational Analysis interview, Chris Barber, May 2026
Reliability is the wrong worry
A common concern about CPO is reliability: putting lasers next to hot compute, surviving repeated thermal cycles, dealing with field replacement. Irrational Analysis argues this is overstated. Done correctly, CPO is more reliable than pluggable transceivers because there are fewer connectors and the optical path is shorter.
The real risk is upstream, and it splits in two. One half is whether there is enough indium phosphide capacity to feed the 1.6T and 3.2T transceiver generations and the CPO ramps that follow. The other half is advanced packaging: getting the photonic engine onto the same substrate as hot logic at yield, with a fiber attach that survives qualification.
Source: Irrational Analysis interview, Chris Barber, May 2026
Strategic read
For Pere, the lens is this: when a bottleneck rotates from logic to optics, the value chain reshuffles. Module assemblers (the Innolights of the world) move into the spotlight, but the binding constraint is upstream at substrate and epitaxy.
The CPO supply story is the clearest current example of the more general pattern: AI infrastructure cost is not where the brand value sits. It sits two or three layers deeper, in materials and equipment with names most software people have never heard.