The chain
Models larger than one accelerator exchange state for every token. The link begins as a system-design decision and ends in qualified optical inputs whose capacity cannot be recreated by adding module lines.
What matters now
Once the gate climbs upstream, downstream floor space cannot clear it. Allocation follows qualified chips and materials.
A captive laser fab is overhead when inputs are abundant. When the input binds, it becomes an allocation competitors must contract for.
Co-packaged optics and silicon photonics change lasers per port. Another module assembler does not.
How the constraint climbed
A selected optical chronology. It follows commitment and the location of scarcity, not every product announcement.
A buyer reserves the lines
Applied Optoelectronics files a Microsoft agreement whose payments secure a ten-year irrevocable reservation of dedicated production lines. The buyer secures capacity rather than merely placing module orders. AAOI filing
EML demand separates from the module
Lumentum reports record EML orders from multiple customers and substantial 200G EML orders. The order signal is now visible at the laser-chip layer. Lumentum
The operating discussion moves upstream
Lumentum's earnings discussion places pressure around EML supply, InP processing and specialized packaging rather than final module assembly. The binding work has moved into qualified component manufacturing. Lumentum call
The material becomes visible
TrendForce identifies InP substrate supply as an emerging AI optical-interconnect bottleneck. This is the current map's floor—not a claim that the branch or theme ends there. TrendForce
The current read
Scarcity moved toward the input
Module assembly was the visible limit. EML chips and then InP became the slower qualified layers. Each move narrowed the supplier set and lengthened the expansion clock.
Captive versus merchant supply
The useful comparison is not module capacity. It is who owns qualified laser output, who relies on merchant allocation, and which contractual state secures access.
Supply or architecture
- InP substrate, epitaxy and laser-fab additions qualify.
- Yields and hermetic packaging throughput improve.
- CPO or silicon photonics cuts lasers per useful port faster than port counts rise.
Four disconfirming signals
- Merchant laser lead times normalize.
- Capacity rights stop carrying value in new agreements.
- Module ASPs hold while laser pricing falls.
- CPO deployment removes total laser demand rather than relocating it.