InP wafer and epitaxy capacity is limited to a handful of suppliers worldwide, making it a recurring bottleneck whenever AI datacenter buildouts spike demand for optical links.