Lumentum is shipping 30% less than customer orders for photonics components essential to AI datacenter networks. The company's optical communications systems (OCS) backlog has surged past $400 million, with most units scheduled for second-half delivery.
All EML production capacity is committed through the end of calendar 2027 under long-term supply agreements. Lumentum increased its indium phosphide manufacturing capacity by over 20% in the December quarter alone, with plans for further expansion. Despite these additions, the demand-supply gap continues widening.
EML chips convert electrical signals to optical light in AI datacenter transceivers, enabling high-speed data transfer between GPUs and servers. The components use indium phosphide wafers requiring specialized fabrication processes that take 18-24 months to scale.
Lumentum's position as a foundational supplier to virtually every AI network gives the shortage systemic implications. Hyperscalers including Microsoft, Google, and Meta are building clusters requiring hundreds of thousands of optical transceivers per facility. Each next-generation AI training cluster needs 10-15x more optical interconnect bandwidth than previous server generations.
The photonics bottleneck compounds GPU supply constraints that have dominated AI infrastructure headlines. Even companies securing Nvidia H100 or B200 chips face delays networking them into functional training clusters without adequate optical components.
Coherent Corp, II-VI, and other photonics manufacturers report similar capacity constraints. The specialized nature of indium phosphide fabrication creates high barriers to new entrants, limiting near-term relief. Traditional silicon photonics offer lower performance for AI workloads requiring 800G and 1.6T transmission speeds.
Long-term supply agreements extending through 2027 give photonics makers unprecedented revenue visibility but leave AI infrastructure buyers scrambling for allocation. Some hyperscalers are pursuing captive photonics capabilities, though developing competitive manufacturing takes years.
The shortage will likely constrain AI training capacity expansion through at least 2026, potentially creating a new chokepoint independent of GPU availability. Datacenter operators report optical components now consume 15-20% of total infrastructure budgets, up from 8-10% two years ago.