The Real Bottleneck in the Optical Interconnect Cycle Is InP
Damnang
· Damnang’s Substack
· April 22, 2026 at 07:35
· ⏱ 11 min read
| Read on Substack ↗
Summary
The article argues that Indium Phosphide (InP) is a structural bottleneck in AI optical interconnects because demand from AI data centers is accelerating dramatically (10x TAM growth from GB300 to Rubin Ultra) while supply expansion faces structural constraints—long qualification cycles, equipment lead times, manufacturing know-how barriers, and raw material concentration in China. This imbalance means the market has not yet fully priced in the persistent shortage across the InP supply chain.
•InP is irreplaceable for high-speed 1310nm/1550nm lasers and EMLs used in AI data center optical modules; CPO architecture further increases InP demand as external laser sources (ELS) scale with lane count, not module count.
•Goldman Sachs estimates combined scale-up and scale-out optical component TAM grows nearly tenfold from GB300 NVL72 to Rubin Ultra NVL576.
•GPU-to-optical-module attach rates are projected to rise from 2-3 in GB200/GB300 era to 4-6 starting from Vera Rubin.
•Silicon photonics is expected to reach roughly half the datacom optical module market within a few years, but this does not reduce InP dependency—it shifts the demand to external lasers.
•Structural supply constraints include 12-24 month customer qualification cycles, long MOCVD tool lead times and process stabilization, accumulated manufacturing know-how as a barrier, and geopolitical risk from China's concentration of indium supply.
•The shortage spills across the entire InP chain: major optical IDMs expanding capacity tighten substrate/equipment availability, pushing external foundry capacity to the limit.