Every AI optical transceiver requires indium phosphide laser chips. Demand exceeds supply by nearly four to one. Equipment lead times stretch eighteen to twenty-four months. The AI infrastructure buildout depends on a compound semiconductor whose supply chain touches a country that controls two-thirds of the raw material.
Every GPU in every AI data center communicates through light. The optical transceivers that carry signals between chips at 800 gigabits per second rely on indium phosphide laser chips for the one function no other material can perform: generating the photons. Silicon photonics handles passive waveguides. Silicon nitride handles low-loss routing. Neither can produce light. That requires a III-V compound semiconductor manufactured in specialized fabs with equipment lead times of eighteen to twenty-four months.
The AI optical transceiver market reached $16.5 billion in 2025. TrendForce projects $26 billion in 2026, a fifty-seven percent increase driven by hyperscaler GPU deployments. Shipments of 800G-and-above transceivers will jump from 24 million units to 63 million in a single year. When the industry transitions from 800G to 1.6 terabit transceivers, indium phosphide usage per module nearly doubles. Millions of InP laser components per hyperscaler, per year.
The Gap
Global effective InP wafer production capacity stands at roughly 600,000 to 750,000 wafers per year. Demand exceeds two million wafers. The supply gap exceeds seventy percent. Prices for a two-inch optical-grade InP substrate surged from $800 in early 2025 to $2,300 to $2,500 by April 2026. 800G transceiver production falls forty to sixty percent short of demand through 2027, according to McKinsey, driven by electro-absorption modulated laser chip bottlenecks. LightCounting expects some relief by mid-2026 as new capacity qualifies, but the structural deficit persists into 2029.
The constraint is physical throughput. Indium phosphide substrate manufacturing requires metal-organic chemical vapor deposition reactors and molecular beam epitaxy systems with order books that extend through 2027. Skilled operators are scarce. Adding money does not add capacity on a timeline shorter than two years.
The Supply Chain
The raw material and the finished substrate occupy different geographies with different bottlenecks. China produces sixty to sixty-eight percent of the world's refined indium metal. The refined metal flows into substrate manufacturing dominated by Japan and the United States: Sumitomo Electric, AXT Inc, and JX Nippon Mining collectively control over eighty percent of global InP substrate capacity. Over ninety percent of high-end InP substrate production sits outside China. China's domestic share of six-inch InP substrates is below five percent.
Both halves of the chain are constrained. AXT, the second-largest substrate maker and a US-headquartered company, reported Q4 2025 revenue held back by delays in Chinese export permits for its Tongmei subsidiary, which manufactures in China. The US producer of InP substrates is itself dependent on Chinese regulatory cooperation for part of its output. The bottleneck compounds: raw indium supply controlled by one country, substrate manufacturing dependent on specialized equipment from another set of countries, and the largest US substrate producer caught between them.
The Lock-In
NVIDIA invested $4 billion in March 2026, splitting the sum between Coherent and Lumentum, the two companies that manufacture InP laser chips for AI optical interconnects. Both are sold out through 2027. Lumentum's optical connectivity backlog exceeds $400 million with multi-hundred-million-dollar co-packaged optics orders. NVIDIA has pre-allocated electro-absorption modulated laser capacity at both suppliers, pushing competitor delivery timelines past 2027.
The investment creates a two-tier market. Hyperscalers aligned with NVIDIA have secured optical supply. Everyone else is scrambling for allocation on a timeline that stretches beyond the planning horizon of most AI infrastructure buildouts. Co-packaged optics, which integrate laser chips directly onto the silicon die, are the next transition. That transition increases total InP demand rather than reducing it, because silicon photonics modules still require InP laser sources for light generation.
The Expansion Race
Substrate makers are investing. AXT completed a $632.5 million capital raise and plans to double InP substrate capacity in both 2026 and 2027. Sumitomo Electric is expanding six-inch InP wafer production by forty percent. JX Advanced Metals committed 1.5 billion yen to increase capacity at its Isohara plant by twenty percent. The top five suppliers account for roughly seventy percent of 2024 revenue.
The expansions will help. They will not close the gap. Even with every announced investment, capacity additions lag demand by the equipment lead time. The reactors needed to grow InP crystals and epitaxial layers take eighteen to twenty-four months to deliver, install, and qualify. Supply shortfalls of thirty to forty percent on 1.6 terabit transceivers are projected through 2029.
What This Changes
The AI supply chain bottleneck has migrated. GPUs were the constraint in 2023. High-bandwidth memory was the constraint in 2024. Networking became the constraint in 2025. In 2026, the constraint has reached a base material that most people in the industry have never handled. The $650 billion AI infrastructure buildout depends on a compound semiconductor with a seventy percent supply gap, manufactured from a raw material that one country dominates, processed in specialized fabs whose equipment lead times exceed the planning cycles of the companies that need them.
The winners are identifiable: Coherent and Lumentum hold the laser chip position. AXT, Sumitomo, and JX hold the substrate position. Indium recyclers and alternative-source miners gain relevance as primary supply tightens. The losers are also identifiable: any hyperscaler that did not lock optical supply before 2026 faces a timeline problem that capital alone cannot solve.
The falsifiable claim: if InP transceiver supply catches demand by mid-2027 without Chinese export relaxation, the bottleneck thesis fails. The second test: if a viable non-InP laser source for AI optical interconnects reaches production scale by 2028, the material dependency thesis fails. Neither outcome is supported by the current equipment order books, the capital expansion timelines, or the physics of light generation in silicon.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)