China's InP Wafer Capacity Shortage Constrains AI Data Center Optical Interconnect Supply Chain
Nashnova编辑部
China produces over 70% of the world's raw indium, yet holds only ~10% of the finished indium phosphide (InP) substrate market — a gap turning critical as AI data-center interconnect speeds climb from 400G toward 3.2T.
What is InP, and why can't AI data centers do without it?
InP — indium phosphide, a semiconductor used to make lasers and photodetectors — is the core substrate inside data-center optical modules. This means → the chips that send and receive light signals in a data center are grown on this material.
Sumitomo Electric's roadmap shows intra-data-center link speeds rising from 400G through 800G and 1.6T to 3.2T, with some server connections shifting from copper to fiber. In plain terms = every speed upgrade lifts optical-module demand, and InP substrate consumption rises in step.
Three companies — Sumitomo Electric, JX Advanced Metals, and AXT — together hold over 90% of the global finished InP substrate market.
China has 70% of the indium — so why can't it make enough wafers?
Domestic producers face four bottlenecks: insufficient high-purity feedstock, low wafer yields, limited large-diameter manufacturing capability, and lengthy customer-qualification cycles. This means → owning the ore is not the same as producing finished wafers; an entire precision-manufacturing chain sits in between.
Yunnan Germanium's subsidiary Xinyao Semiconductor, one of few volume producers, currently makes 2-to-4-inch substrates at roughly 150,000 wafers per year. In April it announced a ~RMB 189 million (~$28.1 million) expansion to add 300,000 wafers/year, including a 6-inch pilot line of 6,000 wafers/year — but 6-inch output remains in test-sample quantities only.
Sanan Optoelectronics runs its own InP epitaxy and photonic-chip lines, yet still sources 100% of its substrates externally. Other Chinese chip makers are still qualifying domestic substrate sources. In plain terms = even China's largest photonic-chip producer depends on others for this core material.
Does Beijing Tongmei count as Chinese domestic capacity?
Beijing Tongmei is a wholly owned subsidiary of U.S.-based AXT. Its output primarily serves the parent company and export markets. This means → the factory is in China, but its production is not available as a domestic-substitution resource.
AXT plans to expand Beijing Tongmei between 2026 and 2028 to fulfill a 6-inch InP substrate supply agreement with Coherent — the added capacity is already spoken for by an overseas customer.
This reflects a distinction easy to overlook: being "located in China" and being "available to China's supply chain" are not the same thing.
How are the dominant overseas suppliers racing to expand?
JX Advanced Metals plans to invest up to ¥120 billion (~$7.53 billion) over four years, adding lines at Hitachi-Naka while keeping its Isohara site running. The target: seven to ten times current capacity. This means → customer orders have already exceeded the company's earlier forecasts, and the scarcity of qualified capacity is strengthening incumbent suppliers' pricing power.
Sumitomo Electric targets InP substrate capacity at 2.4× its 2023 level by 2028, with 4-to-6-inch large-diameter, high-quality substrates as its core competitive edge.
Buyers are locking in supply early. Lumentum signed a six-year capacity-reservation deal with AXT, paying an initial $43.5 million deposit. Coherent signed a three-year 6-inch substrate development agreement with AXT-Beijing Tongmei. In plain terms = downstream companies are putting real money down on future wafers — a sign that the supply crunch is not a forecast but a present reality.
Can China's domestic capacity catch up? How wide is the time window?
Upgrading from 3-inch to 4-inch to 6-inch substrates requires new equipment, recalibrated thermal processes, and fresh yield ramps at every step. From line construction to qualified volume supply typically takes two to three years or longer.
Substrate expansion cycles run far slower than GPU deployment timelines — announced capacity is not the same as qualified wafers a device maker can use today. This means → even with investment in place, the supply gap will not close quickly.
Whether China's domestic InP capacity can leap from pilot line to volume production within two to three years is the critical proof point for narrowing this structural gap.
Content is for reference only, not financial advice.