I remember the moment I first realized the grid was cracking. It was late July, and a private chat among a small group of open-source hardware auditors I’ve been part of since 2017 started buzzing. Someone posted a raw snippet from a Nomura report, not a headline. It was a price prediction for Indium Phosphide substrates—two-inch and three-inch. The numbers were staggering: 42% to 78% increases. My first thought was that this was a typo, or perhaps a brief speculative flurry in a niche corner of the materials market. But the deeper I dug, the more I felt that familiar, chilling sensation from my early days auditing The DAO’s successor in 2017 when I found 42 critical logic flaws that exploited trust assumptions. We are not looking at a simple price hike. We are looking at the structural bottleneck for the entire AI infrastructure stack, a bottleneck that could decide who gets to build the next generation of intelligence.
The material in question, Indium Phosphide (InP), is not the stuff of consumer laptops. It is a III-V compound semiconductor, a niche player with a history of less than 30 years in commercial production, used almost exclusively for photonic integrated circuits (PICs). It processes light, not just electrons. For years, it was the quiet backbone of long-haul telecommunications and the occasional high-end data center link. Today, it is the soul of the 800G and 1.6T optical transceivers that are now the neural fibers connecting GPU clusters. Every time a cluster of Nvidia H100s or GB200s trains a model, it is essentially performing a symphony of data exchange. The bottleneck is no longer the GPU core; it is the I/O. And that I/O passes through an InP laser, specifically an Electro-Absorption Modulated Laser (EML). My own analysis of the supply chain, based on public filings and technical roadmaps from Sumitomo Electric, IQE, and AXT, reveals a system grinding at 90-95% utilization. There is no slack. A single hiccup in a MOCVD reactor in Wales or a crystal growth furnace in Japan can ripple through the entire AI economy.
Let me translate the Nomura data into a language that speaks to the hardware conscience. The report predicts a 42-76% increase for two-inch InP substrates and a 78% increase for three-inch. This asymmetry is not random. It is a signal. The industry is trying to force a migration from two-inch to three-inch wafers to improve economics and yield. But the yield on three-inch is still significantly worse—my back-of-the-envelope, using open-source defect data, suggests a yield gap of at least 15-25% between two-inch and three-inch substrates. This means the higher price is not just profit; it is an insurance premium against the difficulty of scaling a boutique process. The real alarm, however, is in the EML epitaxial wafer segment. Nomura suggests a 50-75% price increase for two-inch EML epiwafers. This is the material that actually gets turned into the laser. The gap between substrate price and epiwafer price confirms my suspicion that the real bottleneck is not the raw crystal, but the deposition of the active layers in a MOCVD reactor. Those reactors, mostly from Aixtron and Veeco, have a lead time of 12-15 months. They are also subject to export controls under the Wassenaar Arrangement. For my Chinese colleagues in the photonics labs in Wuhan and Beijing, this is a nightmare scenario. The ability to produce a high-quality MOCVD tool for InP is still 3-5 years behind the leaders in the US, UK, and Japan. The industry is facing a double bottleneck: a physical one (MOCVD capacity) and a geopolitical one (access to that capacity).
This leads to the contrarian angle that keeps me up at night. While the market is bullish on InP—seeing this as the next NAND-like boom cycle for AI—I see the seeds of a profound vulnerability. The Nomura report itself draws the SanDisk analogy, implying a cyclical peak followed by a crash. But the real risk is not just a price correction. It is the acceleration of a replacement technology. Silicon Photonics (SiPh) has been the shadow for years, always promising but never quite delivering. But the current price pressure on InP is now giving the CSPs—the hyperscalers like AWS, Google, and Meta—an enormous incentive to finally push SiPh over the finish line. If a single major customer like Google or Amazon Web Services certifies a complete 1.6T SiPh transceiver architecture in 2026, the demand for InP EMLs could collapse by 30-40% overnight. The market is pricing in a perpetual supply constraint. I see a inelastic demand curve under severe stress, which creates a classic bubble dynamic. Furthermore, the geopolitical risk is not priced in. A US export control on InP substrates, which I estimate has a 60% probability in the next 12 months, would create a bifurcation of the market: a Western market paying higher prices for secure supply, and a Chinese market facing an 18-month supply gap. The Chinese government’s response, via the third phase of the Big Fund, is pouring money into domestic InP production, but the technical gap in three-inch and four-inch substrates is a chasm, not a gap.
From my audit work on the Chromie Squiggle NFT project in 2021, I learned that when a system is under extreme pressure, its architectural flaws become visible. The InP supply chain is an architectural flaw in the AI narrative. It is a small, boutique, high-margin market that was never designed to serve the hungry, continuous demands of a trillion-parameter model training regime. The current price surge is the market screaming that the infrastructure is unbalanced. The takeaway for any builder or investor is this: do not conflate a price spike with a sustainable new normal. The true battle for AI's future will be fought not just in the GPU memory bandwidth, but in the quiet, slow, and highly chemical world of the MOCVD reactor. The winner will not be the one who pays the 78% premium for three-inch InP, but the one who builds a more resilient, alternative path for light. And right now, that path remains dangerously dark.