The market is not pricing in a material shortage. It is pricing in a structural choke point.
A Serenity declaration, citing a Nomura research report, claims AI photonic material prices will rise 42-78% for 2-inch and 3-inch Indium Phosphide (InP) substrates. Nomura is specific: 2-inch InP +42-76%, 3-inch +78%, EML epiwafers +50-75%.
This is not just a semiconductor story. This is a story about the physical infrastructure that powers the AI agents and smart contracts crypto wants to mint in 2026. We cannot talk about autonomous economic agents without talking about the tiny, fragile, 3-inch wafers that enable the optical interconnects inside an AI cluster. If InP supply is the new silicon wafer shortage, then every AI-crypto thesis built on inference-at-scale just acquired a hidden, physical tail risk.
Context: The Photonic Supply Chain
InP is a III-V compound semiconductor. It is not silicon. It is used almost exclusively for high-speed optoelectronic devices: the lasers and modulators inside 800G/1.6T optical transceivers that connect GPUs inside an AI data center. The current tech focus is on InP substrates (2-inch and 3-inch) and epiwafers (specifically EML and CW types). The substrate major was Sumitomo Electric (~40% share), followed by AXT/Beijing Tongmei (~25%), and Mitsubishi Chemical (~15%). For epiwafers, IQE (UK) leads at ~30%, then AXTI (US) at ~25%, and Sumitomo at ~20%. Chinese local epiwafer suppliers are still below 5% of qualified output.
This is a small, concentrated market. Global revenues are likely below $10B. The capex required to scale InP is not trivial. MOCVD machines, which grow the epitaxial layers, have a 12-15 month delivery lead time. InP crystal growth for substrates requires a 4-6 week boule process. The yield on 3-inch InP substrates is still probably 60-75% at best for Japanese producers, lower for Chinese challengers. This is not a market that can double capacity in six months.
The direct read is: this is a tight oligopoly facing an inflection in demand. AI clusters are ordering 800G transceivers in the millions. Each 800G module requires 8 EML lasers. Nomura projects the annual transceiver volume to grow from ~8 million units (2024) to ~20 million units by 2026. That implies epiwafer demand tripling. Supply cannot follow that quickly.
Core Insight: The Three Hidden Signals
I have spent the last three years auditing supply chains for DeFi projects that required real-world collateral. That forensic skepticism is useful here. The Nomura data contains three signals the market is underestimating.
First, the core bottleneck is not the substrate. It is the epiwafer. The report references AXTI and IQE, which are epiwafer producers, not substrate producers. The price differential between 2-inch and 3-inch substrates (+78% vs +42-76%) is a signal. 3-inch wafers allow for more dies per wafer, but the yield penalty is severe. The price hike on 3-inch is a premium for the capacity to transition away from the legacy 2-inch node. The true capacity constraint is in the MOCVD reactors capable of producing 3-inch EML epiwafers at scale. IQE and AXTI control that bottleneck. This is not a substrate play. This is an epiwafer play.
Second, the AI demand is specifically for EML, not CW lasers. The report states EML epiwafers will rise 50-75%, vs. CW at 40%+. This confirms that the dominant architecture for AI interconnect (longer reach, higher baud rate) is electro-absorption modulated lasers. This is not about VCSELs (short range). It is about EML. Any project that assumes cheap, abundant, short-range optical links for AI inference nodes is missing this detail. The demand pull is for high-performance, long-reach, 100G per lane EML. That is physically harder to produce.
Third, the Nomura report itself is a catalyst for speculative capital. The report is likely being circulated to institutional desks. The 2017 analog is useful here: a detailed external report on a constrained supply chain often triggers a front-running cycle. Hedge funds will buy AXT and IQE stock, pushing the price higher. This creates a positive feedback loop: higher stock prices validate the scarcity thesis, which encourages more capex (which takes 18 months) and more inventory hoarding (which accelerates the shortage). The report is not just a forecast. It is a trigger.
Contrarian Angle: The 2027 Silicon Photonics Overhang and the Regulatory Void
The current narrative is unidirectional. Buy the InP supply chain. Bet on the AI optical interconnect boom. This feels like a NAND flash cycle from 2017, as Nomura itself hints with the Sandisk analogy. But there are two contrarian vectors.
First, silicon photonics (SiPh) is a credible alternative. Intel, Cisco, and Marvell are all pushing SiPh transceivers for 800G and 1.6T. A SiPh-based transceiver does not use an InP EML laser for modulation. It uses integrated silicon modulators with an external InP laser source (less pure InP demand) or potentially no InP at all with heterogeneous III-V on silicon integration. If a major CSP (AWS or Google) makes a full-scale switch to SiPh for 1.6T by late 2026, the InP epiwafer demand growth could slow from ~25% CAGR to ~10% by 2028. The market is not pricing this risk. The premium on InP today includes a zero-percent probability of SiPh winning the next generation.
Second, and more relevant for crypto, the regulatory risk is asymmetric. The US BIS can easily classify high-purity InP substrates and epiwafers as "AI-enabling materials." If they do, exports to Chinese module makers (who control ~50% of the global transceiver market) would halt. The report notes this risk. But the market is not pricing the direction of the disruption. If the US restricts exports, Chinese module makers will be forced to switch to domestic InP sources, which are lower purity and lower yield. This will degrade the total throughput of Chinese AI clusters. It also creates a bifurcated market: InP prices in the Western supply chain spike further (more demand chasing fewer qualified Western suppliers) while Chinese prices crash (oversupply of low-grade domestic output). This bifurcation is not a simple shortage. It is a structured imbalance that creates arbitrage for those who can legally navigate both pools.
Takeaway: The Real Thesis Is Not About InP
InP is a proxy. The real question is: how quickly can the physical layer of AI compute scale?
My CBDC work taught me that macro liquidity cycles dictate crypto market cycles. But macro liquidity is now increasingly dependent on physical infrastructure. You cannot have a 50-billion-dollar market for autonomous AI agents in 2027 if the optical interconnects that allow those agents to talk to each other are constrained by a 3-inch wafer node that takes 18 months to expand. The supply chain for AI compute is becoming the de facto monetary policy of the AI-crypto convergence.
The Nomura report is a canary in the MOCVD chamber. The market will chase AXTI and IQE. It will bid up physical hashrate proxies. But the contrarian trade is not to load up on InP stocks. The contrarian trade is to look at the SiPh alternatives, the MOCVD equipment suppliers (Aixtron, Veeco), and the regulatory bifurcation that will split the market into two liquidity pools. The 2017 ICO dream was regulation. The 2025 AI-crypto dream is a physical supply chain constraint. Treat the Nomura report not as a buy signal, but as an audit trigger.