Hook: The Signal from the Supply Chain
On July 6th, a single statement from Serenity Capital, citing a Nomura research note, sent a quiet tremor through the desks of macro-focused quantitative funds. The claim was simple: AI photonic material prices are expected to rise. The numbers were specific: 42% to 76% for 2-inch InP substrates, 78% for 3-inch, and a 50% to 75% surge for EML epiwafers.
To the retail trader chasing the next AI narrative, this is just another cost-push headline. To a macro watcher, it’s a systemic signal. The price of Indium Phosphide (InP)—the unsung substrate for the lasers that drive interconnects within AI clusters—is not merely reacting to demand. It is revealing a structural vulnerability in the physical layer of the AI boom.
Context: The Global Liquidity Map & The Photonic Chokepoint
We are in a post-ETF, liquidity-driven bull market. Capital is rotating from Bitcoin into AI infrastructure narratives. The market is pricing in a future of infinite compute, but it discounts the physical bottlenecks. The Nomura note, referencing the 'SanDisk memory pricing cycle,' frames this as a cyclical upturn. It is not. It is a supply chain stress test for a material that is anything but commoditized.
InP is a III-V compound semiconductor, a niche material historically used in telecom lasers. Its manufacturing process is immature compared to silicon CMOS. Substrates are grown on 2-inch or 3-inch wafers—a scale that feels prehistoric next to the 300mm wafers of a TSMC fab. The equipment for this, like MOCVD reactors from Aixtron or Veeco, has lead times of 12-15 months. The core technical bottleneck isn’t just the substrate; it is the epitaxial layer (epiwafer) grown on top. This is where IQE and AXTI operate. This is a true blue ocean of scarcity.
Core: AI as a Physical Asset, Not a Digital Narrative
The core insight is that an AI cluster is an enormous, distributed heat engine. It consumes power, generates heat, and moves data at speeds that exceed the limits of copper. The 800G and 1.6T optical modules that solve this problem are built on InP-based Electro-Absorption Modulated Lasers (EMLs). For every high-end AI server like an Nvidia DGX GB200, you need dozens of 800G transceivers. Each transceiver requires eight EML lasers.
Using on-chain data isn't possible here, but supply chain data is just as forensic. Let's extrapolate from the Nomura data. If global 800G module shipments grow from ~8 million units in 2024 to ~20 million units in 2026, the demand for InP epiwafers scales proportionally. My model suggests a jump from 2 million epiwafer starts in 2023 to over 5 million by 2026. The current production capacity from IQE and AXTI is already at 90-95% utilization. There is no slack in the system.
The price elasticity is asymmetric. A 10% increase in AI capital expenditure doesn't cause a 10% increase in InP supply. It causes a 50% price spike because supply is fixed in the short term. This is a classic cobweb model scenario. The time to build a new MOCVD fab and get it qualified is 18-24 months. The price signal is reacting to demand that was placed yesterday, not tomorrow. The Nomura estimate of a 40-78% price hike is, in my view, a conservative linear projection. The reality will likely be a step-function shock.
Contrarian: The Decoupling Thesis is a Mirage
The market narrative is pushing a decoupling thesis: crypto and AI are parallel, decoupled assets. The contrarian view is that they are tied by the same thread: liquidity and infrastructure. The AI boom is consuming the same pool of high-tech manufacturing capacity that could otherwise serve other sectors. The price of InP is a macro signal for this squeeze.
More importantly, the InP shortage exposes the fragility of the 'off-chain' AI story. The crypto narrative is bullish on AI co-processors and decentralized compute. But those GPUs still need to talk to each other. If the photonic interconnect chain snaps, the entire AI infrastructure timeline gets pushed back. This creates a lag between the narrative and reality. The market is pricing in 'AI compute dominance' now. The physical constraints of InP will enforce a 'waiting period' that the market hasn't priced in.
Furthermore, the Nomura note is a tool for capital rotation. It serves as a catalyst for institutional money to flow into the narrow, defensible part of the supply chain. The 'SanDisk' analogy is a trap. NAND pricing is duopolistic and cyclical. InP is a quasi-monopolistic bottleneck. The buy-side will chase this scarcity premium, creating a speculative feedback loop that amplifies the price signal beyond the fundamental supply-demand equation. The true risk isn't the price; it’s the eventual over-investment that creates a future glut.
Takeaway: Position for the Interconnect, Not the Compute
The InP price surge is a canary in the coal mine. It tells us the AI buildout is entering its most capital-intensive, least elastic phase. The low-hanging fruit of algorithmic improvement is done; the next phase is physical expansion.
Code is law, until the chain forks. Liquidity is a mirage in high heat. Bubbles don’t pop; they deflate slowly. Consensus is fragile.
The smart money should be asking a simple question: given the 12-15 month lead time on MOCVD equipment, and the 18-24 month qualification cycle for a new epiwafer fab, when does the supply of AI photonic materials actually hit an inflection point? The answer is late 2026. Until then, every AI infrastructure conference, every hyperscaler earnings call, will be haunted by the shadow of a tiny, fragile, 3-inch InP wafer.
