Is AI's Next Bottleneck a "Beam of Light"? How the Indium Phosphide (InP) Shortage Is Reshaping Optical Networking

One-sentence takeaway: AI data centers are shifting from electrical to optical interconnects, and the core material behind optical links — indium phosphide (InP) — has highly concentrated capacity and long expansion lead times. Supply-chain voices now warn InP could be "scarcer than memory," and this shortage is redrawing the map of the optical networking industry.

A supply-chain warning puts "a beam of light" in the headlines

In mid-August, "indium phosphide shortage" became a trending topic. Taiwanese business media asked whether "the InP shortage will be worse than memory," while others wondered if InP is "the next rare earth." To most people, indium phosphide sounds like an obscure chemical — but to AI data centers, it is the critical material behind 800G/1.6T optical modules and co-packaged optics (CPO). Without it, the fiber links between GPUs simply cannot light up.

What is InP, and why does AI suddenly depend on it?

Indium phosphide is a III-V compound semiconductor with high electron mobility and a direct bandgap — the core material for laser diodes (EMLs, CW lasers) and high-speed photodetectors. Traditional data centers move electrical signals over copper, but at 800G, 1.6T, and 3.2T bandwidths, copper hits its limits on power and distance. The industry is shifting to optical interconnects: convert signals to light, run them through fiber, convert back.

That means every AI server interconnect now needs a laser source. And as modules move from pluggable form factors to co-packaged optics — where the optical engine sits right next to the switch chip — the demand for InP lasers only grows in both quantity and power. In short: the faster AI compute scales, the more "light" it needs — and light comes from InP.

Two headlines reveal the real story

First: Sivers Semiconductors (Sweden) and SemiNex (US) announced a $3.4 million joint development program for "next-generation InP light sources for AI data center interconnects and co-packaged optics" — spanning high-power external lasers, DFB laser arrays for wavelength multiplexing, and SOA gain stages. Customer sampling and early production are targeted for H2 2027, aligned with the CPO scale-up timeline. Second: Coherent stated on its earnings call that "indium phosphide capacity continues to be our primary constraint," revealing that June-quarter InP laser output grew roughly 80% year over year. The company is shifting from 3-inch to 6-inch wafers — roughly four times the output at about half the wafer cost, with better yields.

Applied Optoelectronics (AAOI), meanwhile, cited a 20–40% demand imbalance in transceivers, with lasers as the bottleneck. All three companies point to the same conclusion: the AI optics bottleneck is moving upstream — from module assembly to the light sources and InP material itself.

The ripple effects: from wafers to server lead times

InP is hard to fix quickly because the supply problem is structural: global InP epitaxy and wafer capacity is concentrated in a handful of players, and new capacity takes 12–18 months from investment to volume production — far longer than silicon wafers. AI server lead times are therefore stretching from "waiting for GPUs" to "waiting for GPUs, optical modules, and lasers." For operators, the shortage delays AI cluster buildouts; for investors, the key is separating real beneficiaries — companies with InP epitaxy, laser packaging, or CPO light-source capability — from stocks that merely carry an "optical networking" label.

There is no quick fix for the InP shortage. But the supply chain's response — 6-inch transitions, cross-company partnerships, locking in capacity early — makes the direction clear: AI's next step is bottlenecked by "a beam of light," and indium phosphide is what makes that beam possible.

FAQ

Q1: What is indium phosphide (InP), and why does AI need it?

InP is a III-V compound semiconductor that efficiently emits and detects light — the core material for optical-communication lasers (EMLs, CW lasers) and high-speed photodetectors. As AI data centers push from 800G to 1.6T/3.2T bandwidths and copper reaches its limits, optical interconnects take over, and every optical channel needs an InP laser.

Q2: Why is InP said to be "scarcer than memory"?

Memory (HBM, DRAM) capacity is spread across multiple large makers and expands relatively fast. InP epitaxy and wafer capacity is highly concentrated, with 12–18 month expansion lead times, while AI optical demand is exploding — the supply gap cannot close quickly.

Q3: How is CPO different from traditional optical modules?

Traditional modules are pluggable components inserted into switch faceplates. CPO packages the optical engine next to the switch chip, eliminating faceplate connections to cut power and latency — but it demands more lasers with higher power and reliability.

Q4: Is the $3.4M Sivers–SemiNex program significant?

The amount is modest, but the signal matters: high-power external lasers, DFB arrays, and SOA gain stages are becoming strategic positions for CPO-grade InP light sources, with production timing (H2 2027) aligned to CPO scale-up — a clear sign of early supply-chain positioning.

Q5: How can investors tell real optical-networking winners from hype?

Look at where a company sits in the chain: InP epitaxy/wafers, laser design and packaging, or CPO light-source capability means real exposure; a "photonics" label without products or orders is hype. Also track what companies call their bottleneck on earnings calls — profit tends to follow the constraint.