The story everyone tells about AI hardware is a chip story: more GPUs, faster GPUs, who gets them. But walk down the supply chain and the real constraints aren't the famous chips — they're the boring inputs underneath, the materials that feed the racks. Most of them are made by a handful of companies, on lines that are already sold out, and the AI build-out is pulling on all of them at once. This is a map of those choke points, organized by layer, with the one number that matters for each.
Layer 1 — the parts that move the light
Inside a data center, data between machines increasingly travels as light over fiber, through optical modules. Four of the parts that make and protect that light are badly supply-constrained.
- EML chips. An EML (electro-absorption modulated laser) is the core light-making chip in a high-speed optical module — it sets the speed, and it's what 800G and 1.6T (gigabit/terabit-per-second) modules are built around. Supply is concentrated in a few foreign players (Lumentum, Coherent), the process is hard to scale, and demand has run away from it: Sumitomo's EML output is on the order of 20 million chips, against demand that leaves roughly a 30% gap. Prices have moved accordingly — 40–50% on many parts, up to 80% on some. The sharpest version is 200G EML, the grade 1.6T modules are built from: each module needs at least eight of them, and 2026 supply of roughly 50 million 200G chips covers only about 7 million 1.6T modules against demand north of 25 million — so silicon photonics has to fill more than 70% of the gap. China's Source Photonics is the only domestic maker into a top North-American cloud supply chain, but still runs 2-inch wafers while the leaders move to 4/6-inch.
- CW light sources. A CW (continuous-wave) laser — a high-power DFB laser that emits a steady beam — is the light supply for silicon- photonics modules and CPO (co-packaged optics, where the optics sit right next to the switch chip). As silicon photonics becomes the mainstream design, demand is rigid and rising, and the leaders' capacity is locked: Lumentum's is spoken for by core customers (including Nvidia) out to 2028. The price crept from about ¥3.5 a unit to ¥4–5, and the market is expected to stay tight for three to five years.
- Pump lasers. A pump laser drives the optical amplifier (EDFA) that boosts a signal for long-distance hops — the heart of a DCI (data-center-interconnect) link. Lumentum and Coherent together hold over 90% of the market, a clean duopoly. Their lines were booked through the first half of 2027 by late 2025, with customers now trying to lock 2027–2029; planned expansion of 70–100% still trails demand. The gap runs over 30%, and pump-laser volume has grown around 80%.
- Faraday rotators. The magnetic crystal at the heart of an optical isolator — the part that stops back-reflected light from destabilizing the laser, so every module needs one. Supply is concentrated and, unusually, weaponized: Coherent holds about 50% (~50,000 pieces/month), a Japanese joint venture is next (~30,000/month), and Chinese makers are far smaller. Coherent treats the big module makers (Innolight, Eoptolink) as direct rivals and sells rotators on a “resource-swap” basis — mostly to firms that can trade something back, like EML chips. Price ran from about $120 a piece in 2023 to $175 in 2025 (+40%) at a 70–80% gross margin, and global capacity only climbs from ~100,000 to 160–180,000/month in the second half of 2026 at the earliest.
Layer 2 — the boards everything sits on
Under the chips is the printed circuit board, and the laminate it's pressed from. At AI signal speeds the board itself becomes a bottleneck, and its raw materials are short.
- Low-DK glass cloth. The reinforcing fabric inside high-end copper-clad laminate. ("Low-DK" = low dielectric constant — it loses less signal at high frequency.) AI servers need it, so the big fiberglass makers shifted capacity to these high-end grades, squeezing everything else. The second-generation cloth (the kind that goes into M8-class material) has been tight since January 2026 and stays tight all year. The real chokepoint is the top grade — “Q-cloth” (third-generation quartz cloth) for M8/M9 laminate, where nearly 100% of 2026 output is earmarked for Nvidia’s Rubin high-end and supply comes from a handful of makers (Feilihua, Sinoma / Taishan). Even the second-generation cloth feeding Google and Meta servers runs short: ~3 million meters/month of 2026 demand against roughly 1.2–1.3 million meters/month of Chinese capacity.
- HVLP4 copper foil. An ultra-low-profile copper foil — very smooth, so high-frequency signals lose less energy crossing it. It's hard to make, yields are unstable, and only AI-class boards need it. The projected supply gap is 48% in 2026 and 43% in 2027. In raw tonnage: 2026 demand tops 1,200 tons/month (driven by Nvidia Rubin and AWS Trainium) against only ~300–400 tons/month of yield-adjusted capacity; lead supplier Mitsui (~300 tons/month) hit a quality stumble, and converting a line to HVLP4 cuts its output by at least 30%.
- M8 / M9 CCL. Copper-clad laminate is the base material of the motherboard; M8 grade serves 800G/1.6T, M9 the next-generation (Rubin-class) platforms. It's hard to make — and worse, its two key inputs (the Low-DK cloth and HVLP foil above) are themselves short. Demand keeps climbing: monthly M9 demand in 2027 is projected at 5–6× 2026's M8 shipments.
Layer 3 — feeding the watts
AI racks draw enormous power, and the components that condition that power are getting squeezed.
- GaN and SiC power devices. Gallium-nitride and silicon-carbide chips switch power far more efficiently than ordinary silicon — they're what high-efficiency server power supplies (HVDC, PSU) and on-board voltage regulators (VRM) increasingly use. As chip power climbs, demand climbs with it, but capacity is gated by the substrate underneath (slow crystal growth, slow yield gains), and the devices still cost well above plain silicon.
- High-cap MLCC and IC substrates. MLCCs (multilayer ceramic capacitors) and IC substrates (the rigid carriers a chip is mounted on) do the filtering and power delivery around GPUs, CPUs and power modules. An AI server uses several times the capacitors a normal one does, high-end capacity is short, and expansion is slow. IC substrates have already taken two rounds of price increases since the start of 2026, with some high-end carriers up a cumulative 40% and another round under discussion.
- Power-management ICs. The unglamorous converters and drivers (Texas Instruments, Infineon, MPS) that regulate voltage across the server board. They aren’t AI parts at all — they’re mature chips shared with cars and industrial gear — which is exactly the problem: the AI surge is grabbing capacity off an existing market, so lead times have stretched from a normal 10–12 weeks to 30–40 weeks (eight to nine months), with no mature domestic second source to fall back on. It’s the weakest-plank bottleneck: the board can’t ship even when every exotic part has arrived.
Layer 4 — the materials beneath the chips
Go one layer deeper than the chips themselves and you hit raw materials with their own, harder bottlenecks.
- Indium-phosphide substrate. InP wafers are the base the optical chips above (EML, CW lasers) are grown on. The bottleneck is upstream: high-purity indium is a strategic metal, and China has restricted its export to Japan — which caps output at major suppliers like Sumitomo. The bar on purity and uniformity is high, so few firms can mass-produce it. Sumitomo holds about 40% of the global market, AXT is second, and Chinese producers (Yunnan Germanium among them) are expanding.
- Tungsten hexafluoride (WF6). A gas used to fill the tiny vertical holes and contacts in advanced chips — especially HBM (the stacked memory beside AI accelerators) and 3D NAND. Its demand scales directly with how tall the HBM stack and how many layers the NAND, supply sits with a few firms (Kanto Denka, SK Materials, Merck), and the contract price has climbed six quarters in a row.
The map
One table, the whole picture — the scarce inputs, where they go, and why they're stuck:
| Material | Used in | Why it's scarce |
|---|---|---|
| EML chip | high-speed optical modules | few suppliers, high tech barrier, demand boom (~30% gap) |
| CW light source | silicon photonics, CPO | capacity locked to 2028; tight 3–5 years |
| Pump laser | DCI links, optical amps | >90% duopoly; booked toward 2029 |
| Faraday rotator | optical isolators | ~50% one supplier; a competitive weapon; 70–80% margin |
| Low-DK glass cloth | high-end PCB laminate | capacity shifted to high grades; sold out |
| HVLP4 copper foil | AI/switch PCB layers | unstable yield; 48% / 43% gap (2026 / 2027) |
| M8 / M9 CCL | AI server motherboards | process barrier + its own inputs are short |
| GaN / SiC devices | server power, VRM | substrate-limited; demand tracks chip power |
| MLCC / IC substrate | board + power filtering, packaging | usage up sharply; slow to expand |
| Power-management IC | server boards (PCBA) | AI grabbing mature capacity; lead times 10–12 → 30–40 wks |
| InP substrate | EML / CW laser chips | upstream indium restricted; high purity bar |
| WF6 gas | HBM, advanced process | scales with HBM/NAND layers; concentrated supply |
The pattern
Read the table top to bottom and the same shape repeats at every layer: a part only a few firms can make, a process that's slow to scale, and AI demand arriving faster than anyone can add capacity. That's why this connects straight to the Nokia / DCI story — pump lasers and InP substrate show up on both maps. In a build-out gated by supply, the winners aren't whoever has the cleverest design; they're whoever owns the scarce input or locked it first. Each row above is a place where that's true right now.
An analysis piece: figures are drawn from supply-chain research (industry Q&A notes and sell-side reports through mid-2026) and stated as the thesis, not independently re-derived.