The story everyone tells about AI is a chip story. But walk down the supply chain and the real constraints aren’t the famous chips — they’re the boring inputs underneath. This is the map of the choke points, layer by layer.
↓layer 1 — moving the light
two materials hold up the entire AI build
the lead supplier
locked to Rubin
retires drilling
the same two
upstream of all
EML chips. The core light-making chip in optical modules. Sumitomo ships ~20 million, demand leaves a 30% gap. Prices up 40–50%, up to 80% on some parts.
Lumentum & Coherent dominate.
200G EML is the sharpest version — each 1.6T module needs eight of them. 2026 supply of ~50 million chips covers only ~7 million modules against demand north of 25 million. Silicon photonics fills >70% of the gap.
= 7M vs 25M demand.
CW light sources. A continuous-wave DFB laser feeds silicon photonics & CPO. Lumentum’s capacity is locked through 2028. Price crept from ¥3.5 a unit to ¥4–5. Tight for three to five years.
booked to 2028.
Pump lasers. Drive the EDFA amplifier inside DCI links. Lumentum + Coherent hold over 90%. Booked through 1H 2027; customers now locking 2027–2029. Expansion of 70–100% still trails demand. Gap over 30%.
volume up ~80%.
Faraday rotators. Magnetic crystal at the heart of an optical isolator — stops back-reflected light destabilizing the laser. Coherent: ~50% (~50,000/month). Japanese JV: ~30,000/month. Price: $120 (2023) → $175 (2025), +40%, at 70–80% gross margin.
as rivals. resource-swap only.
→layer 2 — the boards
Low-DK glass cloth. Reinforcing fabric inside high-end laminate. (Low-DK = loses less signal at high frequency.) Top grade — “Q-cloth” — nearly 100% of 2026 output earmarked for Nvidia Rubin. Even Gen-2 cloth: demand ~3 million m/month vs 1.2–1.3 million capacity.
Rubin took it all.
HVLP4 copper foil. Ultra-low-profile copper — very smooth, so high-frequency signals lose less energy. Gap: 48% in 2026, 43% in 2027. 2026 demand >1,200 tons/month vs 300–400 tons/month yield-adjusted capacity. Mitsui line conversion cuts output by 30%+.
yield is unstable.
M8 / M9 CCL. Copper-clad laminate — the base of the motherboard. M8 = 800G/1.6T, M9 = Rubin. Worse, its inputs (Low-DK cloth + HVLP foil above) are themselves short. Monthly M9 demand in 2027 is projected at 5–6× 2026’s M8.
because its parts are short.
⚡layer 3 — the power
GaN & SiC power devices. Gallium-nitride / silicon-carbide chips switch power more efficiently than silicon — the parts in HVDC, PSUs, VRMs. Substrate-gated (slow crystal growth, slow yield gains).
crystal grows slow.
High-cap MLCC + IC substrates. AI servers use several times the capacitors a normal one does. IC substrates have taken two rounds of price increases since the start of 2026; high-end carriers up a cumulative 40%, another round under discussion.
expansion slow.
Power-management ICs. TI, Infineon, MPS — mature chips shared with cars and industrial. AI grabs capacity off an existing market. Lead times: 10–12 weeks → 30–40 weeks (8–9 months).
even when exotic parts arrive.
⌖layer 4 — beneath the chips
Indium-phosphide substrate. The wafer the EML & CW chips are grown on. Upstream indium is restricted by China’s export to Japan — caps Sumitomo. Sumitomo holds ~40% of the global market.
without InP wafers.
Tungsten hexafluoride (WF6). Gas used to fill tiny vertical holes in advanced chips — especially HBM (stacked memory beside AI accelerators) and 3D NAND. Contract price up six quarters in a row.
Kanto Denka, SK, Merck.
※the pattern
Same shape at every layer: a part only a few firms can make, a process slow to scale, AI demand arriving faster than anyone can add capacity. 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.