Where the money is in a printed circuit board

A printed circuit board is four industries stacked on top of each other, and only one of them has pricing power. It is not the one that makes the board. What follows comes out of a World Economy Graph query on the PCB chain — graph structure, live company financials, and web research — and the useful result was not the ranking of suppliers but the discovery that the tier everyone names is the tier that earns least.

Copper is a dependency, not a chokepoint

Every PCB in the graph carries a REQUIRES edge to copper. The edge is real and almost useless: its specificity is 0.019, the lowest of any input into the board, because copper is required by nearly everything else in manufacturing too. A dependency shared by all of industry tells you nothing about who wins inside one chain of it.

The edges that discriminate sit further down the volume ranking and much higher on specificity: fiberglass at 0.447, tin at 0.412, nickel at 0.295, epoxy resin at 0.292. This is a habit worth keeping when reading any dependency graph — the loudest edge is usually the least informative one, and the question to ask of a dependency is not how large it is but how concentrated it is on this chain rather than on all of them.

The four tiers

Laminate. The real material chokepoint is copper-clad laminate: glass fabric impregnated with epoxy resin and bonded to copper foil. It is not a commodity, because its dielectric properties decide whether a board can carry high-speed AI-server signalling at all. Nan Ya Plastics and Kingboard Holdings both produce it directly, and both make their own copper foil and epoxy resin as captive feedstock instead of buying on the open market. That backward integration is the moat. A fabricator without captive laminate is a price-taker to six or seven firms — Nan Ya, Kingboard, Isola, ITEQ, Ventec, Panasonic — and in a tight demand cycle those six set the price for everyone downstream.

Bare board fabrication. The largest tier by volume and the most commoditised. Kinwong, Shennan Circuits, Dongshan Precision, Bomin, Suntak and dozens of others produce standard multilayer boards on essentially undifferentiated process technology; margins reflect it. This is a scale-and-utilisation business, not an IP business. One step up is HDI and flexible PCB, where Zhen Ding is the clearest name — Apple's key flex supplier, and the highest-specificity customer link in the tier at 0.338. Its 22.4% gross margin on TWD 193.4bn of trailing revenue is meaningfully better than a commodity fabricator's, because fine-pitch lines and bending reliability are harder to replicate than a standard multilayer stack.

IC and ABF substrate. The narrowest tier, and the one that sits directly under the AI accelerator. A substrate is not a board the chip sits near; it is the interposer the package is built on, and it needs sub-10-micron line and space control that only a few facilities in the world can hold. Ibiden, Unimicron and Shinko are reported to hold roughly 75–85% of global ABF substrate capacity between them, and the ABF film itself comes almost entirely from Ajinomoto, put at around 95% share of that one input. A monopoly inside a triopoly inside an oligopoly is a rare structure, and it explains how the tier swung from oversupply in 2023 to a reported 5–8% shortage a year later, with Nvidia, AMD and the hyperscalers booking capacity years ahead.

Assembly. Hon Hai and Quanta both depend on the finished board at 0.99 confidence, marked critical — unremarkable for the world's largest contract assemblers, but it is the mechanism by which a shortage three tiers down turns into a slipped server shipment. Fabrinet is the sharper example: it supplies Nvidia directly, which means part of the optical-module and server-assembly work behind Nvidia's AI systems runs through this tier rather than through Nvidia.

The margin gap is the evidence

Ibiden and Unimicron are both substrate names. The thirteen-point gap between them is a difference of mix, not of category: Ibiden skews to advanced package substrate for AI and data-centre customers, Unimicron carries a diversified book of standard board, HDI and substrate. That is the cleanest quantified statement available about this chain — the scarcity is in the substrate, not in board manufacturing generally.

Scale inside the narrow tier is just as skewed. AT&S, Apple's IC substrate supplier, is roughly two orders of magnitude smaller than Ibiden by market capitalisation. It follows prices; it does not set capacity.

Geography, and a disagreement worth keeping

Estimates of 2024 production value put China at 33–35%, Taiwan at 28%, Japan at 13–14% and South Korea at 9–10% — but two industry sources diverge on the exact shares, because one counts company headquarters and the other counts production location. That is a measurement disagreement rather than noise, and it is worth carrying forward unresolved instead of averaging away: it decides whether "China's PCB industry" reads as a third of the world or as barely that.

What is not in dispute is the split by tier. Taiwan and Japan dominate advanced substrate; China dominates commodity volume. So a shock that hits Chinese PCB mostly hits the low-margin end of the chain, and a shock to Taiwan or Japan hits the chokepoint.

What would break this

The case for the advanced tiers rests on AI-server demand outrunning substrate capacity through 2026–2028. That is a forecast, not a measurement, and it can break in either direction. Capacity is being added — Ibiden, Unimicron and Shinko are all reported to be expanding ABF lines — and Ajinomoto's near-monopoly on the film is a single point of concentration that a large enough capex cycle, or a Chinese entrant, could erode over three to five years.

The geopolitical risk most discussed is also the wrong one. China's 2025 export controls cover rare earths, battery materials and graphite, not PCB inputs; on current evidence they do not touch this chain. The live risk is ordinary cyclicality. Fabrication is capital-intensive and prone to the same boom-bust utilisation swings as semiconductors, and a slowdown in hyperscaler capex would hit the HDI and substrate premium first — precisely because that is where the excess demand sits.

What follows

"PCB" is the wrong unit of analysis. Buying it gets you the tier with the least pricing power, because that is where most of the volume and most of the listed names are. The part worth owning is two layers down, has three credible suppliers, and depends on one company's film.