Foundry Model
The split of the chip industry into companies that design and companies that manufacture, which made modern semiconductors possible and concentrated them in Taiwan.
Nvidia does not make Nvidia chips. Apple does not make Apple chips. Qualcomm, Broadcom, AMD, and most of the AI industry own no factories at all. They design, and someone else manufactures. That division of labor is called the foundry model, and it is the organizing fact of the modern semiconductor industry.
Two ways to build a chip
For the industry's first thirty years, a chip company did everything. Intel, Motorola, Texas Instruments, and Japan's giants designed circuits and built them in their own fabrication plants, an arrangement now called the integrated device manufacturer, or IDM. It worked while a fab cost tens of millions of dollars. It stopped working when fabs began costing billions, because each new generation of manufacturing demanded volume that no single company's own product line could supply.
Morris Chang proposed the alternative in 1987: a pure-play foundry, a factory that takes orders from anyone and designs nothing itself. Because it never competes with its customers, a chip designer can hand it the crown jewels without fear. Because it aggregates demand from hundreds of designers, it can run the newest and most expensive process at full utilization. The mirror image is the fabless company, which owns no plant and spends its capital entirely on engineering. Between them sits a third layer most people never hear about: the electronic design automation tools from Synopsys, Cadence, and Siemens, and the licensable circuit blocks from firms like Arm, without which a small design team could not produce a working chip at all.
Why it concentrates
The model has a built-in tendency toward monopoly. Manufacturing at the leading edge has enormous fixed costs and near-zero marginal cost per wafer, so the foundry with the most customers can afford the most advanced tools, which attracts the next round of customers. Yield, the share of chips on a wafer that actually work, improves with volume and accumulated process knowledge that cannot be bought or copied quickly. The result is not a competitive market but a staircase, with fewer players standing on each higher step. Ten companies could build leading-edge logic in 2002. Today three can, and only one, TSMC, does it at scale for outside customers, with Samsung a distant second and Intel trying to enter the business it declined to invent.
That concentration is why chips became geopolitics. The choke points are not diffuse: one Dutch company, ASML, makes the extreme ultraviolet lithography machines; a handful of Japanese and American firms make the tools and chemicals; one Taiwanese company turns them into the processors that run the world's phones, data centers, and weapons. Export controls work at all because there are so few places to apply pressure. Reshoring is slow and expensive for the same reason: the CHIPS Act can fund an Arizona fab, but it cannot quickly reproduce the dense cluster of suppliers, engineers, and institutional habits that took Taiwan thirty years to build.
The split Chang made was a business decision about capital efficiency. It ended by putting the most advanced manufacturing on Earth in one place, and turning a commercial dependency into the most dangerous strategic question of the century.