The word everyone uses and rarely defines
“Foundry” appears in nearly every article in this cohort, usually without explanation, on the assumption a reader already knows what it means. Most readers have a rough sense — “the company that makes the chips” — that is close enough for casual reading and not quite precise enough to understand why TSMC, Samsung Foundry, Intel Foundry, GlobalFoundries, UMC, and SMIC are all doing recognizably the same job while competing so differently. This briefing is the reference piece: read it alongside any of the company-specific foundry briefings in this cohort.
The core distinction: who owns the design
A semiconductor foundry manufactures chips designed by other companies, for a fee, without owning the design itself. The customer — Nvidia, AMD, Apple, Qualcomm, or any of dozens of others — hands over a finished digital design file. The foundry’s job is purely to turn that file into working physical silicon: patterning it onto wafers, testing it, and (often through a partner) packaging it. The foundry has no stake in whether the chip succeeds in the market beyond getting paid for manufacturing it, and typically has no visibility into what the chip is even for beyond what it needs to know to manufacture it correctly.
This is different from an integrated device manufacturer, or IDM — a company like classic-era Intel or Texas Instruments, which designs its own chips and manufactures them in its own fabs, for its own products, from end to end. An IDM’s fab exists to serve one company’s roadmap. A foundry’s fab exists to serve whichever customer’s design is currently loaded onto it — this week Nvidia’s, next week a startup’s, with no obligation that the two ever have anything in common.
Why this was considered a strange idea in 1987
When Morris Chang founded TSMC in 1987 specifically as a pure-play foundry — manufacturing exclusively for others, designing nothing of its own — it did not fit the industry’s existing mental model [1]. At the time, IDMs like Intel and NEC only sold spare fabrication capacity to outside customers opportunistically, when their own production had room to spare. A company built from the ground up around serving outside customers, with no in-house product line competing for the same capacity, had no clear precedent, and its viability was genuinely uncertain at founding [2].
The consequence: an entire industry that doesn’t own factories
The pure-play foundry model is the structural precondition for the fabless design industry — companies that design chips and never build a single fab of their own. Nvidia, AMD, Qualcomm, Broadcom, and Apple’s silicon design team are all fabless in this sense: their entire business depends on a foundry being willing to manufacture a design for a fee without wanting a stake in its success, and without treating another customer’s competing design as a conflict of interest [3]. Without TSMC’s founding bet, the AI accelerator market as it exists today — dozens of companies, each designing distinct silicon, none of them needing to build or own a fab — would not be structurally possible in its current form.
The layer most explainers skip: packaging
A modern chip’s journey does not end when the wafer leaves the foundry. Advanced packaging — integrating a finished logic die with memory stacks, substrates, and interconnects into a sellable product — is often handled by a separate category of company called an OSAT (outsourced semiconductor assembly and test), such as ASE Technology or Amkor, rather than by the foundry itself, though TSMC increasingly does both [4]. Readers tracking this cohort’s packaging-specific briefings should hold the foundry/OSAT distinction in mind: a chip can leave a foundry finished and still not be a sellable product until a separate company, in a separate factory, finishes assembling it.
Why this explainer matters for everything else in this cohort
Every company profile in this cohort’s foundry track — TSMC’s dominance, Intel’s struggle to find an external customer, Samsung’s dual identity, UMC’s deliberate positioning, SMIC’s yield constraints — is a variation on the same underlying question: how well does a given foundry execute the pure-play model TSMC invented, and how much does a given customer trust that execution enough to bet a product line on it. Understanding the foundry concept itself is what makes the rest of those stories legible as one connected industry rather than a list of unrelated company facts.
