A semiconductor foundry is a company that manufactures chips designed by other companies, rather than designing and selling chips of its own. One foundry — Taiwan’s TSMC — makes almost every advanced AI chip in the world, from Nvidia’s GPUs to Apple’s processors, which is why a single company’s factories effectively set the pace of the entire AI industry.
Design and manufacturing are different businesses
Making a chip involves two very different jobs: designing the circuit and physically fabricating it in a factory called a “fab.” Historically, companies like Intel did both — a model known as an integrated device manufacturer, or IDM. Building and running a leading-edge fab costs tens of billions of dollars, so most chip designers eventually gave that up.
That split created the foundry model: “fabless” companies — Nvidia, AMD, Apple, Qualcomm — design chips but own no factories. “Pure-play” foundries — TSMC, GlobalFoundries, UMC, Samsung’s foundry division — own the factories and manufacture chips for many different customers, competing with none of them because they never sell a finished product under their own brand. Nvidia is the clearest example of a fabless AI chip designer that depends entirely on outside foundries.
Why one company ended up dominating
Fabricating a chip at the leading edge — the smallest, densest transistor geometries, currently around 2-3 nanometers — requires machines that cost hundreds of millions of dollars each, years of process engineering, and enormous manufacturing scale to be profitable. Very few companies can clear that bar. TSMC holds roughly 70% of the global pure-play foundry market overall, and an even larger share — commonly estimated above 90% — of the most advanced nodes that go into AI accelerators, meaning it is close to the sole source for the chips inside frontier AI systems.
The bottleneck is packaging, not just fabrication
Fabricating the silicon wafer is only half the job. Modern AI chips need the processor and its high-bandwidth memory stacked and wired together with extreme precision, a step called advanced packaging. TSMC’s version, called CoWoS (“chip-on-wafer-on-substrate”), builds the microscopic bridge that lets a processor and its memory exchange data fast enough for AI workloads. Even a perfectly fabricated chip is unusable until it passes through this packaging step, and TSMC controls most of the world’s capacity to do it — a constraint that has become as important to AI chip supply as fabrication itself.
Why the concentration matters
Because so much of the world’s advanced chip supply runs through one company’s factories, mostly located on one island, any disruption — a natural disaster, a geopolitical crisis, or simply not enough packaging capacity — can slow AI hardware production everywhere at once. That risk is a major reason the United States, the European Union, Japan, and others have been funding domestic chip factories and pushing companies to build AI inference chips and other custom silicon, even though none of these efforts is expected to rival TSMC’s leading-edge capacity for years. It is also why AI hardware sits inside broader debates over AI compute shortages and AI export controls — the questions of who can build advanced chips, and who is allowed to buy them, are really questions about foundry capacity.
In the news
TSMC’s position at the center of the AI supply chain was on display when the company posted record June revenue as AI chip demand continued to outstrip what its factories could supply.
FAQ
Is TSMC the only foundry that can make advanced chips? No, but it is by far the largest. Samsung and Intel also operate leading-edge foundries, and Intel has been opening its fabs to outside customers, but neither has matched TSMC’s yields or scale at the most advanced nodes.
Why doesn’t Nvidia just build its own factories? A leading-edge fab costs tens of billions of dollars and takes years to reach high yields. It’s cheaper and faster for a fabless designer to buy capacity from a specialist than to build and operate one.
Does “foundry” always mean the most advanced chips? No — foundries also make older, cheaper chip generations used in cars, appliances, and industrial equipment. The AI-driven attention on foundries is specifically about capacity at the leading edge.
Sources: TSMC (Wikipedia), Foundry model (Wikipedia), TSMC quarterly revenue disclosures and foundry market-share reporting.