Lesson 7 — Who Does What? Why? Who Pays?
How Do Chips Actually Work?
Learning Material
1 pagesLesson 7 — Who Does What? Why? Who Pays?
Understanding the Complex: How Do Chips Actually Work?
When Jensen Huang — born in Taiwan, raised partly in Thailand, educated at Oregon State and Stanford — founded Nvidia in 1993, he had a clear vision: build chips for computer graphics. What he didn't plan for was that the architecture he developed for rendering 3D video game graphics would, decades later, become the most sought-after hardware in the world for training artificial intelligence.
Huang's career trajectory illustrates something important about how the semiconductor industry is organized: not as a single entity, but as a highly specialized ecosystem of companies, each occupying a distinct role, dependent on the others, rarely overlapping. Understanding who does what — and why the division of labor evolved as it did — is essential to understanding why changing the industry's geography is so difficult.
Three models: fabless, IDM, and foundry
The semiconductor industry today is organized around three business models that emerged over the past four decades.
Fabless companies design chips but own no manufacturing equipment. They outsource all production to contract foundries. Nvidia is the most prominent example, but the category includes Qualcomm (which designs chips for smartphones), AMD (processors and graphics), Apple (the M-series and A-series chips in its devices), Broadcom, Marvell, and hundreds of smaller designers. The fabless model concentrates resources on design — the primary source of value and intellectual property — while avoiding the enormous capital costs of operating a fab.
Integrated Device Manufacturers (IDMs) both design and manufacture their own chips. Intel and Samsung are the two most prominent IDMs, though Samsung also operates as a foundry for other customers. The IDM model gives a company tight control over the design-manufacturing interface and historically allowed companies to optimize both together. Its disadvantage is the enormous capital investment required: a modern fab costs $10–20 billion to build and several billion per year to operate.
Foundries manufacture chips designed by others and do nothing else. TSMC is by far the largest, but GlobalFoundries, SMIC, and UMC also operate at significant scale (though all are behind TSMC at the leading edge). The foundry model requires absolute neutrality with regard to customers — a foundry cannot favor one customer's design over another's or use information from one to help another — and builds its competitive moat through manufacturing expertise rather than design.
The role of Electronic Design Automation (EDA)
Before a chip can be manufactured, it must be designed — and the complexity of designing a modern chip with billions of transistors exceeds anything that could be done by hand. This has given rise to a specialized software industry: Electronic Design Automation.
EDA software allows chip designers to specify circuits at a high level of abstraction, simulate their behavior, optimize their layout, and prepare the design for manufacturing. The EDA industry is dominated by three companies — Synopsys, Cadence, and Mentor (now part of Siemens) — that collectively provide the tools without which modern chip design is impossible.
The concentration in EDA has its own geopolitical implications. US export controls have restricted Chinese companies' access to advanced EDA software, imposing significant constraints on China's ability to design leading-edge chips even if it could manufacture them.
Government: the invisible funder
The semiconductor industry did not emerge spontaneously from free markets. It was, from its earliest days, substantially funded by government — specifically by the US military.
The integrated circuit was developed in part because the US military wanted smaller electronics for missiles and spacecraft. DARPA — the Defense Advanced Research Projects Agency — funded much of the foundational research in semiconductor manufacturing during the 1970s and 1980s, including the Sematech consortium that helped US companies retain manufacturing competitiveness.
Government involvement has intensified in recent years. The US CHIPS and Science Act of 2022 allocated $52.7 billion in subsidies to domestic semiconductor manufacturing and research, with the goal of rebuilding US fab capacity. Taiwan has for decades provided subsidized electricity, tax incentives, and infrastructure to TSMC. South Korea has massive subsidy programs for Samsung and SK Hynix. Japan, the European Union, and China all run comparable programs.
The EU Chips Act, passed in 2023, commits €43 billion to expanding European semiconductor manufacturing capacity from roughly 8 percent of global production to 20 percent by 2030. Whether that goal is achievable remains an open question — the target is ambitious given the lead times for new fab construction.
IMEC: the research institution you've never heard of
Located in Leuven, Belgium, IMEC — the Interuniversity Microelectronics Centre — is one of the world's leading semiconductor research organizations. It serves as a pre-competitive research hub where chip companies, equipment makers, and materials suppliers collaborate on technologies that are too expensive for any single company to develop alone.
IMEC is unusual in that it works with both US and Asian companies, maintains relationships across geopolitical lines, and focuses on research rather than production. When companies want to explore whether a new manufacturing process might work — before committing to the cost of a full development program — they often do so at IMEC. It occupies a critical niche: neutral ground in an increasingly contested industry.
The labor question
One dimension of the semiconductor industry that receives less public attention than geopolitics is labor. Operating an advanced fab requires an extraordinary concentration of skilled workers: process engineers, equipment technicians, materials scientists, quality engineers. TSMC employs roughly 60,000 people in Taiwan. Training this workforce took decades.
When TSMC opened a new fab in Arizona — partly in response to US government incentives — it encountered significant challenges recruiting enough skilled workers locally. The company ultimately brought in several hundred workers from Taiwan to staff the facility during its initial operating period, a move that generated some political friction but reflected the simple reality that the expertise required cannot be instantly trained from scratch.
Next lesson: What's Contested? — the four major debates in the chip industry, from Moore's Law to geopolitics to the future of RISC-V.
Reading time: approx. 9–10 minutes