A semiconductor foundry is a company that manufactures chips but doesn’t design them. It’s essentially a factory-for-hire: other companies create the chip designs, then the foundry handles the enormously complex job of turning those designs into physical silicon. TSMC, Samsung Foundry, and GlobalFoundries are the most prominent examples. This division of labor between designing and manufacturing is the backbone of the modern chip industry.
Why Foundries Exist
Building a chip factory, known as a fab, is one of the most expensive construction projects in any industry. Globally, semiconductor companies plan to invest roughly $1 trillion through 2030 in new fabs. A single leading-edge facility can cost tens of billions of dollars, and labor costs in the U.S. run four to five times higher per hour than in Asia. Very few companies can afford that kind of investment while also spending billions on chip design and research.
The foundry model solves this problem by splitting the industry into two camps. “Fabless” companies like Apple, NVIDIA, AMD, and Qualcomm focus entirely on designing chips. They come up with the architecture, the circuitry, and the specifications, but they own no factories. When the design is ready, they hand it off to a foundry, which manufactures the chips and ships them back. The fabless company then sells the finished product to device makers or consumers.
This split lowered the barrier to entry dramatically. A startup with a clever chip idea no longer needed to build a multi-billion-dollar factory before selling a single product. It could partner with a foundry that offered flexible production volumes, test the market at low cost, and scale up if the product succeeded. TSMC’s founder, Morris Chang, even maintained a policy of reserving production capacity specifically for startups, ensuring new fabless firms always had a chance to get off the ground.
Foundries vs. Integrated Manufacturers
The alternative to the foundry model is the Integrated Device Manufacturer, or IDM. An IDM designs, manufactures, and sells its own chips all under one roof. Intel has historically been the textbook example, though that’s changing (more on that below). Texas Instruments and some divisions of Samsung also operate this way.
The tradeoff is straightforward. An IDM controls every step, which can mean tighter optimization between design and manufacturing. But it also means carrying the staggering cost of running its own fabs. A pure-play foundry, by contrast, spreads that cost across dozens or hundreds of customers. Because it has no chip products of its own, a pure-play foundry has a strong incentive to invest heavily in manufacturing quality and capacity. Its entire revenue depends on keeping customers happy.
NVIDIA’s early history illustrates why this matters. In its early days, NVIDIA relied on SGS Thompson, an integrated manufacturer that also sold its own chips. When NVIDIA’s third product saw unexpected success, SGS Thompson couldn’t manufacture enough to meet demand. NVIDIA renegotiated and moved production to TSMC, where it has remained ever since. A dedicated foundry, with no competing product lines, was better positioned to scale production for a fast-growing customer.
How Chips Get Made in a Foundry
Chip manufacturing is split into two broad phases: front-end processing, where circuit patterns are created on a silicon wafer, and back-end processing, where the wafer is cut into individual chips and packaged for use in devices.
Front-end processing involves around 11 major steps, repeated and layered many times over. It starts with oxidizing the wafer’s surface and depositing ultra-thin films of material. A light-sensitive coating is applied, then exposed to precisely patterned light, a process called lithography. The exposed areas are chemically developed and etched away to create circuit features. Ions are implanted to alter the electrical properties of specific regions, surfaces are polished flat, and electrodes are added to connect everything. Each layer goes through inspection before the next one begins.
Back-end processing is comparatively simpler but still critical. The finished wafer, which may contain hundreds of individual chips, is sliced (diced) into separate pieces. Each chip is wired to its package connections, sealed in a protective mold, and given a final inspection before shipping.
The Equipment That Makes It Possible
The most critical and expensive tool in a modern foundry is the lithography machine. At the leading edge, this means extreme ultraviolet (EUV) lithography systems made by ASML, a Dutch company that is the sole supplier of these machines worldwide. EUV light has a wavelength short enough to print circuit features just a few nanometers wide, enabling the most advanced chips in production today.
Chipmakers use EUV systems to print the most intricate layers of their 7nm, 5nm, and 3nm chips. The remaining, less complex layers are printed using older deep ultraviolet (DUV) systems. By reducing the number of process steps required at each node, EUV lithography cuts defects, lowers costs, and shortens production time. If you own a recent smartphone, gaming console, or smartwatch, the processor inside was almost certainly made using EUV technology. ASML’s next-generation platform, called EXE, is expected to support high-volume manufacturing in 2025 and 2026, pushing chip scaling further into the next decade.
Why Yield Is Everything
In foundry manufacturing, “yield” refers to the percentage of chips on a wafer that actually work. A yield near 100% means almost every chip comes off the line functional. A low yield means the foundry is throwing away expensive silicon, and profitability drops fast.
The relationship between defects and yield is exponential, not linear. Once a foundry pushes its defect rate below a critical threshold (roughly 0.3 faults per square centimeter), yield skyrockets. Getting to that point requires a process called yield learning: systematically identifying and eliminating one source of defects after another until the vast majority of chips meet specifications.
The speed of this learning process matters more than almost any other factor in determining whether a foundry makes money on a new chip technology. Research from Portland State University found that moving the yield ramp up by six months more than doubles the cumulative profit of a semiconductor venture, while delaying it by six months eliminates two-thirds of the profit. Most of a foundry’s earnings on a given process node come during a brief window right after yield surges, when capacity is still constrained and the foundry can charge premium prices. This is why foundries invest so aggressively in getting new process nodes to high yield as quickly as possible.
Process Nodes and the Push to Smaller
Foundries compete largely on their ability to manufacture chips at smaller “process nodes,” measured in nanometers. Smaller nodes generally mean faster, more power-efficient chips that pack more transistors into the same area. Samsung’s second-generation 3nm process, for example, cuts power consumption by 50% compared to its 5nm technology while improving performance by 30% in a 35% smaller footprint.
At the 3nm node and below, foundries have moved to a new transistor design called gate-all-around (GAA), which wraps the transistor’s control gate around the channel on all sides instead of just three. This gives better control over electrical current at tiny scales, reducing leakage and improving efficiency. Samsung began mass production of its 3nm GAA process in 2022. TSMC has followed with its own versions. The race to 2nm and beyond is underway, with both companies and Intel targeting production in the coming years.
Intel’s Shift Into the Foundry Business
One of the biggest recent shifts in the industry is Intel’s move toward offering foundry services to outside customers. Historically a pure IDM, Intel announced its “IDM 2.0” strategy under CEO Pat Gelsinger, creating a standalone business unit called Intel Foundry Services (IFS). The goal: become a major provider of foundry capacity in the U.S. and Europe.
Intel’s pitch to potential customers includes access to its process technology and packaging capabilities, committed capacity in Western markets, and a portfolio of intellectual property spanning x86, ARM, and RISC-V chip architectures. At the same time, Intel itself plans to expand its use of third-party foundries like TSMC for some of its own products, blurring the old line between IDM and fabless models. Whether Intel can compete with TSMC and Samsung as a contract manufacturer remains one of the semiconductor industry’s open questions, but the attempt signals just how central the foundry model has become.
The Cost of Building Fabs Outside Asia
Governments in the U.S. and Europe are pushing to bring more chip manufacturing onshore, but the economics are challenging. According to McKinsey, even after subsidies, a standard mature logic fab built in the United States costs roughly 10% more to construct and carries up to 35% higher operating costs than a comparable facility in Taiwan. European fabs cost about the same to build as in Taiwan but significantly more than in mainland China. The gap comes primarily from construction and labor expenses, which remain two to five times higher in Western countries across every stage of the building process, from site preparation to tool installation.
These cost differences explain why Taiwan and South Korea dominate leading-edge foundry production today, and why the subsidies flowing from programs like the U.S. CHIPS Act and European Chips Act are so large. Reshoring chip manufacturing is as much a geopolitical strategy as an economic one.