Foundry model

Based on Wikipedia: Foundry model

The Most Expensive Buildings in Human History

A modern semiconductor fabrication plant costs more than twenty billion dollars to build. That's not a typo. A single factory, making chips smaller than a human hair can see, requires more capital investment than most countries spend on their entire military budgets. The machinery inside must be so precise that the buildings themselves are designed to absorb earthquakes without transferring even microscopic vibrations to the equipment.

And here's the thing: these astronomical investments only make sense if the factory runs at nearly full capacity, around the clock, for years on end. Let a fab sit idle for even a few weeks, and you're hemorrhaging money faster than almost any other business on Earth.

This economic reality created one of the most consequential business model innovations of the past half-century: the foundry model. It's the reason your smartphone exists, the reason artificial intelligence became practical, and increasingly, the reason geopolitics keeps semiconductor supply chains on the front page.

How Chips Used to Be Made

For the first few decades of the semiconductor industry, if you wanted to design a computer chip, you also had to manufacture it. Companies like Intel, Texas Instruments, and Motorola did everything in-house. They researched new manufacturing techniques. They designed new chip architectures. They built and operated their own fabrication plants. They were what the industry calls Integrated Device Manufacturers, or IDMs.

This made a certain kind of sense. Chip design and chip manufacturing are intimately connected. The way you design a circuit depends on what manufacturing processes are available. The way you manufacture depends on what designs you're trying to build. Keeping both under one roof meant engineers could optimize across the entire chain.

But it also meant that only very large, very wealthy companies could play the game. Building a fab required hundreds of millions of dollars even in the 1980s. A startup with a brilliant chip design but no factory? Out of luck. A small company that only needed a few thousand chips? Not worth anyone's time to manufacture.

Morris Chang's Radical Idea

In 1987, a semiconductor executive named Morris Chang did something that seemed, at the time, slightly crazy. He founded a company in Taiwan that would only manufacture chips. It wouldn't design any products of its own. It would simply take other companies' designs and turn them into physical silicon.

The company was called Taiwan Semiconductor Manufacturing Company, or TSMC. It was the world's first dedicated, pure-play semiconductor foundry.

The word "foundry" comes from metalworking, where it describes a facility that pours molten metal into molds to create castings. The semiconductor industry borrowed the term because the basic idea is similar: you bring your design, the foundry turns it into a physical product. But instead of molten iron, you're dealing with silicon wafers, ultraviolet light, and manufacturing tolerances measured in nanometers—billionths of a meter.

Chang's insight was that the economics of chip manufacturing had reached a tipping point. Fabs were so expensive that even major companies couldn't keep them fully utilized with their own designs alone. Meanwhile, a growing ecosystem of small design companies was hungry for manufacturing capacity they couldn't afford to build themselves.

What if one company specialized entirely in manufacturing, serving many design companies? The foundry could keep its fabs running at full capacity by aggregating demand from dozens or hundreds of customers. The design companies could focus purely on innovation, without the crushing capital requirements of building their own factories.

The Great Unbundling

TSMC's founding triggered a fundamental restructuring of the semiconductor industry. Over the following decades, chip-making split into three distinct categories.

First, the traditional Integrated Device Manufacturers persisted. Intel continued designing and manufacturing its own processors. Samsung maintained both design and fab capabilities. Texas Instruments kept its integrated structure. These companies believed the tight coupling between design and manufacturing gave them advantages worth the capital intensity.

Second, a new category emerged: pure-play foundries like TSMC. These companies invested billions in manufacturing expertise and capacity while deliberately staying out of the chip design business. GlobalFoundries, United Microelectronics Corporation (UMC), and Semiconductor Manufacturing International Corporation (SMIC) joined TSMC in this model. By not designing their own products, they could promise customers something valuable: we will never compete with you.

Third, and perhaps most revolutionary, came the fabless semiconductor companies. These firms designed chips but owned no manufacturing capacity whatsoever. They outsourced all fabrication to foundries.

The fabless model unleashed an explosion of innovation. Suddenly, a small team with a good idea could design a chip and have it manufactured without raising billions in capital. AMD, once an integrated manufacturer, sold its fabs and went fabless. Nvidia, the company now at the center of the artificial intelligence revolution, has never owned a factory. Qualcomm, whose technology lives in billions of smartphones, is fabless. Apple designs its own chips for iPhones and Macs but manufactures none of them—TSMC does.

Why Dedicated Foundries Win

Earlier attempts at foundry services had existed before TSMC, but they operated differently. A service called MOSIS, launched in 1981, let university researchers and small startups manufacture designs by piggybacking on commercial companies' excess capacity. When a conventional chip manufacturer had spare room in a production run, MOSIS could slip in some wafers for outside customers.

This worked, but barely. The manufacturing was always secondary to the foundry's main business of making its own products. Customer service was an afterthought. Worse, fabless companies worried constantly that the foundry might steal their designs, or that the foundry's own product divisions might use manufacturing priority as a competitive weapon.

TSMC's dedicated model solved these problems elegantly. Because TSMC designed no products of its own, customers knew their designs were safe. There was no internal product division that might get priority over outside customers. The foundry's entire business depended on making customers successful, so customer service became a core competency rather than an afterthought.

TSMC also pioneered what's called "customer-owned tooling," or COT. Previous foundries often required customers to use proprietary design tools, locking them into that specific manufacturer. TSMC let customers use industry-standard design software, making it easier to switch foundries if needed. This sounds like it would hurt TSMC's business, but the opposite happened: customers felt safer committing to TSMC precisely because they weren't locked in.

The Economics of Scale

The foundry model's advantages compound as technology advances. Each new generation of semiconductor manufacturing is roughly twice as expensive as the previous one. The latest cutting-edge fabs cost over twenty billion dollars each. At these price points, even giant companies struggle to justify dedicated facilities.

Consider the math. If a fab costs twenty billion dollars and needs to run at ninety percent capacity for ten years to pay for itself, you need to manufacture roughly two hundred billion dollars worth of chips. Very few companies have that much demand for their own designs. But a foundry aggregating demand from Apple, Nvidia, AMD, Qualcomm, and dozens of other customers? Suddenly the economics work.

This is why, paradoxically, the most advanced chips in the world are no longer made by traditional integrated manufacturers. Intel, once the undisputed leader in manufacturing technology, has fallen behind TSMC. Samsung runs foundry services but hasn't matched TSMC's leading edge. The pure-play model, freed from the distraction of product design and powered by aggregated global demand, has proven better at pushing the boundaries of physics.

The Geopolitical Dimension

The foundry model's success created an unexpected concentration of strategic importance. TSMC manufactures the majority of the world's most advanced semiconductors. Nearly all cutting-edge artificial intelligence chips run on TSMC silicon. Most smartphone processors come from TSMC fabs. The company's facilities in Taiwan have become arguably the most strategically significant factories on Earth.

This concentration makes governments nervous. If something disrupted TSMC's operations—a natural disaster, a conflict over Taiwan, even a pandemic-driven logistics breakdown—the ripple effects would cascade through every technology-dependent industry. Cars, phones, medical devices, military systems, data centers: all depend on a handful of facilities on a single island.

In response, nations are racing to build domestic foundry capacity. The United States passed the CHIPS Act, offering tens of billions in subsidies to build fabs on American soil. The European Union has similar programs. China has invested massive sums in SMIC and other domestic foundries, though they remain generations behind the leading edge.

TSMC itself is building new fabs in Arizona and Japan, hedging against the concentration risk that has made it simultaneously indispensable and vulnerable.

The Intel Question

The foundry model's triumph poses an existential question for Intel, historically the world's most important semiconductor company. For decades, Intel succeeded by tightly integrating design and manufacturing. Its engineers could optimize across the full stack in ways competitors couldn't match. The company's manufacturing prowess was legendary.

But Intel stumbled on the transition to smaller transistor sizes, falling behind TSMC and Samsung. Suddenly the integrated model looked less like an advantage and more like a liability. Intel was spending billions maintaining fabs that weren't producing the most advanced chips.

Intel's response has been to embrace the foundry model from both sides simultaneously. It has started using TSMC to manufacture some of its own chip designs—an admission that its internal fabs can't match TSMC's capabilities for every product. More radically, Intel is trying to become a foundry itself, offering manufacturing services to other companies' designs.

This is a profound strategic bet. Intel is essentially trying to become a full participant in the foundry ecosystem after decades of going it alone. The company's recent emphasis on its "18A" and "14A" manufacturing processes reflects this ambition—those designations indicate Intel's roadmap for regaining manufacturing leadership, with the goal of attracting foundry customers who currently rely on TSMC.

Whether Intel can transform its culture and operations fast enough remains one of the great open questions in technology.

The Hidden Complexity

Behind the clean categories of "fabless," "foundry," and "IDM" lies a messier reality. Many companies operate hybrid models. Samsung is both a foundry serving outside customers and an IDM making its own memory chips and smartphone processors. AMD is mostly fabless but maintains some specialized manufacturing for certain products. Apple designs its own chips but also buys standard components from traditional IDMs.

Intellectual property flows in complicated ways through this ecosystem. Foundries don't just run manufacturing equipment; they develop proprietary processes, design rules, and libraries that customers rely on. Chip designers don't just hand over blueprints; they license technology from companies like ARM that provide standard processor architectures. Patent cross-licensing agreements crisscross the industry, creating a web of dependencies that lawyers spend careers untangling.

Security and trust remain ongoing concerns. When you send your chip design to a foundry, you're trusting them with potentially your most valuable intellectual property. Foundries go to enormous lengths to demonstrate trustworthiness—TSMC's customer isolation procedures are legendarily rigorous. But the possibility of design theft or espionage never fully disappears, and it influences which foundries companies are willing to use.

What the Foundry Model Enabled

It's worth stepping back to appreciate what the foundry model made possible. The smartphone in your pocket contains chips from a dozen different companies, all manufactured on the same TSMC production lines. The diversity of the semiconductor industry—thousands of companies pursuing specialized niches—only exists because foundries let them manufacture without billion-dollar capital requirements.

The artificial intelligence boom runs on GPUs designed by Nvidia, a fabless company, and manufactured by TSMC. Without the foundry model, Nvidia would have needed to build its own fabs to pursue its vision of graphics processors repurposed for neural network training. That would have required a fundamentally different company with fundamentally different capabilities and risk tolerances.

Even Apple's strategic pivot to designing its own chips depended on foundries. When Apple decided it could build better processors than Intel for its Macs, it didn't need to build factories. It just designed the chips and handed the manufacturing to TSMC, the same foundry making iPhone processors.

The foundry model turned semiconductor manufacturing from a prerequisite for chip design into a service you could purchase. That seemingly simple shift enabled a generation of innovation.

The Future

Semiconductor manufacturing is approaching physical limits that may reshape the industry again. Transistors are now so small that quantum effects create fundamental challenges. Each generation requires new materials, new techniques, and exponentially more expensive equipment. Extreme ultraviolet lithography machines—essential for the latest processes—cost over one hundred fifty million dollars each and are made by exactly one company in the Netherlands.

These escalating costs put pressure on the foundry model from multiple directions. Only TSMC and Samsung can afford to pursue the absolute cutting edge. Smaller foundries like GlobalFoundries have explicitly abandoned the leading-edge race, focusing instead on less advanced but still profitable processes. The consolidation could leave just one or two companies capable of manufacturing the most advanced chips.

New architectures may provide an escape valve. Chiplets—small specialized chip modules combined into larger packages—let designers mix different manufacturing processes in a single product. Advanced packaging techniques can connect chips manufactured at different foundries. These approaches could sustain innovation even if the traditional path of shrinking transistors reaches its end.

What seems certain is that semiconductor manufacturing will remain too expensive and too strategic for individual companies or nations to go it alone. The foundry model emerged because the economics demanded specialization. Those same economics, intensified, will continue to shape whatever comes next.

Morris Chang's insight from 1987—that chip design and chip manufacturing could be profitably separated—has proven to be one of the most consequential business model innovations of the information age. Every device you use, every cloud service you depend on, every artificial intelligence system making headlines builds on that foundation. The foundry model is invisible infrastructure for the modern world.