The most important tech companies are much larger than their original markets or what anyone imagined. There is something about them that defies anyone’s ability to estimate their market size. These companies not only expand their market, they seem to make their markets.
They do it by attacking the abstractions their industry is built on, not their competitors’ products.
They change what it takes to participate in the industry, creating shared infrastructure out of the core of their competitors’ businesses, and allowing an ecosystem of new companies to form that both win in the market and are dependent on them.
TSMC is an example of this.
The Bundle
Before TSMC, semiconductor companies had to be Integrated Device Manufacturers (IDMs). They designed chips and also owned the fabs that manufactured them. This was a burden but also their core strategic edge. Operating fabs was incredibly expensive and difficult to do efficiently. But both their cost and difficulty were also moats that suppressed competition.
A company could design a chip without owning a fab, but inevitably they’d have to ask an IDM to manufacture it. That meant partnering with a company whose primary business model was making competing chips. Capacity was unreliable since IDMs sought to fully utilize their fab capacity. So any capacity availability tended to be temporary. And any sufficiently successful customer would find themselves facing stiff competition in their former manufacturer, or more likely be acquired.
The prohibitive cost of operating a fab limited the set of companies that could design chips to a very small set of large conglomerates with low competition. The basis of competition was capital not innovation, and the economics tilted competitors towards conservatism.
Having a fab was the most important prerequisite to designing chips. The consequence of these pressures was that designing chips and owning fabs were bundled together. To enter one business you had to enter both.
TSMC unbundled them.
The Abstraction
TSMC flipped this industry logic, pioneering the pure-play foundry model. They built fabs and manufactured chips but only for other companies. They committed to not competing with their customers, and making being the foundry for semiconductor companies their core business. And they turned the fab into an interface, allowing designers to build for it without needing to know what happens inside it.
TSMC converted fabrication from the core capability required of every semiconductor company into shared infrastructure available to any chip designer.
Independent of clear market demand, Morris Chang and Taiwan had reason to pursue this business model. Taiwan had deep experience in wafer manufacturing but almost none in semiconductor design. So it was one of the few approaches to building in semiconductors they were well positioned for.
And there was reason to believe that a pure-play foundry model was both possible and attractive.
Design and manufacturing were tightly integrated in the IDM incumbents. Originally this was out of necessity, but Mead and Conway’s work on VLSI design methodology showed the two could be decoupled. Conway also demonstrated that distinct chip designs could be made on the same wafer. And Mead augured the rise of a fabless semis industry where innovators designed semiconductors while offloading the fabrication to a separate specialized company.
The IDMs delayed their adoption to protect their business models. The complexity of designing chips was growing beyond the ability to manually handle without Mead and Conway’s approach as the number of transistors per chip reached tens to hundreds of thousands.
This “orthogonalization of concerns” made it conceivable for TSMC to separate design from manufacturing and handle the latter for many disparate chip designs. Though this was and is far less trivial to implement in practice.
The rise of a pure-play foundry was also increasingly attractive because of the heavy cost of building fabs. Each successive generation of semiconductor process nodes cost more than the prior ones. And increasingly became too expensive to be reliably covered by any one company’s products. The catastrophic failure that would result from a new chip not having the demand needed to cover its fab’s costs also imposes a conservatism on companies that design and manufacture their own chips. Both on their designs and their building of fab capacity.
Keeping up progress and yield improvement meant continually investing in state-of-the-art fabs, and amortizing that cost took more volume than any one company could supply.
And while the fabless market barely existed, Morris Chang felt confident there was latent demand because of how many chip designers wanted to create new companies and chips but never did due to not having a fab.
Turning companies into a market
TSMC changed the shape of companies in the semiconductor industry.
TSMC being the foundry for semiconductor companies turned the high upfront capex costs of running a fab into a variable cost for them. Allowing these companies to start cheaply and only scale their fabrication costs as their business scaled.
The costs of operating fabs had dictated the industry shape prior to TSMC, a small set of IDMs that both designed and manufactured their own chips. TSMC made it possible for a new shape, companies that solely focused on chip design and left manufacturing to foundries like TSMC.
It took time for TSMC’s approach to bear fruit, because its ideal customers did not yet exist. Instead its early business was the dregs of existing IDM work, the chips that others did not want to produce themselves. But TSMC’s existence increasingly enabled a new breed of chip companies to emerge who were fabless from the foundation of their chip efforts. Companies like Nvidia, Qualcomm, Broadcom, and more.
For these companies, using TSMC was not a price optimization but existential to the structure of their company and strategy. They never thought to build their own fabs, instead focusing their risk and innovation at a higher level of chip design.
TSMC’s approach allows both their customers and TSMC to optimize independently on each side of design and manufacturing. This ecosystem approach continually grows stronger relative to the traditional IDM structure.
Traditional IDMs increasingly found themselves in a precarious bind. IDMs have to be world class at two distinct and increasingly expensive activities, deciding how to split their resources across them. Even worse, they need to synchronize the investment cycles of both, correctly predicting which chips will be successful and exactly how much fab capacity they will need.
Meanwhile they are cornered from both sides by the TSMC ecosystem. By separating design and manufacturing, TSMC can optimize manufacturing across the aggregate demand of the market, while thousands of lightweight fabless semiconductor companies independently explore chip designs and attack IDMs in every product category including ones too speculative for IDMs to invest in. TSMC can build its fab capacity without worrying about whether any individual chip succeeds or fails. TSMC is okay with many of those startups failing while able to support the scaling of the few that find traction.
TSMC turned chip design into a true market, unleashing a horde of fabless startups to do what the market does best, identifying the most promising products via competition and survival of the fittest.
Each IDM was effectively internally deciding which chips deserved fab capacity. TSMC’s ecosystem lets the market make and judge those bets in parallel. It’s no surprise that markets are far more efficient and innovative than the exec suite of most conglomerates. Where markets can function they tend to dominate, and TSMC creates the conditions for chip design to be a working market where chip design companies compete on their chip designs not their capital.
Manufacturing your customers
TSMC was built on the belief that there was latent demand to create fabless chip companies. But it was not a fluke that they turned out to both be right and at the perfect time.
TSMC pulled forward and created the future it sought.
TSMC doesn’t just manufacture chips, it manufactures new customers for itself.
It removed the barriers to entry to starting a semiconductor company fundamentally, lowering the costs by orders of magnitude and what expertise was required. This opened the market to a much larger set of companies and potential founders that couldn’t have considered it before.
As more fabless companies start and scale on TSMC they strengthen the TSMC ecosystem. They drive revenue and volume to TSMC that it reinvests in better manufacturing nodes, yields, and pricing. They push TSMC into new advanced techniques that can then be used by others and help improve its PDKs and processes to make it easier for new startups to work with TSMC. Through the proof of each successful company and the improvement in reliability and benefits of scale they normalize fabless startups to investors such that now it would be concerning to investors to hear a company plans to build their own fabs. And perhaps most of all, this large and expanding number of fabless companies draws an expanded ecosystem around it, spurring many vendors that make it even easier for fabless companies to start like better EDA software and reusable IP blocks.
And all of these lead to fabless companies becoming increasingly viable. This loop is self-reinforcing. TSMC doesn’t just get better and get more customers with more scale. It expands the potential customer base itself.
Most scale advantages improve the product as the customer base grows. Network effects make the product better. Economies of scale make the product cheaper. But what TSMC does changes the nature of the product and the customers themselves. TSMC lowers the minimum viable shape of a chip company. It converts the core fixed costs of a chip company into a variable one, allowing firms that otherwise couldn’t exist to compete. And it reinvests what it captures from that ecosystem into getting better, so these companies can take more of the market. It expands the potential customer base and reformats the entire ecosystem to build around TSMC.
Abstract warfare
TSMC’s strategy is particularly difficult for incumbents because it is not attacking their product. It is attacking their raison d’être and Coasean boundaries. TSMC moved the line between what a semiconductor company should do itself and what it should buy from the market.
IDMs could move production to TSMC. But owning a fab is precisely what makes them powerful. Dropping that capability voluntarily puts them in the morass of fabless companies, competing against companies that have been built around that model from inception. They would be letting go of their primary business with no guarantee of anything to show for it.
Or they can remain vertically integrated and continue to struggle against TSMC’s superior economies of scale and the power of the fabless ecosystem’s market discovery.
The struggle can be seen in AMD and Intel, both longstanding IDMs that have had to maneuver over the last decades to respond to the effect of TSMC on the market.
AMD’s founder Jerry Sanders famously said “Real men have fabs.” But by 2009 AMD relented, spinning off its foundry as a separate company, GlobalFoundries. Though this was in reaction to being unable to handle the growing costs of running its own fabs, it still resisted becoming fabless. The intent was to have the benefits of integrated design and fabrication without the financial burden of running the fabs on its balance sheet. It created long term contracts binding its production to GlobalFoundries while letting it raise outside capital from Abu Dhabi’s Mubadala and sell to others. Instead of being the best of both worlds, this became the worst. AMD could not control GlobalFoundries but their fates were tethered together. GlobalFoundries’ struggles hurt AMD and the contracts penalized AMD for using other foundries. Eventually GlobalFoundries gave up on competing with TSMC on the leading nodes allowing AMD to truly become fabless and adopt TSMC for its products.
Unlike AMD, Intel resisted giving up manufacturing as its core. This created significant problems for it when its foundry struggled to transition from 14nm to 10nm and again to 7nm. Because its design and manufacturing were tightly integrated, Intel’s foundry issues became product issues.
Intel’s supposed advantage in manufacturing became a large liability, keeping it on inferior nodes while competitors used TSMC’s. This then compounded into worse economics for its fabs which were primarily supported by its products causing them to fall further behind TSMC’s progress.
Intel in recent years has been adjusting its approach, under first the leadership of Pat Gelsinger and now Lip-Bu Tan1. There’s much one should write about that, but that’s for another piece. The common thread across both tenures has been keeping manufacturing and design under one roof while forcing each to operate as though it weren’t.
Intel has split the products and foundry businesses into separate entities within Intel. Each is responsible for its own P&L, the foundry charges the products group market prices, and both are free to work with outside foundries and customers. Intel is hoping market discipline forces both businesses into being best in class while maintaining the advantages of co-design.
AMD tried to keep the benefits of integration without the abstraction’s discipline and got neither. It separated the businesses de jure but not de facto and suffered all the problems of an IDM with none of the control. Once the separation became real, both businesses did better. Intel’s bet is that it can keep both halves only by making them each live under the rules TSMC wrote.
TSMC is so effective because it shifted the fundamental abstractions of the semiconductor industry. By attacking incumbents’ business models rather than their products TSMC turned their strengths into baggage. And one way or another every company has had to adjust to this new market structure.
Markets are made
The best companies must treat their market as endogenous. They don’t just grow their share of the market. They continually grow the market itself and create new customers for themselves. Otherwise they’d eventually saturate and their growth would slow.
TSMC grows the fabless market and the fabless market grows TSMC. This is why the best companies are consistently larger than the market expects them to be. Demand elasticity is the only thing that can truly surprise us.
TSMC made its customers, but it did not make their customers. It made new kinds of chips easier to build, but it still needed the companies building them to find new demand. For many years that was mobile, with Apple and others pulling enormous volume through TSMC. More recently it is AI and the datacenter buildout, with Nvidia building on TSMC and the labs building on Nvidia. Each layer makes the one above it possible, and depends on that layer finding demand of its own.
Tech history is about making markets. The defining companies almost all created new ones. Companies can luck into a growing market, but the best way to find that luck is to be the reason the market grows.
KK final note: Silicon Valley spent decades without much silicon. Now semiconductors are the center of attention again after having been in their own bubble universe for a long time. The industry works in many ways that are unintuitive to people who grew up only in software. But it’s also changing in ways unfamiliar to those who’ve spent decades in it. The scale and shape of demand are different, and so are the customers themselves. More on that soon.
Footnote:
- I cannot express how much of a fan I am of Lip-Bu Tan. I once tried to commission TikTok creators to create meme edits of LBT like they do for their favorite TV shows and basketball teams ↩︎