台积电研究专辑——第二篇

I've been posting a bit too much lately. After this post, I'll be going back to lurking.

Yesterday, I liquidated all my options except for TSMC and Tesla. I also reduced my positions in Microsoft and Amazon because the increase in intrinsic value exceeded my preset position limits. My current holdings are Microsoft, TSMC, Amazon, Google, and Tesla.

Let's get back to the topic of TSMC. I believe that if you don't understand the semiconductor industry chain, you will never truly understand TSMC.

Many investors discuss NVIDIA, TSMC, and ASML every day, but very few people can clearly explain the entire process of a chip from design to shipment. This process is precisely what determines who holds the most power in the industry chain and who captures the highest profits.

How is an AI chip born?

Before researching TSMC, I always had a question: Why does NVIDIA, with the world's best GPU design capabilities, not manufacture its own GPUs? Why does Apple invest tens of billions of dollars annually in chip R&D yet still rely on TSMC?

After further research, I realized that the modern semiconductor industry is no longer an industry where a single company can complete all tasks independently. Instead, it is a highly specialized global collaboration system. Each company's advantage is built upon long-term accumulated expertise, and TSMC is one of the most critical nodes in this system.

The best way to understand this is to follow an AI chip from its initial design all the way to its final installation in a server.

Starting with an idea

Suppose NVIDIA prepares to develop the next generation of AI GPUs.

The real work doesn't start with engineers drawing circuit diagrams; instead, the product team first defines the goals: What level of computing performance, power consumption control, and market positioning do they want to achieve?

Then, chip architects begin designing the overall architecture, deciding how many compute units the GPU needs, how cache should be laid out, and how data will be transmitted. These decisions determine what tasks a chip can perform in the future, but at this stage, it remains just a design plan.

In this phase, NVIDIA creates intellectual property, not physical products.

EDA Software: An indispensable tool for modern chip design

With the architecture in place, engineers don't manually draw every circuit. Instead, they use professional EDA software to complete the design, verification, and simulation.

EDA (Electronic Design Automation) can be understood as the "CAD software" of the chip industry, though it is far more complex than architectural design. Without EDA software, designing modern advanced chips would be nearly impossible.

Currently, this field has long been dominated by a few companies, such as Cadence Design Systems and Synopsys. This also shows that what truly matters in the semiconductor industry is not just manufacturing capability; every link has extremely high technical barriers.

IP Licensing: Not everything needs to be redesigned from scratch

Many investors believe that every module on a chip is developed by the company itself. This is not true.

For example, CPU cores, interface controllers, and high-speed communication modules are often purchased as mature IP licenses. A typical example is Arm Holdings. Companies like Apple, Qualcomm, and MediaTek develop their products based on the Arm architecture rather than designing everything from zero. This specialization significantly improves the innovation efficiency of the entire industry.

Why can't NVIDIA build its own foundry?

Once the design is complete, a problem arises: How do we turn the design into a physical chip? Theoretically, NVIDIA could certainly build its own factory. However, in reality, this is almost unfeasible. Building an advanced wafer fab requires massive capital investment and, more importantly, long-term accumulation of manufacturing experience, process control capabilities, and supply chain management skills. Even with funds, it doesn't guarantee stable production of advanced chips. Therefore, NVIDIA chooses to outsource manufacturing to TSMC, focusing its resources on chip architecture and software ecosystems. This division of labor allows both parties to focus on their respective strengths.

Why are lithography machines so important?

Once manufacturing begins, the equipment most familiar to people is the lithography machine. Among them, ASML has become the primary supplier of advanced EUV lithography equipment. However, I used to mistakenly think, "Having ASML means having advanced manufacturing." Later, I realized this understanding was inaccurate. Lithography machines are indeed important, but they are just one key piece of equipment in the entire manufacturing process. What truly determines the final product quality is:

the entire process flow, thousands of manufacturing steps, extremely high consistency control, yield management, and supply chain coordination. If there is only equipment without mature processes, stable mass production is still impossible. Therefore, I gradually realized that ASML's importance lies in providing tools, while TSMC's importance lies in converting these tools into replicable, mass-producible manufacturing capabilities.

What does TSMC truly deliver?

Many people think TSMC delivers wafers. I now believe it truly delivers "certainty." For NVIDIA, whether a GPU successfully hits the market depends not just on excellent design. More importantly, it depends on: Can it be mass-produced on time? Can it maintain high yields? Can it meet the delivery demands of millions of chips? These capabilities cannot be simply explained by a single piece of equipment or technology; they are the result of long-term process accumulation.

Therefore, what customers buy is not just manufacturing services, but the certainty of mass-producing advanced chips on a large scale.

After manufacturing is complete, the chip's journey isn't over

After wafer manufacturing is finished, the chips cannot be directly installed into servers. They must go through several more stages, including dicing, packaging, and testing. Especially in the AI era, the importance of advanced packaging is growing. In the past, packaging was like the final step; now, it already affects chip performance, bandwidth, and power consumption.

Why can't the entire industry chain do without TSMC?

Reviewing the entire process again: Chip design companies handle innovation, EDA companies provide design tools, IP companies provide foundational architectures, equipment companies provide manufacturing equipment, and material companies supply silicon wafers, photoresists, and chemicals. TSMC is the one that truly integrates all these links to achieve advanced process mass production. This is why I increasingly tend to view TSMC as a "platform enterprise" within the industry chain. Being a platform doesn't mean owning all technologies; it means integrating all key capabilities to create value for the entire industry.

My thoughts

My biggest feeling about the semiconductor industry is: Truly great industries rarely involve a single company doing everything. Instead, they form a highly specialized collaboration system. Apple doesn't need to manufacture chips, NVIDIA doesn't need to produce GPUs, ASML doesn't need to design AI processors, and TSMC doesn't need to develop its own CPUs.

Each company focuses on what it does best, which allows the entire industry chain to move forward continuously.

This is why I am increasingly bullish on TSMC.

It is not the company with the highest profits in the industry chain, nor the one with the highest brand recognition, but it stands at a critical position in the entire system. As AI, high-performance computing, and advanced processes continue to develop, the importance of this position is constantly increasing.

Conclusion of this article

If the first

"What kind of company is TSMC?"

then this article answers:

"Why does the entire semiconductor industry chain ultimately converge on TSMC?"

Understanding the entire process of an AI chip from design to manufacturing helps us understand why TSMC is not just an ordinary foundry, but an irreplaceable key link in the global advanced semiconductor industry.

Why are customers willing to hand over profits to TSMC?

If this question cannot be answered clearly, it is difficult to understand why TSMC can maintain profitability far above the average level of the manufacturing industry for a long time.

TSMC's true business model: Why can it earn money others cannot?

When I first read TSMC's financial reports, I had a question. From an asset structure perspective, it is a typical heavy-asset enterprise. Building wafer fabs requires huge capital investments, equipment updates quickly, and depreciation pressure is high. By traditional manufacturing logic, such enterprises usually find it hard to maintain high profit margins over the long term. However, TSMC demonstrates completely different operational characteristics. Even with continuous large-scale capital expenditures, the company maintains high gross margins and return on capital. This shows that its business model cannot be simply classified as traditional manufacturing. Later, I gradually realized that the issue is not "what TSMC produces," but why customers must choose it.

Manufacturing capability itself is a product

Many people think TSMC provides wafer manufacturing services. I now prefer to think it sells a capability. For Apple, it needs to release new iPhones on schedule; for NVIDIA, it needs to stably deliver data center GPUs; the same applies to AMD, Qualcomm, and other companies. What these customers truly buy is not a slice of silicon wafer, but three results:

First, advanced processes can achieve expected performance;

Second, mass production can proceed according to plan;

Third, product yields are sufficiently high.

If these three goals cannot be achieved, even the best chip designs cannot translate into commercial success. Therefore, manufacturing capability itself is TSMC's most important product.

Why is yield more important than price?

After studying semiconductors, I discovered a metric that was previously easy to overlook—yield (Yield). Simply put, yield represents the proportion of produced wafers that meet design standards and can be sold. For advanced chips, changes in yield directly impact customer costs.

Let's take a simplified example.

Assume the manufacturing cost of an advanced GPU is very high. If the yield increases from 70% to 90%, it means the manufacturing cost allocated to each sellable chip decreases significantly. For customers, this value far exceeds minor fluctuations in manufacturing prices. Therefore, the question customers truly care about is often not: Did TSMC charge a few hundred dollars more per wafer? But rather: Can it stably produce the chips I need?

This is why, in the field of advanced processes, the importance of price competition decreases, and stable mass production capability becomes more critical.

Why are customers unwilling to frequently switch foundries?

Price wars occur in many industries, but advanced wafer manufacturing rarely sees this. The reason is that switching foundries is not as simple as changing suppliers. After a chip completes its design, it needs optimization for specific processes. If shifting from one manufacturer to another, it requires re-verifying the design and may affect performance, power consumption, and time-to-market. For a flagship product invested with billions of dollars in R&D, delaying launch by a few months could cause losses far exceeding the saved manufacturing costs. Therefore, the focus of competition in advanced processes is not who offers the lower quote, but who can reduce customer uncertainty. This is fundamentally different from ordinary manufacturing.

Why have capital expenditures increased so much, yet profits haven't been swallowed?

This was the question I wanted to figure out most when researching TSMC. Logically, increasing capital expenditures lead to higher depreciation, thereby compressing profits. However, TSMC's performance over the past years shows things aren't that simple. I believe the key lies in the rising value of advanced processes. As chips become more complex, customers' reliance on advanced manufacturing grows stronger. If a technology helps customers achieve better performance, lower power consumption, or faster time-to-market, customers are willing to pay a higher price. In other words, the value added by advanced processes partially offsets the pressure brought by increased capital investment.

Of course, this doesn't mean capital expenditures carry no risk. If demand for advanced processes slows down significantly in the future while fixed assets continue to increase, profitability could still be affected. Therefore, capital expenditure remains a crucial indicator that must be tracked long-term when analyzing TSMC.

Where does TSMC's true pricing power come from?

Many enterprises possess brand equity, thus holding pricing power. TSMC lacks consumer brands; its pricing power comes from elsewhere. If a customer has only one supplier capable of meeting their needs, price discussions won't revolve solely around costs but will focus more on value. This doesn't mean TSMC can raise prices arbitrarily; it means that when advanced manufacturing capacity is limited and customer products heavily depend on this capability, the factors determining price change. From a business model perspective, this is more solid than relying on brand premiums because it is built upon technology and manufacturing capability.

Why do I think TSMC is more like a platform enterprise?

Writing this, I am increasingly reluctant to view TSMC as an ordinary manufacturing enterprise. Manufacturing enterprises typically rely on orders, while platform enterprises rely on ecosystems. Although TSMC produces wafers, it connects the entire semiconductor ecosystem. Design companies hope it stays ahead, equipment companies need to develop new devices around it, material suppliers need to cooperate with it to optimize processes, and packaging enterprises need to adapt to its new products. Customers' new architectures are often co-developed with it in advance. In other words, TSMC is not merely participating in the industry chain; it is influencing the direction of the industry chain to some extent.

Researching this far, I began to re-understand the word "manufacturing." In the past, I always thought manufacturing competition was about cost. Later, I found that in the field of advanced semiconductors, what truly competes is technology, experience, craftsmanship, reliability, scale, and long-term trust.

These capabilities collectively determine whether a company can sustain value creation. Therefore, I increasingly believe that TSMC's most important asset is not its factories, but the manufacturing system accumulated over decades. Factories can be built, equipment can be bought, and talent can be hired. But integrating these into a continuously leading, stably mass-producing, and constantly iterating manufacturing capability requires a very long time.

After studying TSMC's business model, my biggest cognitive shift is: It doesn't make money by "producing more wafers," but by "continuously providing the world's most advanced and reliable manufacturing capabilities." For customers, they are buying not just manufacturing services, but the certainty that products will hit the market on time, meet performance expectations, and keep costs controllable. This certainty is the core value of TSMC's business model and the main reason I believe it can maintain its competitive advantage for a long time.

Final thoughts

This chapter actually planted a more important question:

If TSMC truly sells "manufacturing capability," why has this capability remained so difficult to replicate for decades?

This is exactly what the next article will discuss.

"Part Four - TSMC's True Moat: Why Have Samsung and Intel Always Struggled to Catch Up?"$Amazon(AMZN.US)