Table of Contents

TL;DR

TSMC controls over 90% of advanced process capacity below 7nm globally. AI accelerator demand pushed its 2025 market cap past $2 trillion, with stock up over 103% in a year. TSMC represents over 40% of Taiwan’s Weighted Index, making the index essentially a proxy for TSMC’s fortunes. Technically, virtually every AI chip on 3nm and 2nm runs through TSMC’s lines — Nvidia H100, H200, B200, Apple M4, AMD MI300, all of them.

What Happened

In 2025, TSMC (NYSE: TSM) ran from approximately $196 at the start of the year to over $390 by year-end, gaining over 100% for the year. Its market cap exceeded $2 trillion, making it one of the top five companies globally by market cap and the first Taiwanese company ever to reach that scale.

The direct cause was the explosion in AI data center demand. Nvidia’s GPUs (B200, H200), Google’s TPUs, Amazon’s Trainium, Microsoft’s Maia — all the AI chips powering ChatGPT, Gemini, Claude, and other large language models depend on TSMC manufacturing. TSMC’s April 2025 monthly revenue grew 17.5% year-over-year, reaching $13.3 billion (approximately NT$423.6 billion).

In full-year 2025, high-performance computing (HPC, primarily AI chips) contributed 58% of TSMC’s revenue, with smartphones contributing 29% — AI has replaced mobile phones as TSMC’s most important growth driver.

Why This Matters

TSMC’s Technical Moat

TSMC’s ability to maintain such high market share comes down to process technology leads that are hard to close:

Samsung’s difficulties: Samsung’s GAA (Gate-All-Around) 3nm process (SF3) is more aggressive than TSMC’s, but yield issues have kept major customers cautious. Qualcomm’s Snapdragon X Elite ultimately chose TSMC, partly due to concerns about Samsung’s yield.

Intel’s transformation struggles: Intel Foundry is ambitious, planning fabs in the US and Europe to attract customers, but the 18A process (approximately equivalent to 2nm) production schedule has slipped repeatedly, with limited early customer orders. Intel’s own processors have also faced performance and power consumption issues.

TSMC’s differentiation:

  • CoWoS (Chip on Wafer on Substrate) advanced packaging: Nvidia’s H100 and H200 integrate HBM high-bandwidth memory via CoWoS. TSMC’s packaging capacity is the most critical link in the AI chip supply bottleneck
  • SoIC (System on Integrated Chip) 3D stacking technology in progress
  • N2 (2nm) process entering volume production end of 2025, AI chip customers already queued

The Strategic Legacy of Morris Chang and C.C. Wei

Morris Chang’s core strategy founding TSMC — “pure-play foundry” — was completely vindicated in the AI era. When he proposed the model in 1987, the semiconductor industry mainstream was IDM (Integrated Device Manufacturer, design and manufacturing in-house). TSMC chose to only manufacture, not design, not compete with customers. This model enabled fabless semiconductor companies like Nvidia, AMD, and Apple to focus entirely on design, outsourcing manufacturing to TSMC.

C.C. Wei has continued reinforcing this moat while aggressively pursuing geographic expansion: Arizona fab (2nm, targeting 2026 production), Japan Kumamoto fab (28nm, opened 2024), Germany Dresden fab (12nm, targeting 2027). Geographic diversification is proactive defense against geopolitical risk.

The Engineering View

Why AI Chips Especially Need Advanced Processes

AI accelerators (Nvidia GPUs, Google TPUs) are fundamentally large arrays of matrix multiplication units. On a 3nm process, you can fit more compute units into the same area, or achieve the same performance at dramatically lower power draw. In the AI era, data center electricity cost has become a primary operating cost — so performance-per-watt directly affects ROI.

Another key factor is memory bandwidth. LLM inference bottlenecks are often not compute, but bandwidth for reading weights from HBM. TSMC’s CoWoS packaging lets GPU and HBM be tightly integrated — the most practical solution today for breaking through memory bandwidth constraints.

TSMC’s Role in the AI Supply Chain

AI applications (ChatGPT, Claude, etc.)
    ↓
AI model training/inference (Nvidia GPU / Google TPU)
    ↓
Chip design (Nvidia, AMD, Google, Microsoft, Amazon)
    ↓
Wafer foundry (TSMC, 90%+ leading-edge market share)
    ↓
Advanced packaging (TSMC CoWoS / SoIC)
    ↓
Data center infrastructure

TSMC’s position in this chain makes it indispensable to AI infrastructure. Any event affecting TSMC’s capacity — earthquake, power outage, geopolitical conflict — would directly disrupt the pace of global AI development.

What to Watch Next

N2 process production progress: TSMC’s N2 yield and production ramp determines when next-generation AI chips (Nvidia’s Blackwell successor, Apple M5) reach market.

CoWoS capacity: CoWoS is the actual current bottleneck. TSMC continues expanding CoWoS capacity, but demand growth is faster — this is the most strained point in the entire AI supply chain.

Geopolitical risk: TSMC’s $2T+ market cap makes it simultaneously Taiwan’s largest asset and largest strategic target. Arizona fab progress and government subsidies (US CHIPS Act) are critical variables in mitigating this risk.

Taiwan stock concentration: TSMC over 40% of the weighted index means Taiwan’s market is essentially a bet on AI chip demand. This is concentration risk for long-term investors, but also reflects Taiwan’s strategic position in the global semiconductor supply chain.

References

🇺🇸 English

Here's something that should make every engineer pause for a second: one company manufactures over 90% of the world's most advanced chips. Not 51%. Not 70%. Over 90%. And in 2025, that company crossed a two-trillion-dollar market cap — the first Taiwanese company ever to reach that scale, and one of only five companies on the entire planet at that valuation. That company is TSMC, and the reason it got there is you — or at least, the AI you've been using.

Every Nvidia GPU training the models behind ChatGPT, Claude, Gemini. Every Google TPU. Amazon's Trainium, Microsoft's Maia — all of them roll off TSMC's production lines. When TSMC's monthly revenue hit 13.3 billion dollars in April 2025, up 17.5% year over year, it wasn't smartphones driving that. AI chips — what the industry calls high-performance computing — made up 58% of TSMC's revenue. Mobile phones, which used to be the crown jewel, dropped to 29%. The AI era didn't just boost TSMC. It restructured the entire company's center of gravity.

So how did TSMC get here? And why can't Samsung or Intel close the gap?

Let's start with the technical reality. At 3nm and 2nm process nodes — the cutting edge — TSMC's yields are simply better. Samsung has an aggressive Gate-All-Around architecture for their 3nm process, which on paper looks more advanced. But yield problems kept major customers away. Qualcomm's Snapdragon X Elite, which was originally a potential Samsung win, ended up going to TSMC partly because of those yield concerns. When you're building a chip that costs hundreds of millions in design investment, a yield difference of even a few percentage points is the difference between a viable business and a disaster.

Intel is a different story — more of a tragedy of ambition. Intel Foundry is genuinely trying, with planned fabs in the US and Europe, and their 18A process is roughly equivalent to 2nm. But production schedules have slipped repeatedly, early customer orders are thin, and Intel's own chips have had performance and power issues that undermine confidence. You can't sell yourself as a foundry if your own products aren't inspiring trust.

TSMC's edge isn't just process nodes though. The real moat in the AI era is CoWoS — Chip on Wafer on Substrate. This is advanced packaging technology that lets you integrate a GPU die with HBM high-bandwidth memory in extremely close physical proximity. Why does this matter for AI? Because the bottleneck in running large language models isn't usually compute — it's bandwidth. Reading billions of model weights from memory fast enough to keep the compute units fed. CoWoS is the most practical solution available today to break through that bandwidth wall. Nvidia's H100 and H200 both depend on it. And TSMC's CoWoS capacity is, right now, the single most constrained point in the entire global AI supply chain. Demand is growing faster than capacity is being built.

Now here's a bit of history worth knowing. Morris Chang founded TSMC in 1987 on a model that the industry thought was weird: pure-play foundry. They would only manufacture chips. No chip design. No competing with customers. At the time, the dominant model was companies like Intel doing everything in-house — design, manufacturing, the whole stack. Chang's bet was that fabless chip companies would rather outsource manufacturing and focus entirely on design. He was right, and it took about three decades for the full vindication to arrive. Nvidia, AMD, Apple — all of them are fabless. All of them depend on TSMC. The AI boom is essentially the compound interest on a strategic decision made in 1987.

TSMC's current CEO, C.C. Wei, has been expanding geographically — an Arizona fab targeting 2nm production in 2026, a Japan fab already open running at 28nm, a Germany fab on track for 2027. This isn't just growth strategy. It's geopolitical hedging. TSMC sitting almost entirely in Taiwan creates concentration risk that every government with an interest in AI infrastructure is very aware of.

Which brings us to the other concentration risk nobody talks about enough: Taiwan's own stock market. TSMC represents over 40% of Taiwan's Weighted Index. Forty percent. That means Taiwan's benchmark index is essentially a bet on AI chip demand. For investors, that's a real risk to understand. But it also tells you something profound about Taiwan's position in the global technology supply chain — a small island nation that has made itself structurally indispensable to the future of AI.

Three things to take away from all of this.

First, TSMC's moat is real and it's deep. It's not just process technology — it's packaging, it's customer trust built over decades, and it's yield reliability that competitors haven't matched. Closing that gap takes years, not quarters.

Second, CoWoS is the actual bottleneck right now. If you're watching the AI chip space, don't just track process node announcements. Watch CoWoS capacity expansion. That's where supply is tightest and where delays will be felt first.

And third — any single event disrupting TSMC, whether technical, natural, or geopolitical, directly disrupts the pace of global AI development. That's not hyperbole. That's the supply chain math. The world built AI on a single point of manufacturing excellence, and that's both a testament to how good TSMC is, and a risk that everyone from chip designers to governments is now very much awake to.

🇹🇼 中文

台積電的市值在 2025 年突破了 2 兆美元。這個數字有多大？它讓台積電擠進全球市值前五大企業，也是台灣有史以來第一家達到這個規模的本土公司。而驅動這一切的，只有兩個字：AI 晶片。

先來說一個核心事實。全球所有 7 奈米以下的先進製程，台積電掌控了超過 90% 的產能。換句話說，你現在用的 ChatGPT、Claude、Gemini——背後跑這些模型的晶片，幾乎全部都是台積電做的。Nvidia 的 H100、H200、B200，Apple 的 M4，AMD 的 MI300，清單可以一直列下去。

2025 年台積電的股價從年初大約 196 美元，一路漲到年底超過 390 美元，全年漲幅超過 100%。這波漲勢的直接原因很清楚：AI 資料中心的需求在爆炸。Nvidia 的 GPU、Google 的 TPU、Amazon 的 Trainium、Microsoft 的 Maia——這些巨頭全都要搶台積電的產線。台積電 2025 年的營收結構也因此翻轉了，AI 晶片所在的高效能運算類別貢獻了 58% 的營收，智慧型手機只剩下 29%。AI 已經正式取代手機，成為台積電最重要的成長引擎。

那三星和英特爾呢？為什麼市場份額沒辦法搶過來？

三星其實技術路線走得更激進，他們的 3 奈米製程在結構設計上比台積電更前衛，但問題出在良率。良率就是生產出來的晶片有多少比例是好的、可以用的。良率不穩，客戶就不敢下單。高通的 Snapdragon X Elite 最後選擇台積電，部分原因就是對三星良率的疑慮。英特爾的情況更複雜，他們想重建晶圓代工業務，在美國和歐洲都有建廠計劃，但關鍵製程的量產時程一再往後推，加上自家處理器也出問題，信心很難重建。

台積電真正的護城河，除了製程技術本身，還有一個關鍵字：CoWoS。這是一種先進封裝技術，讓 GPU 和高頻寬記憶體 HBM 可以緊密整合在一起。AI 模型在推論時的瓶頸，很多時候不是算力不夠，而是從記憶體讀取資料的速度跟不上。CoWoS 就是解決這個瓶頸目前最實際的方案，而且這個技術台積電幾乎是獨家掌握。目前整個 AI 供應鏈最緊張的點，不是晶圓產能，而是 CoWoS 封裝產能，需求成長的速度比台積電擴充的速度還快。

從供應鏈的角度看，台積電的位置非常特殊——它是整條 AI 基礎設施鏈上唯一無可取代的環節。任何影響台積電產能的事件，不管是地震、停電，還是地緣政治衝突，都會直接衝擊全球 AI 發展的節奏。這也是為什麼台積電同時是台灣最大的資產，也是最大的戰略目標。

台積電的全球佈局是對這個風險的主動防禦：亞利桑那廠預計 2026 年量產 2 奈米，日本熊本廠已經開幕，德國的廠也在規劃中。地理分散不只是商業策略，也是地緣政治層面的風險管理。

最後，有一個投資角度的觀察值得提一下。台積電在台灣加權指數的佔比超過 40%，這代表你買台灣股市，本質上有將近一半是在押注 AI 晶片需求。這對長期投資人來說是集中度風險，但換個角度看，它也精準地反映了台灣在全球半導體供應鏈裡的戰略地位。

總結三個核心要點：第一，台積電控制全球 90% 以上先進製程，AI 晶片沒有替代路線，短期內這個格局不會改變。第二，CoWoS 封裝技術是目前整個 AI 供應鏈最緊張的瓶頸，比晶圓產能更關鍵。第三，地緣政治風險是台積電最難量化的變數，全球佈局是它主動降低單點脆弱性的策略，但這需要時間，而風險是現在進行式。

Tags

#tsmc#semiconductors#ai-chips#foundry#semiconductor-industry

Related Articles