Table of Contents

TL;DR

TSMC (TSM) ran from approximately $196 at the start of 2025 to over $390 by year-end, with its market cap breaking $2 trillion — over 100% gain for the year. It wasn’t just Taiwanese buyers: Korean retail investors, US institutional investors, and Japan’s government pension fund all increased their positions. The fundamental reason isn’t complicated: nearly all AI chips in the world run through TSMC’s manufacturing lines, and that situation is hard to change in any foreseeable timeframe.

What Happened

2025 was the year TSMC as an investment thesis was “discovered globally.”

Price performance: TSM ADR gained over 103% for the year; Taiwan-listed TSMC (2330) set multiple all-time highs. By end of 2025, TSMC’s market cap was approximately $2.06 trillion, surpassing Alphabet (Google’s parent), making it the fifth-largest company in the world by market cap.

Revenue numbers: Full-year 2025 combined revenue reached NT$3.8 trillion (approximately $116 billion USD), up over 31% from 2024. High-performance computing (HPC, primarily AI chips) contributed 58% of revenue; monthly revenue set all-time records multiple times through 2025.

Investor activity by region:

  • Korea: Korean retail investors bought large amounts of TSM ADR through HTS (Home Trading Systems), with the KOSPI itself up over 80% in 2025, and semiconductor stocks becoming the hottest topic driven by Samsung and SK Hynix momentum
  • US: Mainstream institutions (BlackRock, Fidelity, Vanguard) continued increasing TSM positions; TSM is one of the rare non-US companies in the S&P 500
  • Japan: GPIF (Government Pension Investment Fund, the world’s largest retirement fund) increased TSMC holdings; the Japanese government’s chip subsidies attracting TSMC to build a Kumamoto fab also lifted Japanese semiconductor stocks

Why This Matters

Structural Reasons for TSMC’s Irreplaceability

TSMC’s 2025 run wasn’t a bubble — it had clear fundamental support:

Supply side: 90%+ of leading-edge capacity Over 90% of chips below 7nm are manufactured by TSMC. Samsung is catching up, but yield issues with their 3nm process have kept major customers — Qualcomm, Nvidia, AMD — from fully switching. Intel Foundry was still in its ramp-up phase in 2025.

Demand side: all AI accelerators bet on TSMC Nvidia H200, B200, GB200; Google TPU v5; Amazon Trainium 2; Apple M4 — these products defined the 2025 AI infrastructure landscape, and all are manufactured on TSMC’s 3nm or 5nm processes. Without these chips, there’s no compute for ChatGPT.

CoWoS packaging: the hardest part to replicate Another bottleneck in AI chips is integrating high-bandwidth memory (HBM). TSMC’s CoWoS (Chip on Wafer on Substrate) technology enables tight packaging of GPU and HBM — it’s the foundation that makes the Nvidia H100/H200 possible. CoWoS capacity was the most constrained supply element in 2025, giving TSMC additional pricing power.

North America Revenue at 75%

75% of TSMC’s 2025 revenue came from North American customers — primarily Apple, Nvidia, AMD, Qualcomm, and Google. That number says TSMC is essentially “a Taiwanese company manufacturing for American tech companies,” and its success is tightly coupled to American tech’s success.

The Engineering Perspective

AI application demand (ChatGPT, Gemini user explosion)
    ↓
Compute demand (cloud providers buying massive amounts of GPUs)
    ↓
GPU orders (Nvidia, AMD order from TSMC)
    ↓
TSMC 3nm/CoWoS capacity
    ↓ (the only path through)
TSMC market cap, stock price

This chain is short and has almost no detours. That’s why a large portion of global “invest in AI” capital flows into TSMC — it’s the unavoidable bottleneck in the AI supply chain.

Taiwan stock concentration risk: TSMC represents over 40% of the Taiwan Weighted Index (TAIEX). This means the TAIEX’s moves are substantially TSMC’s moves. For funds investing in Taiwan, this is concentration risk; but it also reflects TSMC’s real role in Taiwan’s economy — the “guardian mountain” nickname isn’t an exaggeration.

What to Watch Next

N2 volume production progress: TSMC’s N2 process entered trial production at end of 2025, with full production in 2026. Customers include Apple (M5) and Nvidia’s next-gen GPU. N2 yield and production ramp directly affect the 2026 AI chip specification timeline.

CoWoS capacity expansion: TSMC continues investing in CoWoS capacity, targeting doubling by end of 2026. This schedule determines delivery pace for next-gen AI chips like the Nvidia GB300.

Geopolitics: TSMC’s $2T market cap makes it simultaneously Taiwan’s largest asset and largest strategic target. The Arizona fab (targeting 2nm production in 2026), Japan’s Kumamoto fab (28nm, already operating), and the Germany Dresden fab (12nm, targeting 2027) are geographic diversification moves — but Taiwan remains the irreplaceable core in the near term.

TSMC $3T market cap? Motley Fool analysts believe TSMC has a path to $3 trillion market cap before 2030, assuming sustained AI data center demand and N2/A16 process execution on schedule.

References

🇺🇸 English

TSMC just crossed two trillion dollars in market cap. Let that sink in. A Taiwanese semiconductor company is now the fifth-largest company on the planet — bigger than Alphabet, Google's parent. And it got there by more than doubling its stock price in a single year.

So what actually happened, and why should engineers care?

Here's the raw story: TSMC's American depositary shares went from around $196 at the start of 2025 to over $390 by year-end. That's a 103% gain. Revenue hit the equivalent of about $116 billion US dollars for the full year — up 31% from 2024. And it wasn't just Taiwanese investors riding this wave. Korean retail traders were piling in through their brokerage apps. US giants like BlackRock, Fidelity, and Vanguard kept increasing their positions. Even Japan's government pension fund — the largest retirement fund in the world — bought more TSMC. This was a global pile-in.

The question is: why?

The answer is almost frustratingly simple. Nearly every meaningful AI chip on the planet runs through TSMC's factories. Nvidia's H200, B200, GB200. Google's TPUs. Amazon's Trainium chips. Apple's M4. All of them fabbed on TSMC's 3-nanometer or 5-nanometer processes. If you want compute to run ChatGPT, Gemini, or any large-scale AI workload — there is essentially one road, and TSMC owns it.

On the supply side, TSMC controls over 90% of all leading-edge chip manufacturing — anything below 7 nanometers. Samsung is trying to catch up, but their 3-nanometer yields have been rough enough that Nvidia, Qualcomm, and AMD haven't made the jump. Intel's foundry ambitions were still in ramp-up mode through 2025. So there's no real alternative at scale.

Now here's the part that's less talked about but arguably more important: packaging. It's not enough to just make a fast GPU — you also need to connect it tightly to high-bandwidth memory. TSMC's CoWoS technology does exactly that. It's what lets you stack a GPU die and HBM memory modules into a single dense package, which is the physical foundation of chips like the H100 and H200. CoWoS capacity was the single most constrained resource in the AI supply chain in 2025. And TSMC owns that too. That scarcity gave them significant pricing leverage.

Think of the chain like this: AI applications explode in usage, cloud providers scramble to buy GPUs, Nvidia and AMD flood TSMC with orders, and TSMC is the only passage through. It's a short chain with almost no detours. That's why so much global "invest in AI" capital ultimately flows to TSMC — because it's the unavoidable bottleneck.

One number that frames this perfectly: 75% of TSMC's 2025 revenue came from North American customers. Apple, Nvidia, AMD, Qualcomm, Google. TSMC is essentially a Taiwanese company manufacturing the backbone of American tech. When American tech wins, TSMC wins.

There's also an interesting lens if you follow Taiwanese markets: TSMC makes up over 40% of the Taiwan stock index. So when you hear "Taiwan's market hit an all-time high," you're really mostly hearing "TSMC hit an all-time high." That's concentration risk, but it also honestly reflects TSMC's actual weight in Taiwan's economy.

Looking ahead, three things to watch. First, TSMC's N2 process — their next-generation node — is ramping into full production in 2026. Apple's M5 and Nvidia's next GPU generation both depend on it. Yield rates and production speed there will define the AI chip spec race for the next two years.

Second, CoWoS capacity. TSMC is targeting a doubling of CoWoS output by end of 2026. That timeline directly controls how fast next-gen AI accelerators like Nvidia's GB300 can ship.

Third, geopolitics. A $2 trillion company headquartered on a geopolitically sensitive island is a risk that everyone is watching. The Arizona fab is targeting 2-nanometer production in 2026. Kumamoto in Japan is already running. Dresden in Germany targets 2027. These are real diversification moves — but Taiwan remains the irreplaceable core for the foreseeable future.

Some analysts think TSMC has a credible path to $3 trillion before 2030, assuming AI data center demand holds and the N2 process ramp goes smoothly. Whether or not that target hits, the underlying story is the same.

Three takeaways. One: TSMC's rise isn't hype — it's a direct reflection of AI infrastructure demand, and the company sits at the only chokepoint that matters. Two: the CoWoS packaging business is just as strategically important as the chips themselves — it's where the physical constraints of AI hardware actually live. And three: geographic diversification is happening, but slowly — Taiwan remains the center of gravity for years to come, which means the geopolitical risk isn't going away.

This is what it looks like when a single company becomes load-bearing infrastructure for a global technology wave.

🇹🇼 中文

台積電在 2025 年底市值突破了 2 兆美元，股價從年初的 196 美元一路飆到 390 美元以上，全年漲幅超過 100%。這件事的重點不只是「台灣之光」這種情感層面，而是背後有非常清晰的結構性邏輯。

先說一個數字：全球 7 奈米以下的晶片，超過 90% 是台積電製造的。三星在追，英特爾也在爬坡，但高通、Nvidia、AMD 這些大客戶就是沒有全面轉換的意願——不是不想，是台積電的良率和製程成熟度，目前沒有人追得上。

那 2025 年的需求面發生了什麼？很簡單：AI 基礎設施大爆發。Nvidia 的 H200、B200、GB200，Google 的 TPU，Amazon 的 Trainium，蘋果的 M4——這些定義了 2025 年算力格局的產品，全部跑在台積電的 3 奈米或 5 奈米製程上。沒有這些晶片，就沒有 ChatGPT 跑得動的算力。邏輯鏈非常短：AI 用戶爆炸 → 雲端業者瘋狂買 GPU → GPU 廠商向台積電下單 → 台積電是唯一的出口。

還有一個更細節但很關鍵的點：CoWoS 封裝技術。AI 晶片除了本身的運算核心，還需要把高頻寬記憶體緊密整合進去，這個封裝工藝叫 CoWoS，台積電掌握得最好。2025 年 CoWoS 產能是整個 AI 供應鏈最緊張的環節之一，這讓台積電在議價上有更大的主動權。

所以你會看到一個有趣的現象——不是只有台灣投資人在買台積電。韓國散戶透過交易系統大量買進台積電的 ADR；美國貝萊德、富達、Vanguard 持續增持；日本的政府退休基金，全球最大的退休基金，也在加碼。大家都在找「最純的 AI 基礎設施曝險」，而台積電就是那個答案。

台積電 2025 年 75% 的收入來自北美客戶，本質上它是替美國科技業代工的台灣公司。這個角色讓它和美國科技業的成敗高度綁定，也讓它在地緣政治上處於一個微妙的位置。

往後看，有幾個觀察點：2 奈米製程 N2 進入量產的速度，決定蘋果 M5 和 Nvidia 下一代 GPU 的規格；CoWoS 產能能否如期在 2026 年底倍增，決定 AI 晶片的供貨節奏；還有亞利桑那、熊本、德勒斯登幾個海外廠的進度，這是台積電在地緣政治壓力下分散風險的佈局，但短期內台灣依然是不可取代的核心。

總結三個核心要點：第一，台積電的漲勢有基本面支撐，不是炒作——先進製程的護城河是真實的技術門檻；第二，AI 晶片的供應鏈幾乎沒有繞過台積電的路，這讓它成為全球投資者佈局 AI 最直接的標的；第三，CoWoS 封裝和 N2 製程的量產進度，是判斷台積電 2026 年能否繼續這個故事的關鍵觀察指標。

Tags

#tsmc#semiconductors#ai-chips#taiwan-stocks#investing

Related Articles