Executive Summary · Independent Case Study

TSMC: the indispensable factory of the AI era

A neutral, fully-cited teardown of the world's dominant chip foundry — the company that physically builds the silicon behind Apple, Nvidia and the AI boom — and the three questions that decide what it becomes next.

TWSE: 2330 · NYSE: TSMHsinchu · founded 1987As of 2 June 2026

TSMC does not design or sell chips of its own — it manufactures other companies' designs, and it does so at a scale and precision no rival can match. In FY2025 it earned US$122.4bn of revenue at a 59.9% gross margin, held an estimated ~66–70% of the foundry market and over 90% of the leading edge.[3]·[11]·[9] The question is no longer whether TSMC dominates — it is whether the AI cycle, the geopolitics of Taiwan, and the cost of going global all break its way at once.

$122.4bn

FY2025 revenue

+35.9% YoY

59.9%

FY2025 gross margin

62.3% in Q4

~66–70%

Foundry market share

>90% at leading edge

58%

Revenue from AI / HPC

first time, 2025

Figures from TSMC's FY2025 results and industry trackers.[2]·[3]·[9] Revenue in US$.

A cyclical business, supercharged by AI

TSMC annual revenue, US$ bn, 2020–2025. Note the 2023 downcycle before the AI-driven surge. Hover for YoY growth.

Sources: TSMC results; revenue history per market trackers.[12]·[3]

The three questions this study is organised around

This is a compilation, not an argument. On each question the evidence genuinely cuts both ways; we lay out the strongest version of each side and leave the weighing to you.

Is the leading-edge monopoly durable — or will Samsung, Intel and China close the gap?

For: TSMC holds >90% of the 3nm/2nm leading edge; reported N2 yield ~65% vs Intel 18A ~55% and Samsung SF2 ~40%, with 500+ customers. Against: Intel's 18A is improving and won Microsoft; SMIC pushes 7nm on DUV; the capital arms race never stops.[9]·[10]·[22]·[23]

Read the evidence →

Is the AI super-cycle durable enough to justify $52–56bn a year — or an overbuild?

For: AI/HPC was 58% of 2025 revenue, Nvidia reportedly locked >60% of 2026 packaging, and hyperscaler capex keeps rising. Against: CEO C.C. Wei said TSMC is 'very nervous' and careless capex 'would be a disaster'; revenue leans on two customers and peak margins.[3]·[18]·[4]·[31]

Read the evidence →

Can TSMC globalise without eroding its margins — or Taiwan's 'silicon shield'?

For: US/Japan/Germany fabs are customer- and subsidy-funded, and TSMC keeps R&D and the most advanced nodes in Taiwan ('Made with Taiwan'). Against: overseas fabs dilute gross margin 2–4pp, and advanced-process capacity in Taiwan is projected to fall from ~94% to ~79% by 2026.[27]·[29]·[28]·[15]

Read the evidence →

⚖️

What reasonable people disagree about

Whether TSMC's Taiwan concentration is its greatest strength or its greatest fragility — and whether the silicon shield protects Taiwan or makes TSMC a target. Bulls see an irreplaceable monopoly riding a multi-year AI wave with record margins and pricing power; bears see a capital-intensive, cyclical business with two dominant customers, a 2–4pp margin drag from globalisation, and exposure to the single most dangerous geopolitical flashpoint on earth.[3]·[13]·[27]·[28]

🔍

How to read this study

Every load-bearing claim carries a bracketed citation to the Sources page (tiered, with stance and language tags). Taiwanese-Chinese sources are quoted in the original alongside translation. See Methodology & Limitations for what is disclosed vs. estimated and where this study may be wrong.

Company & Timeline

The company that turned manufacturing into the moat

TSMC's founding insight was deceptively simple: separate chip design from chip-making, and become the trusted factory for everyone. Four decades later, that bet underpins the entire fabless semiconductor industry.

Founded 1987 · HsinchuChair & CEO: C.C. Wei

TSMC pioneered the pure-play foundrymodel — manufacturing chips for others without competing with them — which let companies like Apple, Nvidia and AMD go “fabless.” Born from state industrial policy (the government held 48.3%at founding) and ITRI research, it became Taiwan's most valuable company and a linchpin of the global economy.[1]·[7]

What TSMC actually does

TSMC is a contract manufacturer of semiconductors. Fabless customers — Apple, Nvidia, AMD, Qualcomm, Broadcom, MediaTek — design the chips; TSMC fabricates them in its “fabs” (fabrication plants) on the most advanced process “nodes” in the world, and increasingly packages them too. It serves 500+ customers across 10,000+ products, deliberately holding no chip brand of its own so customers never fear it will copy their designs.[10]·[7]

💡

The insight: manufacturing without rivalry

Chang reasoned that Taiwan lagged in chip design and marketing but had a real edge in manufacturing. By committing to never design or sell its own chips, TSMC could become the neutral factory the whole industry could trust — letting Intel, Apple and Nvidia outsource without fear of their designs being copied.[7]·[1]

From ITRI spin-off to “sacred mountain”

TSMC was not a garage startup — it was a deliberate act of national strategy. When Texas Instruments and Intel declined to participate, Taiwan's government funded the bulk of the venture and recruited Chang from ITRI. Domestically, TSMC is now called the 護國神山— the “sacred mountain that protects the nation” — a phrase that captures both pride and the unease of a whole economy resting on one company.[1]·[36]

Selected timeline

1985

Morris Chang, recruited by minister Li Kwoh-ting, leaves the US to head Taiwan's Industrial Technology Research Institute (ITRI).[1]

1987

TSMC is founded (Feb 21) on a new idea — a pure-play foundry that only manufactures, never designs. National Development Fund 48.3%, Philips 27.6% + tech transfer.[1]

1993 / 1997

Lists on the Taiwan Stock Exchange (1993), then becomes the first Taiwanese company on the NYSE (1997, ticker TSM).[1]

2018

Morris Chang retires after 31 years; Mark Liu becomes chairman, C.C. Wei CEO.[1]

2020

First to volume-manufacture 5nm; first to commercialise ASML's EUV lithography at scale.[8]

2022

Begins 3nm (N3) volume production; AI/HPC begins overtaking smartphones as the growth engine.[8]

2024

C.C. Wei becomes chairman and CEO; Arizona Fab 21 reaches volume production; ~83,825 employees.[1]

2025

2nm (N2, gate-all-around) enters mass production in Taiwan; announces +US$100bn US expansion to US$165bn total; revenue hits US$122.4bn.[14]·[15]·[3]

🏔️

Why the founding still matters

The state's founding stake and ITRI roots are why the “is TSMC Taiwan's or the world's?” question is so charged today: the company that globalisation now pulls abroad was, at its origin, an instrument of Taiwanese industrial policy.[1]·[36]

Market & Industry

Where TSMC sits in the chip value chain

The semiconductor industry split itself in two: companies that design chips and companies that build them. TSMC owns the most capital-intensive, highest-barrier link — fabrication — and is extending into packaging.

Pure-play foundryAI / HPC-led demand

TSMC occupies the manufacturing chokepoint of a trillion-dollar industry. Demand is now led by AI and high-performance computing (58% of 2025 revenue), with hyperscaler AI-infrastructure capex cited around US$650bn for 2026— but that same concentration ties TSMC's fortunes to one volatile end-market.[3]·[34]

The fabless–foundry split

Building a leading-edge fab now costs US$15–20bn and a single ASML EUV machine ~US$350m.[20] That economics forced a division of labour: fabless designers (Apple, Nvidia, AMD) focus on architecture and software, while foundries shoulder the capital and process science of manufacturing. TSMC made this split viable — and captured its most defensible link.

The value chain, and what TSMC owns

Tools & materials

ASML (EUV), Applied Materials, Lam, Tokyo Electron; wafer & gas suppliers

Chip design (fabless)

Apple, Nvidia, AMD, Qualcomm, Broadcom, MediaTek

Wafer fabrication

TSMC (+ Samsung, Intel, SMIC, UMC, GlobalFoundries)

← TSMC

Advanced packaging

TSMC (CoWoS / SoIC), OSATs (ASE, Amkor)

← TSMC

End systems

Servers, phones, cars, AI data centres

TSMC owns wafer fabrication and, increasingly, advanced packaging (CoWoS) — the two stages where AI value is captured today.[18]·[20]

How big, and how concentrated

TSMC's revenue is a barometer for the whole electronics cycle and for Taiwan's exports. Its market capitalisation reached roughly US$2.087tn — the world's ~6th-largest company — and it makes up over 40%of Taiwan's entire stock index, which itself ranks among the world's five largest.[5]·[6] Domestically, semiconductors are described as the “new oil” and TSMC as the key enabler of Nvidia's AI dominance.[42]

“Semiconductors function as the 'new oil', a critical input for economic activity — and TSMC is a key supplier for the AI giant Nvidia.”

original · zh ·“半導體是『新石油』…台積電是AI巨頭輝達的關鍵供應商。”

自由財經 (Liberty Times Net) · Taiwanese financial press · 2025/2026 · English is a translation from zh · source

Why the industry position is strong

TSMC sits at the highest-barrier link; entry needs $15–20bn per fab and scarce EUV tools.[20]
Demand is structurally rising with AI/HPC — 58% of 2025 revenue, up from 40% in 2022.[3]·[13]
It increasingly owns advanced packaging (CoWoS), the current AI bottleneck.[18]

Why the position carries risk

Concentration in AI/HPC means a hyperscaler pullback would hit demand fast.[31]
The foundry business is historically cyclical — revenue fell 9% in 2023, and domestic economists warn the AI peak is unpredictable.[12]·[49]
Mature-node oversupply from China threatens pricing at the trailing edge.[39]

📊

Bottom line

TSMC owns the most valuable real estate in semiconductors — but its tenant base has narrowed toward AI. That is a strength while the boom lasts and a concentration risk if it doesn't.[3]·[31]

Business Model & Economics

How a factory earns 60% gross margins

TSMC sells wafers, priced by how advanced the process is. Leading-edge scarcity, near-total utilisation and relentless pricing power turn a capital-heavy manufacturer into one of the most profitable companies on earth — for now.

FY25 gross margin 59.9%~US$30k per 2nm wafer

The economics rest on a simple loop: only TSMC can reliably make the newest chips, so it charges a premium (a 2nm wafer reportedly tops ~US$30,000), reinvests record capex to stay ahead, and repeats. Average wafer prices have risen ~15% a year since 2019.[14]·[19] The risk: those same fixed costs cut hard the other way if demand softens.

59.9%

FY2025 gross margin

62.3% in Q4 2025

54.0%

Q4 2025 operating margin

net margin 48.3%

77%

Revenue from ≤7nm

advanced nodes, Q4 2025

+5–10%

2026 advanced-node price hikes

2nm wafers +~50%

Margins per TSMC results; pricing per trade reporting.[2]·[19]

Revenue by process node

TSMC's money is made at the leading edge. In Q4 2025, 3nm was 28% and 5nm 35% of revenue; nodes at 7nm and below made up 77%.[2] Newer nodes command far higher prices, which is why staying one step ahead of Samsung and Intel is the whole game.

Revenue by technology node (Q4 2025)

Share of wafer revenue. 3nm + 5nm alone are nearly two-thirds. Hover to isolate.

3nm

28%

5nm

35%

7nm

14%

≥10nm / other

23%

Who pays the bills — and the concentration that creates

In 2025, Nvidia overtook Appleas TSMC's largest customer for the first time, at an estimated ~19% of revenue versus Apple's ~17%; a year earlier Apple was ~25% and Nvidia under 11%.[13] Two customers now drive roughly a third of revenue — extraordinary demand, but also extraordinary dependence.

Nvidia — 19% (19%)
Apple — 17% (17%)
Other customers (≈500) — 64% (64%)

Estimates; figures vary by source (Nvidia quoted ~19–22%, Apple ~17–18%).[13]

Pricing power vs. capital intensity

Because there is no real alternative at the leading edge, TSMC can raise prices and customers largely absorb it: advanced-node quotes rise another 5–10% in 2026, and 2nm wafers reportedly cost 50%+ more than their predecessors.[19] The offsetting reality is brutal capital intensity — ~US$40.9bn of capex in 2025, guided to US$52–56bn in 2026 — and a stated long-term gross-margin target of only ~53%, below today's level.[33]·[40]

Why the model prints money

Leading-edge scarcity → durable pricing power (+15%/yr wafer prices since 2019).[19]
High utilisation + rich node mix lifted FY25 gross margin to 59.9%.[2]·[40]
Switching foundries mid-design is painful, so customers prepay and lock capacity.[14]

Why it's more fragile than it looks

Enormous fixed costs ($52–56bn capex) make profit highly sensitive to utilisation.[31]
Customer concentration: ~36% of revenue from two customers.[13]
Management's own long-term gross-margin target is ~53%, below current levels; overseas fabs dilute further.[40]·[27]

📊

Bottom line

The disclosed margins are not seriously disputed — but they reflect a near-monopoly running near peak utilisation. Bulls read that as durable pricing power; bears as operating leverage that cuts both ways through the cycle.[40]·[31]

Competitive Landscape

A monopoly at the frontier, a crowd at the trailing edge

On the newest nodes TSMC has almost no peer; its rivals are years and yield-points behind. But the industry's structure — supplier power and geopolitics — still constrains even the leader.

~66–70% foundry share>90% at 3nm/2nm

At the leading edge this is effectively a monopoly: TSMC's reported 2nm yield (~65%) is far ahead of Intel's 18A (~55%) and Samsung's SF2 (~40%).[10]The real competitive pressure isn't a rival foundry — it's supplier dependence on ASML and the geopolitical risk of where the fabs sit.

Porter's Five Forces

Click each force to see its rated pressure and the evidence behind it. Ratings are this study's reading of sourced facts, not TSMC's.

Leading-edge foundry

New entrants · Low pressure

A leading-edge fab costs US$15–20bn and each EUV scanner ~US$350m; decades of process know-how are needed. Even Intel — a cash-rich incumbent — runs foundry losses trying to re-enter.[20]·[22]

Click a force to read the rated pressure and its basis. Ratings are this study's analysis of sourced evidence.

Where everyone sits

Only three firms are seriously pursuing sub-3nm — TSMC, Samsung and Intel — and only TSMC combines frontier process leadership with massive scale and customer breadth. SMIC has scale but is capped at 7nm by export controls; UMC and GlobalFoundries left the leading-edge race entirely.[9]·[23]·[44]·[50]

Foundry positioning — process leadership × scale (hover a point)

Hover a company to see the basis for its placement (this study's analysis).

Why the lead looks unassailable

A multi-year, multi-point yield lead at 2nm over both Samsung and Intel.[10]
Rivals are bleeding: Intel Foundry posted a ~$2.3bn quarterly loss; Samsung can't win marquee 2nm customers.[9]·[21]
Export controls cap SMIC at 7nm without EUV, keeping China off the frontier.[23]

Why rivals (and reality) could close in

Intel's 18A is improving fast (~20–30% → 50–55%, reportedly ~7%/month) and won Microsoft.[22]·[48]
China is pouring billions into mature nodes and equipment, threatening trailing-edge pricing.[39]
TSMC cannot advance without ASML; a tool or supply shock hits the whole industry.[20]

⚖️

What's genuinely contested

Not whether TSMC leads, but for how long, and how much it matters that one supplier (ASML) and one island (Taiwan) sit upstream of the entire frontier.[20]·[28]

Strategy, Moats & Globalization

The flywheel — and the price of going global

TSMC's stated strategy is simple: stay one node ahead, keep customers' trust, and reinvest everything. Its revealed strategy now adds a harder problem — manufacturing abroad without giving up the economics or the home-island advantage that made it.

N2 GAA in production$165bn US plan

The moat is a compounding flywheel: process leadership wins the best customers → their volume funds the next node and the CoWoS packaging AI needs → that lead widens. Globalisation is the strategic tax on it — overseas fabs dilute gross margin 2–4pp, which TSMC accepts to de-risk geography and keep customers (and governments) happy.[18]·[27]

The sources of durable advantage

Process leadership. First to 7nm, 5nm and EUV at volume; 2nm (N2) — its first gate-all-around node, ~10–15% faster or 25–30% lower power than 3nm — entered mass production in late 2025.[8]·[17]
Trust through neutrality. By never designing chips, TSMC is the one foundry customers trust with their crown-jewel designs — a moat Samsung (a competitor to its own customers) and Intel structurally lack.[7]
Ecosystem & scale. 500+ customers, 10,000+ products, and the design-tool/IP ecosystem around its nodes create switching costs measured in years.[10]
Advanced packaging. CoWoS — stacking compute and HBM memory — is the current AI bottleneck, and TSMC is scaling it from ~35k to ~120–140k wafers/month by end-2026, with Nvidia reportedly taking >60%.[18]

▸ What is CoWoS, and why does it suddenly matter so much?

CoWoS (“chip-on-wafer-on-substrate”) places multiple compute dies and stacks of high-bandwidth memory on a silicon interposer, so an AI accelerator behaves like one giant chip. Almost every frontier AI GPU needs it, and capacity — not raw wafer fabrication — has been the binding constraint on AI hardware. Owning both the node and the packaging is why TSMC captures so much of the AI value chain.[18]·[34]

SWOT — applied even-handedly

Strengths

>90% leading-edge share; multi-point 2nm yield lead.[10]
Pure-play trust; 500+ customers, 10k+ products.[7]·[10]
Pricing power (+15%/yr wafer prices since 2019).[19]
CoWoS packaging lock-in for AI accelerators.[18]

Weaknesses

Almost all leading-edge output is in Taiwan — concentration.[29]
Customer concentration: ~36% from two customers.[13]
Overseas fabs dilute gross margin 2–4pp.[27]
Capital intensity: $52–56bn capex in 2026.[33]

Opportunities

AI/HPC secular demand — 58% of 2025 revenue.[3]
2nm / A16 ramp + advanced-packaging expansion.[17]·[18]
US / Japan / Germany fabs with host subsidies.[15]·[41]
Price hikes the leading edge can absorb.[19]

Threats

Taiwan–China geopolitics; “silicon shield” debate.[28]
AI overbuild / demand normalisation.[31]
Samsung/Intel catch-up; China mature-node glut.[22]·[39]
ASML single-supplier dependency.[20]

Going global: Arizona, Japan, Germany

On 3 March 2025 TSMC announced an additional US$100bn US investment — lifting its Arizona commitment to US$165bn(three more fabs, two packaging plants, an R&D centre) on top of the prior $65bn — and reports suggest the figure could rise further under a US–Taiwan tariff framework.[15]·[16] It is also building in Kumamoto, Japan and Dresden, Germany, largely funded by customers and host-government subsidies.[41]

Why globalisation is smart strategy

De-risks the single-island concentration customers and governments now demand.[15]
Subsidies + customer co-investment defray the cost; tariffs make local US output valuable.[16]
R&D and the most advanced nodes stay in Taiwan — “Made with Taiwan.”[29]

Why it's a costly hedge

Overseas fabs dilute gross margin 2–4pp — a real, ongoing cost.[27]
US fabs are higher-cost and harder to staff; ramps have slipped historically.[26]
Each node moved abroad feeds the domestic “de-Taiwanization” debate.[28]

🧭

Stated vs. revealed strategy

TSMC saysTaiwan remains its core and R&D never leaves; it is simultaneously committing nine-figure sums abroad. Both are true — the strategy is a dual-core hedge, and its success hinges on keeping the frontier at home while satisfying customers and Washington.[29]·[15]

Peer Comparison

TSMC vs. the field

A like-for-like look at the foundry peer set. Two firms chase TSMC at the leading edge; the rest compete on mature and specialty nodes. Shares and yields are third-party estimates — treat them as directional.

Estimates · as of 2026

No peer matches TSMC on bothprocess leadership and profitability. Samsung and Intel can build advanced nodes but lag on yield and lose money on foundry; SMIC, UMC and GlobalFoundries don't compete at the frontier at all.[9]·[10]·[37]

Estimated foundry revenue share

Global foundry market share (estimated, 2024–25)

Approximate share of foundry revenue. TSMC's ~67% dwarfs the field. Hover to isolate.

TSMC

~67%

Samsung Foundry

~9%

SMIC

~6%

UMC

~5%

GlobalFoundries

~5%

Intel Foundry

~4%

Estimates from trade trackers; TSMC ~64.9% (Q3 2024) rising toward ~66–70% in 2025.[9]·[23]

Side-by-side

Foundry	Leading node / yield	Profitability	Ownership / HQ	Key constraint
TSMC	2nm (N2) ~65% yield[10]	FY25 gross margin 59.9%[2]	Public · Taiwan	Taiwan concentration; ASML[29]·[20]
Samsung Foundry	SF2 ~40% yield[10]	Reportedly loss-making at leading edge[37]	Public · South Korea	Yield; competes with own customers[21]
Intel Foundry	18A ~50–55% yield[22]	Operating losses (~$2.3bn/qtr 2024)[9]	Public · USA	Thin external customer base[22]
SMIC	7nm via DUV (no EUV)[23]	Profitable; lower margins	Public · China	Export controls; ~84% China revenue[23]
UMC	Mature / specialty (≥14nm)[44]	Profitable; mid-margin	Public · Taiwan	No leading-edge presence
GlobalFoundries	Differentiated mature nodes[44]	Profitable; mid-margin	Public · USA	Exited leading edge (2018)

Yields and shares are reported estimates that move quarter to quarter; live market caps are deliberately omitted (see Methodology).

⚖️

Reading the table fairly

Samsung and Intel are credibleat the frontier — they are not vapourware, and per domestic reporting Intel's 18A yield (~55%) has surpassed Samsung's and is narrowing the gap to TSMC's ~65%. The honest gap is one of yield, customers and money-making, not of whether a 2nm process exists.[22]·[47]·[10]

Financials & Growth

Record profits, record spending

2025 was a blowout: revenue up 36%, gross margin back near 60%, profit at records. The same year set up the central financial tension — capex is climbing toward $56bn, raising both the upside and the downside.

FY25 net margin ~48%2026 capex $52–56bn

TSMC's financials are disclosed and audited(unlike many peers' estimates here): FY2025 revenue US$122.4bn (+35.9%), gross margin 59.9%, Q4 EPS NT$19.50 (full-year NT$66.25), an eighth straight quarter of profit growth.[3]·[2]·[11] The watch-item is capex: a record ~$40.9bn in 2025, guided to $52–56bn in 2026.[33]

$122.4bn

FY2025 revenue

+35.9% YoY

59.9%

FY2025 gross margin

+3.8pp YoY

NT$66.25

FY2025 EPS

+46.4% YoY

$40.9bn

FY2025 capex

→ $52–56bn in 2026

Gross margin: cyclical, but structurally high

Margins track the cycle and the node mix. They dipped to 54.4% in the 2023 downturn, then climbed back to a full-year 59.9% in 2025 (62.3% in Q4) as the rich AI/HPC mix and high utilisation kicked in.[2]·[40]Management guides to a long-term gross margin of ~53%, implying today's level reflects unusually favourable conditions.[40]

Gross margin — down in the cycle, up on AI mix

TSMC full-year gross margin, %. The 2023 trough and 2025 recovery. Hover for context.

FY2025 gross margin 59.9% per results; prior years approximate from reported figures.[2]·[40] 2022–24 are approximate.

Capex: the bet that scares and excites

Capital spending is the clearest read on management conviction — and the biggest source of operating leverage. After holding flat in 2023–24, capex jumped to ~$40.9bn in 2025 and is guided to a record $52–56bn in 2026 (~70–80% on advanced logic).[33]·[3] Roughly three-year fab lead times mean this spend is a bet on AI demand lasting into ~2029.[32]

Capital expenditure, US$ bn (2026 = guidance midpoint)

Capex history; 2026E shown at the ~$54bn midpoint of the $52–56bn guide. Hover for detail.

2025 capex ~$40.9bn; 2026 guidance $52–56bn. Earlier years approximate from disclosures/press.[33]·[12]

The bull read on the numbers

Revenue +36% and margin near 60% show pricing power and operating discipline.[3]·[2]
Eight straight quarters of profit growth; EPS +46% in 2025.[2]·[3]
Record capex is customer-demand-led, not speculative, per management.[32]

The bear read on the numbers

$52–56bn capex sharply raises fixed costs and depreciation; FCF gets sensitive to demand.[31]
Long-term gross-margin target (~53%) sits below current levels.[40]
Overseas fabs add a 2–4pp structural margin drag; domestic press already ties capex/overseas costs to a reported employee-bonus squeeze.[27]·[46]

📊

Bottom line

The disclosed financials are excellent and not seriously disputed. The debate is entirely forward-looking: whether the record capex is prudent investment or pro-cyclical overbuild.[32]·[31]

Risks & Geopolitics

The most important company in the most dangerous place

TSMC's risks are unlike a normal company's. Beyond cyclicality and concentration sit a contested geopolitical theory ('the silicon shield'), the physical limits of a small island, and a domestic fear of being hollowed out.

Taiwan concentration護國神山 · 矽盾

The single biggest risk isn't competitive — it's that nearly all leading-edge production sits in Taiwan, beside a geopolitical flashpoint, on an island with real water, power and talentlimits. Whether globalising abroad reduces that risk or hollows out Taiwan's “silicon shield” is genuinely contested.[28]·[29]·[26]

The “silicon shield” — protection or target?

The 矽盾 (“silicon shield”)theory holds that because the world cannot function without Taiwan's chips, the world is compelled to protect Taiwan. As TSMC builds abroad, Taiwanese analysts debate whether that shield is eroding into a “Taiwan–US dual-core” model.[28]

“Traditional silicon-shield logic rests on Taiwan's indispensable, singular capacity… TSMC's US expansion signals a shift from a Taiwan-centric to a Taiwan–US dual-core architecture.”

original · zh ·“過去傳統矽盾的邏輯建立在台灣產能的不可或缺性與單一性上…台積電若再度對美擴大投資，則象徵公司從台灣單一核心轉型為台美雙核心架構。”

STPI / 工研院 iKnow · Taiwan science-policy research institute · 2025 · English is a translation from zh · source

The counter-view, also from Taiwan and from US academics, is that de-Taiwanization fears are overstated for decades: reliable capacity outside Taiwan is a 2030s problem at the earliest, and the island's talent ecosystem is “virtually irreplaceable.”[30]

“Reliable manufacturing capacity outside Taiwan needs at least 2030 or 2035; true resilience may require waiting until mid-century or later.”

original · zh ·“要達到『在台灣之外建立可靠製造能力』的階段，至少需要2030或2035年…恐怕要等到本世紀中葉、甚至更晚。”

Matt Seitz · University of Wisconsin–Madison AI program (via 風傳媒) · 2025 · English is a translation from zh · source

De-Taiwanization, by the numbers

Taiwan's government defines “advanced” as sub-5nm. About 94% of that capacity was in Taiwan in 2025, projected to fall to ~79% by 2026 as ~21% moves overseas.[29] Taipei's response has been to keep the crown jewels home: it amended the National Security Act to protect sub-14nm manufacturing tech, and TSMC stresses that all R&D stays in Taiwan — a shift it frames as “Made with Taiwan,” not “Made in Taiwan.”[29]

The physical limits: water, power, talent

Even setting geopolitics aside, the island strains to support TSMC. Analysts have called water and power “the Achilles heel of Taiwan's semiconductors”; TSMC itself estimates a serious drought could cut revenue 0.7–1.1%, and its chairmen have repeatedly named talent shortage as the industry's biggest constraint amid low birth rates.[26]·[25]·[43] A 7.4-magnitude earthquake in April 2024 cost ~US$92m but was shrugged off within hours — resilience and fragility in the same event.[24]

Why the risks may be overstated

Taiwan's ecosystem is irreplaceable for decades; US-made leading edge is “costly and inefficient.”[30]
R&D and the frontier stay in Taiwan; law now protects sub-14nm tech.[29]
TSMC restored 80% of operations within hours of a major earthquake.[24]

Why the risks are serious

Nearly all leading-edge output is in one geopolitical flashpoint.[28]·[29]
Water, power and talent are structural domestic constraints.[26]·[43]
Globalising abroad dilutes margins and, critics say, the silicon shield itself.[27]·[28]

⚠️

Where this gets genuinely uncertain

The geopolitics cannot be modelled. A Taiwan Strait crisis is low-probability but catastrophic-impact, and reasonable, well-informed people in Taipei and Washington disagree on whether TSMC's global build-out makes Taiwan safer or more expendable. This study presents the competing framings rather than picking one.[28]·[30]

Forward View

Three ways the next few years could break

Not predictions — possibilities, with the conditions that would trigger each and what to watch. TSMC's own management captures the tension better than any outside analyst.

Scenarios, not forecasts

TSMC's future turns on three variables it only partly controls: the durability of AI demand, the geopolitics of Taiwan, and the cost of globalising. The most telling signal is that management is simultaneously spending a record $52–56bn and calling itself “very nervous.”[3]·[4]

“If we did not do it carefully, that would be a disaster for TSMC for sure.”

C.C. Wei · TSMC Chairman & CEO, on AI-bubble risk while announcing record 2026 capex · 15 Jan 2026 · source

The same executive committing record capex also flagged the downside — a candour that frames the whole forward debate.[4]·[32]

Scenarios

Bull case

The AI build-out is durable and TSMC compounds

AI/HPC demand stays structural; 2nm, A16 and CoWoS sustain pricing and a rich mix; Taiwan keeps the frontier and globalisation de-risks geography without heavy dilution. Margins hold in the high-50s and TSMC widens its lead.[32]·[17]·[18]

Watch: Hyperscaler capex holds; 2nm yields ramp; overseas dilution stays ≤2pp.

Base case

Growth moderates, dominance persists

AI growth cools from breakneck to merely strong; overseas fabs shave 2–4pp off gross margin; TSMC still holds ~60%+ foundry share and its leading-edge monopoly. Geopolitics simmers without rupture.[27]·[40]

Watch: Capex digested without overbuild; margin settles toward the ~53% long-term target.

Bear case

Overbuild meets a demand air-pocket — or a Strait shock

AI capex outpaces real demand; utilisation and pricing fall just as record fixed costs land, compressing margins and free cash flow. A rival yield catch-up, a China mature-node glut, or a Taiwan crisis amplifies the hit.[31]·[39]·[28]

Watch: Hyperscaler capex cuts; 2nm price resistance; any cross-Strait escalation.

The questions that actually decide it

Does AI demand stay structural? If hyperscaler capex (~$650bn in 2026) holds, the capex bet pays off; if it air-pockets, TSMC's fixed costs bite.[34]·[31]
Does anyone close the yield gap? Intel's 18A and Samsung's SF2 are the variables; a genuine second source would dent TSMC's pricing power.[22]·[10]
Does the silicon shield hold? The lowest-probability, highest-impact variable — and the one no model can price.[28]·[30]

⚖️

The honest synthesis

On the evidence surveyed here, TSMC is the most dominant of the foundries — and also one of the most exposed to a single demand cycle and a single geopolitical fault line. Both can be true at once. The bull and bear cases share the same facts; they differ on which risk you weight.[3]·[31]·[28]

Methodology & Limitations

How this was made — and where it may be wrong

A research artifact is only as good as its honesty about its own limits. This page states the method, the frameworks, what's disclosed vs. estimated, and the as-of caveat.

As of 2 June 2026Independent

Method

Research proceeded by fan-out web search and direct fetching of primary and reputable secondary sources, working per section from TSMC's own results filings and investor materials outward to trade analysts, press and tertiary context. Because TSMC is a Taiwanese company, a substantial share of the work was done natively in Traditional Chinese (經濟日報, 自由財經, 聯合新聞網, 風傳媒, TVBS, STPI/工研院 iKnow), including deliberate disconfirming searches (去台化 / 矽盾 / 缺水缺電 / 人才短缺 / 護國神山外移), with Chinese quotes shown in the original alongside an English translation. Every URL cited was opened and read during the research run; no link appears unless it was fetched, and a link-checker validated each one while quotes were spot-checked against the source. Each claim was then transcribed into a structured manifest tagging it with a tier, a confidence level and a stance. The load-bearing figures for TSMC are its FY2025 and Q4 2025 financials and operating metrics — revenue, gross margin, capex and node mix — together with the (estimated) ~67% foundry share and >90% leading-edge position on which most of the competitive argument rests.[2]·[3]·[11]

Frameworks used

The study applies the Pyramid Principle (answer-first, balanced) throughout, Porter's Five Forces and a 2×2 positioning map in the Competitive Landscape, peer comparables in Peer Comparison, a value-chain lens and node-mix breakdown in the Business Model, SWOT in Strategy, and scenario analysis in the Forward View — each chosen because the disclosed data could support it even-handedly. Frameworks the data could not fill that way, such as McKinsey 7S or a full BCG portfolio for what is effectively a single-business-line company, were deliberately skipped rather than padded out with guesswork.

Disclosed vs. estimated

A clear line separates what TSMC discloses from what others estimate. The disclosed, high-confidence layer is TSMC's own financials and operating metrics, drawn from its FY2025 and Q4 2025 results and investor materials.[2]·[3]·[11] Comparable-basis figures — market-share percentages (foundry and leading-edge), competitor yield rates for TSMC, Intel and Samsung, CoWoS capacity, customer revenue shares for Nvidia and Apple, and overseas-fab margin dilution — are directional analyst and press estimates rather than TSMC disclosures, and are flagged as such wherever they appear.[9]·[10]·[13]·[18]·[27]Where even a defensible estimate could not be cleanly re-sourced at the as-of date — precise per-customer revenue and a live peer market-cap table — a range is shown or the figure is excluded rather than guessed.

⚠️

Where this case study may be wrong

Competitor yields are estimates. The TSMC N2 ~65% / Intel 18A ~50–55% / Samsung SF2 ~40% figures are reported, not audited; yields move fast and sources diverge.[10]·[21]·[22]
Customer-share figures vary by source. Nvidia's 2025 share is quoted from ~19% to ~22% and Apple's ~17–18%; we show the range rather than assert one.[13]
CoWoS capacity and the 60% Nvidia allocation are press estimates that shift with each earnings cycle.[18]
Margin dilution from overseas fabs (2–4pp) is TSMC's own guidance framing, not a realised outcome; the actual drag depends on subsidies, utilisation and mix.[27]
The “silicon shield” debate is contested and political — we present multiple Taiwanese and Western framings rather than endorsing one.[28]·[30]·[29]
SEC filing pages block automated fetching (HTTP 403); headline FY/Q4 figures were corroborated across TrendForce and press that quote the same filing.[2]·[11]
As-of staleness. This is a point-in-time artifact as of 2 June 2026. TSMC reports monthly revenue and quarterly results; anything after that date (new results, capex revisions, 2nm ramp, tariff/geopolitical events) is not captured.

Neutrality & independence

This is a compilation, not an argument: every section pairs the case for and the case against, with supporting and critical evidence tagged by stance and frameworks applied even-handedly. It is an independentresearch artifact, not affiliated with, endorsed by, or sponsored by Taiwan Semiconductor Manufacturing Company Limited, provided for informational and educational purposes only and not investment advice; critical claims are attributed to named sources and framed as reporting or analysis, never as bald fact. It is point-in-time as of 2 June 2026, and errors are the author's — corrections welcome.

Sources

The evidence base

Every load-bearing claim on this site traces to a source below — fetched and read during the research run. Tiered, with stance and language tags so the balance of the evidence is visible, not asserted.

50 sources3 Tier-141 Tier-26 Tier-334% Chinese-language

Stance mix

21 supporting · 15 critical · 14 neutral

Source tiers

3 primary · 41 reputable · 6 tertiary

Language

17 Chinese (34%) · 33 English

Tier 1 = primary / authoritative (SEC filings, TSMC investor relations and press releases). Tier 2 = reputable secondary (Reuters/CNBC, TrendForce, DIGITIMES, 經濟日報, 自由財經, 風傳媒, STPI/工研院). Tier 3 = tertiary / context (encyclopedias, market trackers, commentary) — used for colour, not load-bearing facts.

Executive Summary · 1

[45]Tier 2neutralMedium confidence
TSMC is the indispensable manufacturer of the AI era — pure-play foundry leader at ~66–70% revenue share and >90% at the leading edge — but its future turns on the durability of the AI cycle, the geopolitics of Taiwan, and whether it can globalise without eroding margins or the 'silicon shield.'
“TSMC riding the AI wave amid global expansion.”
SemiWiki — TSMC 2025 Update (overview)· fetched 2026-06-02

Company & Timeline · 4

[1]Tier 3neutralHigh confidence
TSMC was founded Feb 21, 1987 by Morris Chang at ITRI; Taiwan's National Development Fund held 48.3%, Philips 27.6% (with tech transfer), pioneering the world's first pure-play foundry. Listed TWSE 1993, NYSE 1997. C.C. Wei became chairman & CEO June 2024; ~83,825 employees (2024).
“Morris Chang established TSMC on February 21, 1987... National Development Fund 48.3%, Philips 27.6%... the world's first dedicated semiconductor foundry.”
TSMC — Wikipedia· fetched 2026-06-02
[7]Tier 3supportingHigh confidence
Morris Chang's pure-play insight: Taiwan lagged in IC design and marketing but had manufacturing strength, so TSMC would only fabricate for customers — never design or sell its own branded chips — letting Intel, Apple and Nvidia outsource without fear of having their designs copied.
“TSMC became the world's first dedicated semiconductor foundry... the contract manufacturing model competitors like Samsung and Intel initially couldn't match.”
TSMC — Wikipedia (pure-play foundry model)· fetched 2026-06-02
[8]Tier 3supportingHigh confidence
TSMC was first to volume-market 7nm and 5nm, and first to commercialise ASML's EUV lithography at volume; it now produces 3nm and began 2nm (N2) in late 2025.
“First to market 7nm and 5nm nodes; first to commercialize ASML's EUV lithography at volume.”
TSMC — Wikipedia (process firsts)· fetched 2026-06-02
[36]Tier 2criticalzhMedium confidence
TSMC is Taiwan's '護國神山' (sacred mountain), but its dominance now raises a 'TSMC disease' concentration worry: it is over 42% of the TAIEX (vs Nvidia ~6.7% of the US market, Samsung ~29% of Korea's), so the island's stock market, exports and strategic standing all ride on one company.
“TSMC accounts for over 42% of Taiwan's market weight; a single company carries the stock market, exports, visibility and strategic security — a 'TSMC disease', not the classic Dutch disease.” · orig: “台積電目前佔台股權重超過42%…單一企業承載股市、出口、國際能見度與戰略安全感…形成『台積電病』而非傳統『荷蘭病』。”
聯合新聞網 — 台積電病？台股躍升全球第五大背後的危險集中· fetched 2026-06-02

Market & Industry · 5

[5]Tier 2neutralzhHigh confidence
TSMC's market cap was ~US$2.087tn, the world's ~6th largest company; it makes up over 40% of Taiwan's weighted index, and Taiwan's market (~US$4.3tn) surpassed the UK.
“TSMC market cap US$2.087 trillion, ranked 6th globally; TSMC comprises over 40% of Taiwan's market weight.” · orig: “台積電市值2.087兆美元（65.71兆台幣），排名全球第6大企業…佔台股比重逾40%的台積電”
自由財經 — AI助威！台股市值直逼加拿大台積電躍居全球第6大企業· fetched 2026-06-02
[6]Tier 2neutralzhHigh confidence
Taiwan's stock market rose to ~US$4.95tn, the world's 5th largest (behind the US, mainland China, Japan, Hong Kong); TSMC accounts for ~42% of the TAIEX and was up ~49% year-to-date on the AI boom.
“TSMC currently accounts for approximately 42% of Taiwan's weighted index; the stock has surged 49% year-to-date.” · orig: “目前台積電占台灣加權指數約42%…台積電股價今年來已大漲49%。”
聯合新聞網 — 台積電驚人漲幅帶動台股市值躍全球第五大· fetched 2026-06-02
[34]Tier 3supportingMedium confidence
The semiconductor foundry model splits the industry: fabless designers (Apple, Nvidia, AMD, Qualcomm) design chips while pure-play foundries manufacture them; advanced-node demand is driven by AI, HPC and smartphones, with hyperscaler AI infrastructure capex cited around US$650bn for 2026.
“With hyperscaler capex commitments of $650 billion for AI infrastructure in 2026 alone, the constraint environment is expected to persist through 2027.”
AI in Asia — TSMC's AI Packaging Capacity· fetched 2026-06-02
[42]Tier 2neutralzhMedium confidence
TSMC's revenue tracks the global electronics cycle and Taiwan's electronics exports; semiconductors are described domestically as the 'new oil' and TSMC as the key supplier enabling Nvidia's AI dominance.
“Semiconductors function as the 'new oil'... TSMC is a key supplier for AI giant Nvidia.” · orig: “半導體是『新石油』…台積電是AI巨頭輝達的關鍵供應商。”
自由財經 — 台積電與台股、台灣經濟· fetched 2026-06-02
[49]Tier 2criticalzhMedium confidence
Taiwanese commentators warn the AI cycle's peak is unpredictable: the head of CIER (連賢明) cautioned that an overheated market usually corrects harder than expected, and that US industrial policy may not always align with Taiwan's interests — a caution against extrapolating the boom.
“No one can foretell the peak or the cooling; after markets overheat, the correction is usually larger than imagined.” · orig: “沒有人可以預告高點與退燒之時…金融市場過熱後，修正風險會隨之而來，且修正幅度會比想像更大。”
聯合新聞網 — 台積電病：AI景氣循環隱憂· fetched 2026-06-02

Business Model & Economics · 4

[13]Tier 2criticalMedium confidence
In 2025 Nvidia overtook Apple as TSMC's largest customer (~19–22% of revenue vs Apple ~17–18%); in 2024 Apple was ~25% and Nvidia under 11%. HPC was 55% of Q4 2025 revenue, up from 40% in 2022.
“Nvidia ended 2025 as TSMC's largest customer, ~19% of revenue, nudging Apple's 17% into second; HPC made up 55% of Q4 revenue.”
CNBC — Nvidia set to supplant Apple as TSMC's largest customer· fetched 2026-06-02
[14]Tier 2supportingzhHigh confidence
TSMC's 2nm (N2) entered mass production in Q4 2025 at a reported ~US$30,000 per wafer, with six anchor customers — Apple, AMD, Qualcomm, MediaTek, Broadcom and Intel — competing for capacity.
“N2 mass production expands as scheduled in Q4 2025; foundry pricing ~US$30,000 per wafer; six clients: Apple, AMD, Qualcomm, MediaTek, Broadcom, Intel.” · orig: “台積電2奈米（N2）製程…於2025年第4季擴大量產…代工報價上看每片約3萬美元…蘋果、超微、高通、聯發科、博通與英特爾。”
經濟日報 — 台積電2奈米第4季如期放量每片約3萬美元六大客戶搶產能· fetched 2026-06-02
[19]Tier 2supportingMedium confidence
TSMC has raised average wafer prices ~15%/year since 2019; advanced-node quotes rise another 5–10% in 2026, and 2nm wafers reportedly cost ~50%+ more than predecessors — evidence of leading-edge pricing power.
“TSMC intends to raise advanced-node wafer prices by 5% to 10% starting 2026; 2nm wafer prices surge over 50% vs predecessors.”
Tom's Hardware — TSMC wafer prices up 15%+ each year since 2019· fetched 2026-06-02
[40]Tier 2supportingHigh confidence
TSMC's gross margin rose to a full-year 59.9% in 2025 (62.3% in Q4), reflecting leading-edge mix, high utilisation and pricing power — though management's stated long-term gross-margin target is ~53% and overseas fabs are a drag.
“Q4 2025 gross margin 62.3%, operating margin 54.0%.”
TrendForce — TSMC Q4 2025 margins· fetched 2026-06-02

Competitive Landscape · 9

[9]Tier 2supportingHigh confidence
TSMC held ~64.9% of foundry revenue in Q3 2024 (vs Samsung 9.3%); projected ~66% in 2025, and upwards of 90% market share at the 3nm and 2nm leading edge.
“TSMC 64.9% (up 2.6 points), Samsung 9.3%... Intel reported $2.3B foundry loss in Q4 2024.”
TrendForce — The 2nm Foundry Battle: TSMC Leads· fetched 2026-06-02
[10]Tier 2supportingzhMedium confidence
On comparable advanced-node yields, reporting put TSMC's N2 at ~65%, Intel's 18A at ~55% and Samsung's SF2 at ~40%; TSMC serves 500+ customers across 10,000+ products.
“TSMC N2 yield ~65%, Intel 18A ~55%, Samsung SF2 ~40%; over 500 customers, 10,000+ products.” · orig: “台積電 N2 良率65%、英特爾 18A 55%、三星 SF2 40%…客戶超過500家，生產逾萬種產品。”
經濟日報 — 整理包／台積電2奈米將量產全球先進製程賽局· fetched 2026-06-02
[20]Tier 2supportingMedium confidence
Barriers to leading-edge entry are extreme: each ASML EUV scanner costs ~US$350m and a cutting-edge 3nm fab runs US$15–20bn, anchoring TSMC's position and its pricing power.
“Each ASML EUV machine costs around $350 million... a 3nm fab costs between $15 billion and $20 billion.”
Tom's Hardware — EUV economics underpin TSMC pricing power· fetched 2026-06-02
[21]Tier 2supportingMedium confidence
Samsung's 2nm (SF2) yields were reported still below the level needed for stable mass production (GAA yield ~10–20%, ~40% in some reports), making it hard to win major foundry customers.
“Yields are reportedly still below the level required for stable mass production, raising questions about securing major foundry customers.”
DIGITIMES — Samsung 2nm yields reportedly remain below mass-production threshold· fetched 2026-06-02
[22]Tier 2criticalMedium confidence
Intel returned to group profit but its foundry unit stayed loss-making; 18A yields improved from ~20–30% early in the year to ~50–55%, and Intel claims external 18A customers (incl. Microsoft).
“Intel's foundry business remains firmly in the red while 18A yields persist; 18A improved from ~20-30% to 50-55%.”
AnySilicon — Global Foundry Earnings Q3 2025· fetched 2026-06-02
[23]Tier 2neutralMedium confidence
SMIC became the world's third-largest foundry (~5–6% share), producing 7nm via DUV multi-patterning without EUV; ~84% of its revenue is from China and yields/costs lag the leading edge.
“7nm on DUV requires 34 lithography steps vs nine with EUV, raising costs and lowering yields; ~84.1% of revenue from China.”
Andy Lin — How is SMIC after the US embargo?· fetched 2026-06-02
[39]Tier 2criticalMedium confidence
China is investing heavily to localise mature-node and equipment capacity; Chinese foundries are adding trailing-edge supply that risks oversupply and price pressure at mature nodes, even as EUV-dependent leading edge stays out of reach.
“China invests billions to close critical chokepoints in its semiconductor supply chain.”
EE Times — China Invests Billions to Close Critical Chokepoints· fetched 2026-06-02
[48]Tier 2criticalzhMedium confidence
Intel says its 18A yield is improving ~7% per month and will reach mass production by end-2025 with Panther Lake — though domestic reporting notes the yield sat around 10% until mid-year, far below the 70–80% needed for volume, underscoring how hard catching TSMC is.
“18A yield improves ~7% per month; but Panther Lake's yield stayed around 10% until mid-year, far below the 70–80% needed for mass production.” · orig: “18A製程良率每月提升7%…Panther Lake的良率一直維持在10%左右，遠低於量產所需的70%至80%。”
自由財經 — 英特爾稱18A良率月升7% 將踏入2奈米賽道挑戰台積電· fetched 2026-06-02
[50]Tier 2neutralzhMedium confidence
Domestic trade press frames the 2nm race as TSMC-leading-but-contested: TSMC held 64.9% of the foundry market (advanced nodes = 74% of wafer sales), Samsung's SF2 showed ~30% initial yield, and Intel's foundry lost US$5.3bn in H1 2024 after abandoning its 20A node.
“TSMC's Q3 2024 foundry share was 64.9%, advanced nodes 74% of wafer sales; Samsung SF2 initial yield ~30%; Intel foundry lost US$5.3bn in H1 2024.” · orig: “台積電在2024年第三季晶圓代工市場佔有率達64.9%…三星2奈米製程(SF2)測試初始良率約30%…英特爾晶圓代工部門2024年上半年累計虧損53億美元。”
科技新報 — 先進製程2奈米之戰，台積電維持領先，三星、英特爾能否彎道超車？· fetched 2026-06-02

Strategy, Moats & Globalization · 7

[15]Tier 1neutralHigh confidence
On Mar 3, 2025 TSMC and President Trump announced an additional US$100bn US investment, lifting total planned Arizona investment to US$165bn (on top of the prior $65bn) — three more fabs, two advanced-packaging facilities and an R&D centre.
“TSMC intends to expand its U.S. investment by an additional US$100 billion, bringing total planned investment to US$165 billion.”
TSMC — Intends to Expand US Investment to US$165 Billion (press release)· fetched 2026-06-02
[16]Tier 2neutralMedium confidence
Under a 2026 US–Taiwan tariff framework, reports said TSMC's Arizona commitment could ultimately rise toward US$465bn; the CFO said the $165bn expansion 'isn't done.'
“TSMC's Arizona chip expansion isn't done after $165B U.S. investment, CFO signals.”
CNBC — TSMC set to expand its $165 billion U.S. investment· fetched 2026-06-02
[17]Tier 2supportingHigh confidence
TSMC's 2nm (N2) is its first gate-all-around (GAA) nanosheet node, claiming 10–15% speed gain or 25–30% lower power and ~15% higher density vs N3; volume production began quietly in Q4 2025 in Taiwan (Baoshan/Kaohsiung).
“N2 is designed to deliver a 10%–15% performance gain at the same power, 25%–30% lower power, and 15% higher density vs N3E.”
Tom's Hardware — TSMC begins volume production of 2nm-class chips· fetched 2026-06-02
[18]Tier 2supportingMedium confidence
CoWoS advanced-packaging capacity is set to expand from ~35,000 wafers/month in late 2024 to ~120,000–140,000 by end-2026; Nvidia reportedly secured over 60% of TSMC's 2026 packaging capacity for its Rubin GPUs.
“CoWoS capacity is expanding from 35,000 wafers per month in late 2024 to 120,000–140,000 by end-2026; NVIDIA secured over 60% of 2026 packaging capacity.”
FinancialContent — TSMC Commits $56B Capex to Double CoWoS Capacity· fetched 2026-06-02
[27]Tier 2criticalMedium confidence
Overseas fabs dilute margins: TSMC estimates near-term gross-margin dilution of ~2%, widening to 3–4% as overseas production scales — a structural cost of geographic diversification.
“TSMC estimates near-term margin dilution of around 2%, which could further expand to 3-4% as production scales.”
Nasdaq / Zacks — TSM's Global Expansion: Growth Enabler or Margin Risk?· fetched 2026-06-02
[35]Tier 2supportingzhMedium confidence
TSMC's 2nm initial capacity (Baoshan + Kaohsiung) was ~45,000–50,000 wafers/month by end-2025, rising above 100,000/month in 2026 and toward ~200,000 by 2027 incl. Arizona Fab 21 P2; Apple took ~half of initial N2 capacity.
“Baoshan and Kaohsiung combined ~45,000–50,000 wpm by end-2025, over 100,000 in 2026, ~200,000 by 2027 with Arizona.” · orig: “寶山與高雄合計2奈米月產能約4.5萬至5萬片，2026年上看逾10萬片…2028年可望達約20萬片。”
經濟日報 — 台積電2奈米產能規劃· fetched 2026-06-02
[41]Tier 2neutralMedium confidence
TSMC also builds abroad in Japan (JASM, Kumamoto) and Germany (ESMC, Dresden) alongside Arizona, pursuing a multi-region footprint backed by host-government subsidies and customer co-investment.
“TSMC's global expansion spans Arizona, Kumamoto (Japan) and Dresden (Germany).”
SemiWiki — TSMC 2025 Update: Riding the AI Wave Amid Global Expansion· fetched 2026-06-02

Peer Comparison · 3

[37]Tier 2supportingMedium confidence
TSMC's full-year 2025 gross margin of 59.9% far exceeds peers: Intel Foundry runs operating losses, Samsung Foundry is reportedly loss-making at the leading edge, while specialty/mature foundries (UMC, GlobalFoundries) earn lower structural margins without leading-edge nodes.
“Growth diverges as AI, specialty nodes and 2nm ramps reshape the market; Intel foundry remains in the red.”
AnySilicon — Global Foundry Earnings Q3 2025 (peer margins)· fetched 2026-06-02
[44]Tier 2supportingHigh confidence
On the leading edge TSMC competes mainly with Samsung Foundry and Intel Foundry; on mature/specialty nodes with SMIC, UMC and GlobalFoundries. Only TSMC, Samsung and Intel are pursuing sub-3nm, and TSMC leads on yield and customer base.
“TSMC maintains dominance through established volume production of advanced nodes, while Samsung and Intel compete on 2nm readiness with uncertain yields.”
TrendForce — 2nm battle peer set· fetched 2026-06-02
[47]Tier 2criticalzhMedium confidence
Per domestic reporting citing KeyBanc, Intel's 18A yield reached ~55% (above Samsung's ~40%) and is narrowing the gap to TSMC's ~65%; meanwhile a big customer 'releases orders with one hand and chases with the other' — Intel outsources some chips to TSMC while pushing its own node.
“Intel 18A yield has reached 55%, up from 50% last quarter, surpassing Samsung's 40% and approaching TSMC's 65% on 2nm.” · orig: “Intel 18A目前良率已達55％，較上季的50%出色…超越三星且快速逼近台積電2奈米水準。”
聯合新聞網 — 英特爾18A良率超車三星台積電大客戶左手釋單右手追趕· fetched 2026-06-02

Financials & Growth · 6

[2]Tier 2supportingHigh confidence
Q4 2025: revenue NT$1,046.09bn (US$33.73bn), net income NT$505.74bn, diluted EPS NT$19.50; gross margin 62.3%, operating margin 54.0%, net margin 48.3%. Node mix: 3nm 28%, 5nm 35%, 7nm 14%; ≤7nm advanced = 77%.
“Net profit NT$505.74 billion, diluted EPS NT$19.50; gross margin 62.3%, operating margin 54.0%; 3nm 28%, 5nm 35%, 7nm 14%, advanced (7nm and below) 77%.”
TrendForce — TSMC 2025 Q4 Net Profit Soars 35% YoY to Record NT$505.7B· fetched 2026-06-02
[3]Tier 2supportingHigh confidence
FY2025 revenue reached US$122.42bn; AI and HPC processors were 58% of 2025 revenue (a first). 2026 capital budget guided to US$52–56bn (~70% advanced logic, 10% specialty, 10–20% packaging). Results published Jan 15, 2026.
“AI and HPC processors accounted for 58% of TSMC's 2025 revenue of $122.42 billion... the company will invest $52-$56 billion in capex.”
Tom's Hardware — TSMC 'very nervous' about AI bubble; $52–$56B capex· fetched 2026-06-02
[11]Tier 1neutralHigh confidence
TSMC's official Q4/FY2025 results filing (SEC Form 6-K): full-year 2025 revenue US$122.4bn / NT$3.8tn (+35.9% USD), full-year gross margin 59.9%, EPS NT$66.25.
“Full year 2025 revenue increased 35.9% in US dollar terms to US$122 billion; full-year gross margin 59.9%; EPS NT$66.25.”
SEC Form 6-K — TSMC Q4/FY2025 results (primary filing)· fetched 2026-06-02
[12]Tier 3neutralHigh confidence
TSMC annual revenue history (US$): 2020 ~$45.5bn, 2021 ~$56.8bn, 2022 ~$75.9bn, 2023 ~$69.3bn, 2024 ~$90.1bn, 2025 ~$122.4bn — a 2023 dip then an AI-driven surge.
“2020 $45.51B; 2021 $56.82B; 2022 $75.88B; 2023 $69.30B; 2024 $90.08B.”
CompaniesMarketCap — TSMC revenue history· fetched 2026-06-02
[33]Tier 2neutralMedium confidence
TSMC's full-year 2025 capital expenditure was ~US$40.9bn (NT$1.3tn); the US$52–56bn 2026 guidance is a ~30% increase and a record.
“Full year 2025 capital expenditures were TWD 1.3 trillion or $40.9 billion; 2026 budget US$52–56 billion.”
Futu — TSMC Earnings Call: 2026 Capex Targeting High End of $52–56B· fetched 2026-06-02
[46]Tier 2criticalzhMedium confidence
Domestic reporting ties pressure on TSMC's profit-allocation (a reported ~15% cut to the employee-bonus ratio that angered some staff) directly to the surge in capex and overseas (US/Japan) fab buildout — a concrete sign the globalisation bill is being paid somewhere.
“Continued US and Japan fab expansion and an AI-driven capex surge mean the company must restructure how it allocates profit.” · orig: “美國、日本廠持續擴建，加上AI需求帶動資本支出暴增，公司勢必需要重新調整獲利分配結構。”
聯合新聞網 — 台積電爆『砍15%分紅』惹怒員工？· fetched 2026-06-02

Risks & Geopolitics · 7

[24]Tier 2neutralHigh confidence
A 7.4-magnitude earthquake struck Taiwan on Apr 3, 2024; TSMC estimated ~US$92m of Q2 losses but restored ~80% of operations within hours — illustrating both resilience and the concentration risk of Taiwan-centric production.
“TSMC estimated losses of $92.44 million... resumed 80% of operations within hours.”
Supplyframe — Taiwan Quake Spares TSMC But Rattles Supply-Chain Concerns· fetched 2026-06-02
[25]Tier 1criticalMedium confidence
TSMC officially lists earthquakes, floods, typhoons, droughts and shortages of water, electricity and gas among operational risks; it has estimated a serious drought could cut revenue ~0.7–1.1%.
“TSMC acknowledges threats including earthquakes, floods, typhoons, droughts, and shortages in utilities such as water, electricity and natural gas.”
TSMC — Risk Management (investor relations)· fetched 2026-06-02
[26]Tier 2criticalzhMedium confidence
Per analysts, water and power are 'Taiwan's semiconductor Achilles heel'; ex-chairman Mark Liu acknowledged Taiwan is small and resource-scarce, and TSMC's chairmen have repeatedly named talent shortage as the industry's biggest challenge.
“Shortages of labour, water and electricity are concerns for TSMC's expansion; BCG warned water and power are the Achilles heel of Taiwan's semiconductors.” · orig: “缺工、缺水、缺電成隱憂…缺水、缺電是台灣半導體的阿基里斯腱。”
自由財經 — 日經：缺工、缺水、缺電成台積電擴張隱憂· fetched 2026-06-02
[28]Tier 2criticalzhMedium confidence
Taiwan's 'silicon shield' logic — that the world must protect Taiwan because its chip output is indispensable — is debated as TSMC builds abroad; analysts at STPI/ITRI frame a shift toward a 'Taiwan–US dual-core' model and warn the shield erodes if leading-edge output abroad reaches ~30–50% of US demand.
“Traditional silicon-shield logic rests on Taiwan's indispensable, singular capacity... TSMC's US expansion signals a shift from a Taiwan-centric to a Taiwan–US dual-core architecture.” · orig: “過去傳統矽盾的邏輯建立在台灣產能的不可或缺性與單一性上…台積電若再度對美擴大投資，則象徵公司從台灣單一核心轉型為台美雙核心架構。”
STPI iKnow — 從護國神山到雙核心韌性：台積電美廠擴張對矽盾的考驗· fetched 2026-06-02
[29]Tier 2supportingzhHigh confidence
Per Taiwan's Ministry of Economic Affairs (advanced = sub-5nm), ~94% of advanced-process capacity was in Taiwan in 2025, projected to fall to ~79% by 2026 (~21% overseas); Taiwan amended its National Security Act to protect sub-14nm tech, and TSMC stresses R&D stays in Taiwan ('Made with Taiwan').
“~94% of sub-5nm advanced processes were in Taiwan in 2025, with an estimated 79% remaining by 2026... strategy evolved from 'Made in Taiwan' to 'Made with Taiwan'.” · orig: “2025年約94%的5奈米以下先進製程在台灣，到2026年預計將剩下79%留在台灣…戰略已從『Made in Taiwan』演進為『Made with Taiwan』。”
TVBS — 矽盾2.0來了！台積電海外設廠專家揭留台關鍵防去台化· fetched 2026-06-02
[30]Tier 2supportingzhMedium confidence
Counter-view: a US academic (Matt Seitz) argues reliable chip capacity outside Taiwan is at least a 2030–2035 prospect and full resilience may take until mid-century; a Taiwan think-tank head (Chang Chih-cheng) calls Taiwan's ecosystem 'virtually irreplaceable' and US-made leading-edge 'extremely costly and inefficient' — so de-Taiwanization fears may be overstated near-term.
“Reliable capacity outside Taiwan needs at least 2030 or 2035; true resilience may require waiting until mid-century or later.” · orig: “要達到『在台灣之外建立可靠製造能力』的階段，至少需要2030或2035年…恐怕要等到本世紀中葉、甚至更晚。”
風傳媒 — 台積電赴美投資美國學者：傷矽盾是2050年以後的事· fetched 2026-06-02
[43]Tier 2criticalzhMedium confidence
Talent is a recurring constraint: TSMC's headcount grew from ~51k (2019) to ~76k (2023) toward 80k+ in 2024, yet — per domestic trade reporting — TSMC and MediaTek 'still cry talent shortage', and the deep work-life-balance gap in the US and Germany makes its Taiwan work culture hard to replicate abroad.
“Taiwan's engineering talent mostly flows to high-paying leaders like TSMC and MediaTek, yet these two firms still cry talent shortage.” · orig: “台灣理工人才大多流向台積電、聯發科等祭出高薪的龍頭公司，但這2家公司還是大喊缺人。”
DIGITIMES 中文 — 台積電員工突破8萬人海外招聘work-life balance難解· fetched 2026-06-02

Forward View · 4

[4]Tier 2criticalHigh confidence
Asked about an AI bubble, CEO C.C. Wei said TSMC was 'very nervous' and that careless investment 'would be a disaster for TSMC for sure' — while still committing record 2026 capex, signalling confidence in multi-year AI demand.
“If we did not do it carefully, that would be a disaster for TSMC for sure.”
Tom's Hardware — TSMC CEO on AI-bubble risk· fetched 2026-06-02
[31]Tier 2criticalMedium confidence
Bear case: analysts warn TSM's valuation reflects peak margins and utilisation; record fixed costs from $52–56bn capex make profit and free cash flow highly sensitive to any AI-demand normalisation, and revenue is increasingly concentrated in AI/hyperscaler customers.
“Surging capex and rising fixed costs heighten downside risk if demand normalizes... revenue increasingly concentrated in AI platforms and hyperscalers.”
I/O Fund — TSM Stock and the AI Bubble: the valuation debate· fetched 2026-06-02
[32]Tier 2supportingMedium confidence
Bull framing: TSMC management says aggressive expansion is still insufficient to meet strong AI demand; 2026 capex targets the high end of the US$52–56bn range, with ~3-year fab lead times implying confidence in demand through ~2029.
“TSMC's leadership maintains strong conviction in the multiyear AI megatrend... capacity additions aligned with customer roadmaps.”
MLQ.ai — TSMC signals heavier AI-driven spending, lifts 2026 capex outlook· fetched 2026-06-02
[38]Tier 3supportingMedium confidence
Some analysts argue a slowdown in AI build-out would hit TSMC less than feared because its leading-edge monopoly and diversified end-markets (smartphone, auto, IoT) cushion a downturn — though others see sharp operating-leverage downside.
“Analysis of TSMC's resilience if AI infrastructure spending slows.”
Motley Fool — What Happens to TSMC Stock if the AI Build-Out Slows?· fetched 2026-06-02