Taiwan Semiconductor Manufacturing Company Limited (NYSE: TSM) — The Widest Moat in Technology on Its Thinnest-Ever Valuation Cushion

Ticker: NYSE: TSM (ADR; ordinary shares TWSE: 2330) · Sector: Information Technology — Semiconductors (dedicated foundry) Date: June 6, 2026 Primary source: FY2025 Form 20-F (filed 2026-04-16, fiscal year ended 2025-12-31); IFRS; reported in New Taiwan Dollars (NT$) with USD convenience translation at NT$31.37/US$1.00. 1 ADS = 5 ordinary shares.

⚡ Claude’s Take

This block is the author’s own independent, subjective opinion. It is general information only and is not investment advice. The analysis that follows (Sections 1–15) is deliberately position-free and carries no recommendation or price target — that discipline is intact everywhere below this block.

Verdict: HOLD / great business, full price — accumulate on weakness, do not chase here. Not a short. TSMC is, on the evidence, one of the widest-moat businesses in global equities — a leading-edge foundry near-monopoly with ~70% of all foundry revenue, >90% of the leading edge, 59.9% gross margins that are expanding with scale, ~35% ROE, a US$63B net-cash fortress, and essentially zero dilution. The problem is not the business; it is the entry point. The stock trades at the 99.5th percentile of its own ten-year valuation history (~40x trailing earnings, ~21x forward, ~12.5x book), having tripled off its 52-week low of US$202 to ~US$415. My reverse-DCF says the market is paying for management’s full base case — ~20%+ revenue compounding plus capex normalization — with no multiple-based margin of safety. My four-year scenario work frames an asymmetric payoff: roughly +3%/yr in the base case, +16%/yr if the AI super-cycle and pricing power persist, but −13%/yr if an AI-capex air-pocket or a Taiwan geopolitical de-rating compresses the multiple back toward its historical mean.

The framing is quality-compounder-at-a-full-price, not deep value and not momentum-chase. I would own this business enthusiastically and the stock patiently: accumulate into weakness below roughly US$350 (~18x forward EPS) and back up the truck below roughly US$280 (~14x forward, toward where a de-rating + the demonstrated franchise economics offer real support). At ~US$415 the right action is to hold what you have and wait — you are being asked to underwrite flawless execution at the richest price the franchise has ever commanded.

Conviction: medium. Flips bullish on a 20%+ de-rating with the thesis intact, or hard evidence that capex intensity is normalizing into an FCF inflection while AI demand holds. Flips bearish if Intel 18A/14A or Samsung SF2 demonstrably close the yield gap and win marquee leading-edge customers, if hyperscaler AI-capex visibly rolls over (CoWoS cancellations, cloud-capex cuts), or on a Taiwan-Strait escalation. Tag: “The widest moat in technology, standing on the thinnest valuation cushion it has ever had.”

1. Executive Summary

Taiwan Semiconductor Manufacturing (TSMC) is the world’s dominant dedicated semiconductor foundry — it manufactures leading-edge logic chips to the designs of fabless customers (Apple, Nvidia, AMD, Qualcomm, Broadcom) and owns none of its own chip products. In FY2025 it generated US$121.4B of revenue (+31.6%), a 59.9% gross margin, a 50.8% operating margin, ~US$54B of net income (44.5% net margin), and ~35% ROE — figures that sit at the very top of the global manufacturing universe and, critically, are improving with scale. The clearest evidence of the moat is the incremental margin: on the ~US$29B of revenue added in 2025, roughly two-thirds fell through to operating income.

TSMC’s advantage is the rare case where the moat is the 60% gross margin. It rests on economies of scale fused with customer captivity (Greenwald’s most durable category): a leading-edge fab costs US$20–30B and TSMC’s FY2025 capex alone was US$40.6B (~34% of revenue), a fixed-cost base amortized over a wafer volume no rival can match. That scale funds a process-technology lead (N2/2nm entered volume production in 2025; yields reportedly 10–15 points ahead of Intel 18A and Samsung SF2), which funds an ASP premium (N2 wafers ~US$30,000, “no discounts”), which funds the next node’s scale. Reinforcing layers — multi-million-dollar customer switching costs, the proprietary CoWoS/SoIC advanced-packaging chokehold on AI accelerators, and the co-developed EDA/IP “grand alliance” — widen the position further. At the leading edge TSMC holds >90% share, and that share has been stable-to-rising for a decade while Samsung loses ground and Intel exited merchant leading-edge logic.

The industry structure is excellent at the leading edge (a near-monopoly with escalating capital- and lithography-gated barriers, secular AI demand, and a supply ceiling set by ASML’s EUV monopoly) and poor at the trailing edge (Chinese state-subsidized mature-node overcapacity) — but mature nodes are a minority ~26% of TSMC’s mix, much of it differentiated specialty. The dominant risks are therefore not competitive erosion (TSMC is winning) but (1) the Taiwan-China geopolitical tail — the overwhelming majority of leading-edge capacity sits on one island within missile range of the PRC — and (2) a cyclical AI-capex air-pocket that would temporarily de-rate utilization.

Capital allocation is excellent: TSMC reinvests the bulk of a 35%-ROE earnings stream at incremental returns far above its cost of capital, self-funds the largest capex program in industry history out of operating cash flow, pays a growing-but-conservative dividend (NT$22/share FY2025, +29%; ~28% payout), does essentially no buybacks-at-peak, no dilution, and no empire-building M&A. The debated item — building higher-cost fabs in Arizona, Japan and Germany (a 2–4% gross-margin drag) — is best read as a subsidized, customer-funded supply-security investment, validated by the first Arizona fab turning profitable in 2025.

The entire investment question reduces to price. TSMC trades at ~US$415/ADS (US$2.15T market cap, ~US$2.09T EV on net cash), the richest valuation in its own recorded history, yet the lowest forward P/E in the leading-edge complex (~21x vs ASML 34x). Both are true: the franchise has structurally re-rated from cyclical commodity foundry to secular-growth compounder, and the “cheap” forward multiple is entirely conditional on the embedded EPS leap (US$10.44 → ~US$19.3/ADS) actually delivering. There is no valuation margin of safety. This note takes no position on the security; it lays out the embedded expectations and the evidence on both sides.

2. Business Overview

The model. TSMC is the world’s largest pure-play (dedicated) semiconductor foundry — it fabricates logic chips to its customers’ designs but designs and sells no chips of its own. This is the structurally decisive feature. An Integrated Device Manufacturer (IDM) such as Intel or Samsung both designs and manufactures chips, and therefore competes with the very fabless companies it would court as foundry customers. TSMC carries no such channel conflict: Apple, Nvidia, AMD, Qualcomm and Broadcom can hand it their most valuable IP without fear the foundry will build a competing product. Morris Chang’s founding insight — separating design from manufacturing — is, in Greenwald’s terms, the reason the foundry market exists at all and the reason a pure-play, not an IDM, dominates it.

How it makes money. TSMC sells wafer-fabrication capacity at leading process nodes, priced by node, design complexity, and increasingly by advanced-packaging content. Revenue scales with (a) wafer volume and (b) average selling price per wafer, the latter rising sharply as mix shifts to advanced nodes. FY2025 revenue of US$121.4B converted to a 59.9% gross margin, a 50.8% operating margin, a 44.5% net margin (~US$54B), and ~35% ROE. The pricing mechanics matter to the thesis: a leading-edge 300mm wafer at N3 sells for roughly US$18,000–20,000 and N2 for ~US$30,000, versus a few thousand dollars at mature nodes — so each node migration is simultaneously a volume event and a steep ASP step-up, and the gross-margin ladder (51.6% → 59.6% → 54.4% → 56.1% → 59.9% across 2021–25) tracks utilization and advanced-node mix more than any cost-cutting program. Unlike a fabless customer, TSMC also collects capacity-reservation deposits and, increasingly, customer prepayments that de-risk the capex build — a structural working-capital advantage that funds the fab arms race partly on customers’ balance sheets.

The customer base is broadening within concentration. The apparent paradox of FY2025 — top-10 concentration rising to 78% even as the largest customer (Apple) fell to 19% — is explained by the rise of a new cohort: alongside Nvidia (17%) and AMD, the hyperscalers are now designing their own AI silicon and fabbing it at TSMC — Google (TPU), Amazon (Trainium/Inferentia), Microsoft (Maia), Meta (MTIA) and others. This “custom-ASIC” wave deepens TSMC’s captivity (every credible AI-silicon program in the West, merchant or in-house, routes through TSMC’s leading edge and CoWoS packaging) while diversifying it away from any single end-product cycle. It also means TSMC’s demand is increasingly tied to aggregate hyperscaler AI capex rather than to the smartphone replacement cycle that historically drove it — a higher-growth but also a more capex-cycle-sensitive demand base.

Revenue by platform (FY2025, % of net revenue). The single most important business fact of the last two years is a mix inversion: High-Performance Computing (HPC) is now 58% of revenue (up from 43% in 2023 and 51% in 2024), having overtaken Smartphone at 29% (down from 38%/35%). IoT is 5%, Automotive 5%, Digital Consumer Electronics 1%, Others 2%. AI accelerators — Nvidia, AMD and hyperscaler custom silicon — are the engine, and the growth is concentrated in the most advanced, highest-ASP nodes.

Revenue by technology (% of wafer revenue, FY2025). 3nm reached 24% (just 6% in 2023, 18% in 2024); 5nm 36%; 7nm 14% — so advanced technology (7nm-and-below) was 74% of wafer revenue, up from 69% (2024) and 58% (2023). Mature nodes (16nm and above) are a stable ~26% tail. N2 (2nm), TSMC’s first gate-all-around node, entered volume production in 2025; A16 (16-angstrom) risk production is slated for 2026.

Revenue by geography (FY2025). North America 75% (up from 70%), Asia-Pacific ex-China/Japan 9%, China 9% (down from 12%), Japan 4%, EMEA 3%. The North-America surge reflects US fabless AI customers; China’s decline reflects export controls and the AI mix shift.

Customer concentration. Top-10 customers were 78% of FY2025 revenue (70% in 2023). The largest customer (Apple) was 19% (down from 25% in 2023); the second (Nvidia) 17% (up from 11%) — the AI surge in one number. Top-2 ≈ 36%.

Recurring vs. cyclical. At the franchise level the business is highly recurring — essentially every advanced logic chip designed in the Western ecosystem is taped out at TSMC, and designs lock in for multi-year product cycles. At the volume level it remains cyclical (smartphone/PC end-demand, inventory corrections); the FY2023 −4.5% revenue dip is the reminder. The current AI/HPC boom is masking that cyclicality — a fact that matters more for valuation than for business quality.

Verdict (Business Overview): A genuinely differentiated, structurally recurring franchise with a decisively favorable mix shift toward its highest-value nodes — but with extreme customer concentration and residual end-market cyclicality currently hidden by the AI cycle.

3. Industry Dynamics

The value chain and where the profit pools sit. The logic-semiconductor chain runs in five linked stages, and the location of economic rent explains TSMC’s privilege. At the top sit EDA/IP licensors — Synopsys and Cadence (a design-software duopoly >70% share) and Arm (the dominant instruction-set architecture). Next, fabless designers (Nvidia, AMD, Apple, Qualcomm, Broadcom, MediaTek) capture brand and architectural value but own no fabs. They hand designs to foundries (TSMC, Samsung, Intel, SMIC, UMC, GlobalFoundries), which convert designs into silicon. Wafers pass to OSAT / advanced-packaging (ASE, Amkor — and, critically, TSMC’s own in-house CoWoS). Underpinning all of it are the equipment makers: ASML (lithography), Applied Materials, Lam, Tokyo Electron, KLA.

Profit concentrates at two “tollbooths.” The first is ASML, sole supplier of EUV and High-NA EUV — no chip below ~7nm exists without its machines, which gives ASML monopoly pricing and gates total industry leading-edge supply. The second is TSMC, the dominant leading-edge foundry. Everyone else is, to varying degrees, captive: fabless designers cannot manufacture, and cannot cheaply switch foundries because re-taping-out a chip to a different process is a multi-quarter, multi-million-dollar engineering effort with yield risk. TSMC’s 59.9% gross and 50.8% operating margins are the financial fingerprint of that position.

Market size and growth. The tailwind is exceptional. The global semiconductor market reached ~US$772B in 2025 (+22%) and WSTS forecasts ~US$975B in 2026 (+25%), approaching US$1 trillion, with Logic guided +37% — almost entirely AI-driven. Within that, the top-10 foundries earned ~US$169.5B in 2025 (+26.3%); TSMC’s US$121.4B is ~72% of all top-10 foundry revenue. The growth is concentrated at the leading edge (advanced nodes = 74% of TSMC wafer revenue). Management has raised its long-term target to a ~25% USD revenue CAGR for 2024–2029 and guides AI-accelerator revenue to a mid-to-high-50s% CAGR — aggressive targets that embed substantial AI optimism, but underwritten by a known node pipeline (N2 ramping through 2026, A16 to follow).

Competitive intensity — a near-monopoly at the leading edge. Blended 2025 foundry share: TSMC ~69.9% (a record), Samsung 7.2%, SMIC 5.3%, UMC 4.35%, GlobalFoundries 3.87% — with Intel Foundry not in the top ten. The blended figure understates the dominance: at the leading edge (≤5nm), TSMC’s share is estimated above 90%. Apply Greenwald’s market-share-stability test — a swing under ~2 points over 5–8 years signals formidable barriers — and TSMC’s share has not merely held but risen from the mid-50s% to ~70%. That is a dominant-firm regime, and the barriers compound: (1) capex intensity — a single leading-edge fab is US$20–30B+, with 2026 industry capex near US$200B and TSMC alone at US$52–56B; (2) EUV/High-NA access rationed by ASML (~US$350M per High-NA machine); (3) the cumulative-volume yield learning curve; (4) process talent; (5) ecosystem lock-in. These are escalating, not eroding.

Capital cycle — the two-speed split. Apply Marathon’s supply-side lens and the industry bifurcates. At the leading edge the normal capital cycle is partly suspended: record capital is chasing high returns, but mean-reversion is blunted because only one rational actor (TSMC) can profitably build leading-edge capacity, ASML throughput caps total supply, and AI demand is lumpy-to-the-upside. This is Marathon’s “value in growth” case — a wide moat plus a concentrated, favorable supply side justifies premium economics that would mean-revert in a fragmented industry. The chief capital-cycle risk sits one level up: hyperscaler AI-datacenter capex is itself a textbook boom, and an overbuilt cloud-capex cycle would hit TSMC’s leading-edge utilization hard.

At the trailing/mature edge the picture inverts into a classic negative capital cycle. Chinese foundries — SMIC, Hua Hong, Nexchip — are scaling state-subsidized mature-node capacity aggressively (China’s share of top-10 mature capacity set to exceed 25% by end-2025) and cutting prices. Because this is state capitalism — subsidies, debt forgiveness, no market-clearing — the cycle that would normally punish overcapacity is broken. TSMC is largely insulated (only ~26% of revenue is mature, much of it differentiated specialty with selective 2025 price hikes); the exposure is its overseas specialty fabs (JASM, ESMC) pushing it into the contested mature arena at the wrong point in that sub-cycle.

Regulation and geopolitics — the dominant structural factor. The single largest variable is not competitive but geopolitical: Taiwan-China risk. The overwhelming majority of TSMC’s leading-edge capacity sits on one island within PRC missile range. The “silicon shield” cuts both ways — concentration is simultaneously the chief vulnerability and a deterrent. US export controls (the Dec-2024 Entity-List expansion, the Jan-2025 AI Diffusion Framework and Foundry Due-Diligence Rule; Taiwan adding Huawei and SMIC to its own control list in June 2025) cap China’s leading-edge access and on balance advantage TSMC against SMIC — but carry compliance tail-risk (a potential >US$1B Commerce penalty over chips allegedly rerouted to Huawei via Sophgo). Forced geographic diversification is the cost of customer/government “de-risking”: Arizona (a ~US$165B program; the first fab swung to a NT$16.1B profit in 2025 from a NT$14.3B 2024 loss), Japan/JASM (losses widening), Germany/ESMC, with ~US$4.7B of subsidies secured. The structural cost is a ~2–3% gross-margin dilution in FY2025, guided to a further 2–3% in FY2026.

Verdict (Industry): Structurally excellent at the leading edge (a near-monopoly with escalating, capital- and lithography-gated barriers, secular AI demand, and a supply-constrained ceiling), where TSMC sits almost uncontested across ~74% of revenue; structurally poor and deteriorating at the trailing edge (Chinese state-subsidized overcapacity), where TSMC’s exposure is a minority and largely specialty. The decisive risks are geopolitical and cyclical, not competitive.

4. Competitive Position

Relevant market (Greenwald Step 1). Not “semiconductors” broadly but leading-edge (7nm-and-below) logic foundry — the arena where fixed costs stay fixed and scale compounds. Defining the market correctly is essential: TSMC’s dominance is overwhelming here and merely solid at mature nodes.

The moat, named: Economies of Scale + Customer Captivity (Greenwald’s strongest, most durable category), reinforced by a genuine supply/process advantage and an ecosystem network effect. Each mechanism, pressure-tested against the financial outcome it protects:

1. Economies of scale + capex intensity (the load-bearing layer). A leading-edge fab costs ~US$20–30B; TSMC’s FY2025 capex alone was US$40.6B (~34% of revenue) across >17M wafers of annual capacity. The fixed-cost base of R&D, EUV tooling, and fab construction is amortized over a volume no competitor can match, so TSMC’s per-wafer leading-edge cost is structurally lowest. Falsification test: remove the scale and TSMC cannot fund the R&D/capex arms race — this is the moat that, if broken, breaks the 59.9% gross margin and ~35% ROE. It is the dominant advantage.

2. Process-technology / yield leadership (genuine but secondary). N2 yields are reported at ~65–70% in early 2026, a 10–15-point lead over Intel 18A and ~10–20 points over Samsung SF2. Greenwald cautions that proprietary-technology advantages erode (“in the long run everything is a toaster”) — true, but here the technology lead is fused to scale: only scale funds the next node, and the node lead funds the ASP premium that funds scale. Falsification test: lose the process lead and the ASP premium erodes, compressing margins — but scale and packaging would still hold. Reinforcing, not load-bearing.

3. Customer switching costs (real, financially significant). Moving a chip to another foundry means re-doing the physical design against a different PDK, re-qualifying IP, paying for new advanced mask sets (tens of millions per tape-out), and absorbing yield-ramp risk. The evidence is behavioral: despite a US$16.5B Samsung deal, Tesla routed its top “AI 6.5” chip to TSMC’s N2, and Qualcomm is reportedly reverting to TSMC for 2nm given Samsung’s yields. Customers dual-source for leverage and keep coming back at the leading edge.

4. Advanced-packaging lock-in (CoWoS / SoIC / 3DFabric) — a newer, second moat layer. AI accelerators are now gated not by the transistor alone but by TSMC-proprietary advanced packaging. Nvidia has reportedly booked ~60% of CoWoS demand (~515,000 wafers) for 2026–27, and TSMC is scaling CoWoS from ~35k wafers/month (late-2024) toward ~130k/month (end-2026). Packaging scarcity is itself a barrier — no rival offers an at-scale CoWoS-L equivalent. Falsification test: this directly gates Nvidia/AMD AI revenue; remove it and the highest-margin AI volume becomes contestable. It is becoming load-bearing.

5. The “grand alliance” / Open Innovation Platform (ecosystem network effect). TSMC’s PDKs, IP libraries and reference flows are co-developed with Synopsys, Cadence and IP vendors so that designing for TSMC is the path of least resistance. More customers → more IP ported → easier for the next customer. A genuine (if hard-to-quantify) network effect that lowers customer design cost and raises rivals’ switching cost.

Pricing power as proof. N2 wafers are reportedly priced ~US$30,000 with a stated “no discounts” policy — a 10–20% (some reports higher) premium over N3. A firm that can raise leading-edge prices into a supply-constrained market and watch gross margin climb toward 60% is exercising the pricing power only a real moat confers.

Competitive dynamics vs. rivals.

• Intel Foundry (IFS): 18A entered volume in 2025 but yields lag (~50–55% mid-2025) and external traction is negligible — IFS external revenue was ~US$174M in Q1’26 (~3% of its foundry revenue) versus TSMC’s US$35.9B. Microsoft (Maia), Amazon and a preliminary Apple deal (May 2026) are votes for supply diversification, but magnitude is unproven. Intel is the most credible 18A/14A threat if yields converge by 2027.
• Samsung Foundry: SF2/2nm yields reported mid-50s% (below the ~60% mass-production threshold) versus TSMC 60–70%; Taylor, TX customer production slips to 2027. The Tesla win is real, but the marquee high-end chip went to TSMC.
• SMIC (trailing edge / China): ~5.3% share, #3, but capped near 7nm by EUV export controls — a mature-node and geopolitical concern, not a leading-edge threat.

The longer-dated threats — China indigenization and High-NA EUV. Two developments bear watching beyond the 2027 yield-gap question. First, China’s indigenization drive: blocked from EUV and the leading edge by export controls, Beijing is pouring state capital into a domestic supply chain (SMIC at 7nm via multi-patterned DUV, domestic lithography efforts, Huawei’s Ascend accelerators). This is a mature-node and self-sufficiency threat, not a near-term leading-edge one — SMIC cannot economically reach 3nm without EUV — but over a 5–10-year horizon a state-funded competitor unconstrained by return-on-capital discipline is the kind of supply-side force Marathon warns about, and it is the reason the mature-node sub-segment is structurally unattractive. Second, High-NA EUV: ASML’s next-generation lithography (~US$350M per tool) is the gating technology for the angstrom era (A14 and below). TSMC has been characteristically deliberate — extending current EUV via multi-patterning rather than rushing High-NA — while Intel has bought in early. If High-NA proves decisive and Intel’s early adoption translates into a process lead at A14, that is the one technical path by which the yield gap could invert. Today the evidence runs the other way (Intel 18A yields lag, external traction is negligible), but it is the threat a TSMC bull should monitor most closely.

Greenwald scorecard. Market-share stability: passes decisively (share stable-to-rising for a decade). ROIC/profitability: passes (≈35% ROE, 50.8% operating margin, margins expanding with scale — the unmistakable fingerprint of scale + captivity). Cost-of-entry vs. value: the replacement cost of TSMC’s leading-edge capacity, accumulated process knowledge, and customer-qualified ecosystem is effectively un-buildable by a new entrant at any rational cost — the EPV of the franchise sits far above its reproduction cost, the Greenwald signature of a true barrier-protected business rather than a merely well-run one.

Verdict (Competitive Position): A durable, multi-layered, and currently widening advantage — one of the widest, most identifiable moats in global equities. The rare case where the moat is the 60% gross margin. Two honest qualifiers: customer concentration is extreme (top-2 ≈ 36%, though better read as evidence of captivity than fragility); and the Marathon caution that capex at ~34% of revenue with CoWoS capacity up ~4x in two years is boom-phase deployment whose favorable supply side persists only because Intel and Samsung keep failing to convert capex into competitive yield. The swing variables for durability are (a) whether Intel 18A/14A or Samsung SF2P close the yield gap by 2027, and (b) Taiwan geographic concentration.

5. Growth History and Forward Opportunities

History. Revenue compounded at ~24.5% over five years (NT$1,587B in 2021 → NT$3,809B in 2025), but not in a straight line: +42.6% (2022), −4.5% (2023, the PC/smartphone inventory trough coinciding with the N3 ramp), +33.9% (2024), +31.6% (2025). The composition shifted decisively — HPC from 43% to 58% of revenue in two years, +48% y/y in 2025 alone — making this an AI-led growth story layered on a maturing smartphone base. Growth has been almost entirely organic, funded by reinvested cash flow rather than acquisition or equity issuance.

Quality of the growth. High. It is concentrated in leading-edge nodes at premium ASPs (3nm went 6% → 24% of wafer revenue in two years), it expands margins rather than diluting them (incremental operating margin ~67% in 2025), and it is anchored by multi-year customer design wins and capacity-reservation agreements rather than spot demand. The principal quality caveat is concentration: the growth depends heavily on a handful of US fabless AI customers (Nvidia 17%, plus AMD, Apple, hyperscaler ASICs), so the “quality” is contingent on the AI-accelerator cycle holding.

Forward opportunities.

• AI accelerators / HPC: the core driver. Management guides AI-accelerator revenue to a mid-to-high-50s% CAGR through 2029 and overall revenue to ~25% USD CAGR — implying revenue roughly doubling to ~US$235–250B by FY2029 if the base case holds.
• Node roadmap: N2 (2nm, gate-all-around) ramping through 2026 at premium pricing; A16 (16Å, backside power delivery) to follow. Each node migration is a fresh ASP step-up and a fresh switching-cost reset in TSMC’s favor.
• Advanced packaging (CoWoS/SoIC): a second growth vector with its own scarcity-driven pricing, expanding ~4x by end-2026, increasingly indispensable to AI accelerators.
• Geographic expansion: Arizona (three new fabs + two packaging facilities + an R&D center under the US$165B program), Japan (JASM second fab), Germany (ESMC/Dresden) — adds capacity and supply-security optionality, at a near-term margin cost.

Decomposing the growth. It is worth separating the drivers, because they carry different durability. The FY2025 +31.6% breaks down roughly into HPC (+48% y/y, the dominant contributor — call it ~24 points of the total), smartphone (+11%, ~3 points), and the smaller platforms. In other words, the great majority of TSMC’s recent growth is a single vector: AI/HPC. Strip out HPC and the residual business is growing at a healthy-but-pedestrian high-single-digit rate consistent with its mature-smartphone and specialty exposure. This is the double-edged core of the growth story — the AI vector is what makes TSMC a 30%-grower rather than a 10%-grower, but it also means a HPC air-pocket would not merely slow growth, it would remove most of it. The bull reads the AI vector as a multi-year secular build (data-center buildout, inference scaling, sovereign AI, edge AI); the bear reads it as a capex spike that digests. The node and packaging roadmaps are real, contracted, and visible; the demand underneath them is the assumption.

The mature-node and packaging vectors. Two secondary growth sources deserve mention. Mature/specialty nodes (~26% of revenue — RF, BCD, automotive, image sensors) grow slowly but carry selective pricing power and anchor the overseas specialty fabs (JASM, ESMC); they are a stabilizer, not an engine, and face the Chinese-overcapacity pressure discussed above. Advanced packaging (CoWoS/SoIC), by contrast, is becoming a genuine second growth engine with its own scarcity economics — as AI accelerators integrate more HBM stacks and larger interposers, packaging revenue per accelerator rises, and TSMC captures it at attractive margins because it controls the at-scale capacity. Over time, “wafer plus packaging” content per leading-edge customer is the metric to watch; it is rising, and it deepens the moat as it grows.

Verdict (Growth): High-quality growth — margin-accretive, leading-edge, organic, and durable at the franchise level — but increasingly concentrated in the AI-accelerator cycle. The forward opportunity set is genuine and large; the risk is that management’s ~25% CAGR target embeds an AI-demand persistence that is itself the bull-case assumption (and the bear-case vulnerability). The honest read: the capability to grow (capacity, nodes, packaging, customer captivity) is proven and contracted; the rate of growth is hostage to the one variable TSMC does not control — how much AI compute the world buys.

6. Financial Quality

All figures IFRS, NT$ millions unless noted; FY2025 USD at NT$31.37/US$1.00. Source: FY2025 20-F, audited by Deloitte & Touche (auditor since 1987); prior years cross-checked to public filings.

Operating leverage is the headline. FY2025 gross margin recovered to 59.9% (from the 54.4% 2023 trough) and operating margin hit a record 50.8%. More telling than the levels: on the NT$914.7B of revenue added in 2025, incremental gross margin was ~71.8% and incremental operating margin ~67.1% — roughly two-thirds of every incremental revenue dollar fell through to operating income. That is textbook operating leverage from a fixed-cost-heavy, high-utilization fab base, and it occurred despite a ~2–3-point gross-margin drag from the ramp of overseas fabs. The core Taiwan operation runs at margins high enough to absorb the geographic-diversification tax and still expand consolidated gross margin.

Free cash flow and the capex phase. FCF was US$32.0B in FY2025 (26.3% FCF margin) — after funding US$40.6B of capex. OCF was US$72.5B, and cash conversion (OCF/net income) was 1.34x, the natural result of depreciation running well above working-capital and tax drags. Earnings are firmly cash-backed. The capex line is the central judgment call: FY2025 capex was 33.4% of revenue, and capex/depreciation was 1.87x — still firmly in heavy-build mode (depreciation lags spend ~3–5 years). The decline in intensity (53% → 33% of revenue) even as the absolute number rose is what lifted FCF margin from the mid-teens to the high-20s. The forward risk is explicit: management guides FY2026 capex to US$52–56B, a ~30% step-up that will compress near-term FCF and is the key swing factor for any forward FCF model.

Balance sheet — a deeply net-cash fortress. Cash & equivalents were NT$2,767,856M (US$88.2B); with financial assets and FVOCI instruments, gross liquid resources are ~US$101B. Total debt (bonds + bank loans + current portion) is ~US$32.9B, leaving net cash ≈ US$63B. TSMC carries modest investment-grade debt purely as a cost-of-capital optimization. PP&E is 47% of total assets — the physical fab base and the future depreciation load. Two working-capital tells reinforce earnings quality: inventories were essentially flat (NT$288.1B vs NT$287.9B) despite +31.6% revenue (the opposite of channel-stuffing), and receivables grew only ~3% against +31.6% revenue (pristine collections, no revenue pulled forward via loose terms). The material off-balance-sheet item is the parent guarantee of TSMC Arizona’s CHIPS-related obligations (up to US$6.6B funding + US$5B loans).

Quality of earnings — high and conservative.

• Net income tracks OCF (1.34x); AR and inventory grow far below revenue; SBC is trivial (NT$1,246M, 0.03% of revenue).
• Dilution is effectively zero. Weighted-average shares (~25,928M) were flat 2023–2025; the dilutive effect of potential shares was 0.009%. TSMC funds growth from cash flow and debt, not equity — a sharp contrast to most “growth” semis.
• Non-operating income is recurring, not gimmickry: dominated by interest income of NT$105,739M (US$3,371M) on the cash pile — genuine and repeatable. The FX gain (NT$13,831M) is volatile and should be normalized out of run-rate, but at ~0.7% of pretax it is immaterial.
• Effective tax rate ~17% (13.1% / 17.7% / 17.0% over 2023–25) — below the 20% ROC statutory rate, reflecting a ~5% Taiwan alternative tax on unappropriated earnings offset by R&D/investment credits and prior-year true-ups. (Note: income-tax expense is 9.1% of revenue in the common-size statement, which equals ~17% of pretax income — the two figures are consistent, on different denominators.)
• No one-time items distort FY2025 run-rate (only minor earthquake losses, ~NT$5.3B in 1Q25, booked in COGS).

The depreciation wall — the one mechanical headwind to watch. The capex super-cycle plants a future cost. Today’s US$40–56B of annual spend converts, with a 3–5-year lag, into depreciation that flows through cost of revenue. Depreciation already rose from NT$523B (2023) to NT$680B (2025) and will keep climbing as the AI-era fabs season. As long as revenue and utilization grow faster than the depreciation step-up — as they did spectacularly in 2025 (incremental operating margin ~67%) — margins expand. But the mechanism cuts both ways: in a demand-soft year, fixed depreciation on under-utilized leading-edge fabs is exactly what turned the FY2023 dip into a 5.5-point gross-margin contraction (59.6% → 54.4%). This is the quantitative heart of the residual cyclicality — TSMC’s margins are levered to utilization through a depreciation base that only ratchets upward. The overseas-fab dilution (2–4%) compounds this by adding higher-cost, lower-utilization capacity to the denominator. Management’s ability to hold gross margin near 60% through the FY2026 capex peak — absorbing both the depreciation ramp and the overseas drag — is the single most important number to track on the financials, and the base case assumes pricing power and Taiwan-fab efficiency offset it.

Verdict (Financial Quality): Economics decisively improve with scale. Incremental operating margin of ~67%, near-record gross/operating/net margins, ~35% ROE after a US$63B low-yielding cash drag (operating-asset returns are higher still), ~zero dilution, trivial SBC, conservative working capital, and 1.34x cash conversion. The single genuine pressure is the capex cycle (US$40.6B spent, US$52–56B guided, plus further overseas margin dilution), which will compress near-term FCF even as absolute FCF stays large. This is a high-return business at a high-reinvestment phase — the unit economics do not merely survive scale, they compound with it.

7. Capital Allocation

Verdict up front: TSMC’s capital allocation is among the best in global large-cap industrials, and the central reason the moat compounds rather than erodes.

The reinvestment engine. The defining question for TSMC is not how it returns cash but how it reinvests it. Capex ran NT$1,082,672M (2022), dipped to ~NT$950–956B (2023–24) through the inventory correction, then surged to NT$1,272,411M (US$40,561M) in 2025, with 2026 guided to US$52–56B. Capex has run ~34% of sales for years — alarming in almost any other business, but not here, because of the returns: ~35% ROE on modest leverage means ROIC tracks roughly three to four times a reasonable ~10% WACC. In the Greenwald/Marathon frame, this is the rare durable case where high returns are not competed away — demand captivity plus scale economics let TSMC reinvest enormous sums at incremental returns far above cost of capital. This is high-return reinvestment widening a moat, not a capex boom destined to mean-revert. The one mechanical caution: capex/depreciation has stayed persistently above 1.0x (1.82x/1.46x/1.87x over 2023–25) — continuous net capacity addition — and depreciation is ramping hard (NT$679,684M in 2025), which will pressure margins as assets season. The mitigant is that capacity is largely pre-sold (long-term agreements, prepayments, capacity-reservation deposits) rather than built speculatively.

Dividends — progressive, quarterly, deliberately subordinated to capex. The quarterly cash dividend rose from NT$3.50 (all of 2023) to NT$4.00–4.50 (2024) to NT$5.00–6.00 (2025, NT$22.00 annualized, +29% y/y). Cash dividends paid rose to NT$466,779M (US$14,880M) in FY2025. The dividend is conservatively set — ~27.5% payout, covered ~4.9x by OCF — and at a ~0.9% yield this is explicitly not an income story but a reinvestment compounder returning the residual. The policy hierarchy is clear: capex first, progressive dividend second, both funded internally with no equity raise.

Buybacks, issuance, SBC — clean. TSMC does essentially no buybacks (treasury activity is restricted-stock reclamation, not open-market repurchase). Share issuance is a rounding error, the ~25.9B share count is stable, and SBC is minimal. This deliberately contrasts with US semis peers (Nvidia, AMD, TI, Intel) that lean on buybacks and carry higher SBC dilution. TSMC shareholders are not diluted, and management is not buying back stock at cyclical peaks.

M&A — organic by design. No material acquisitions across 2023–2025. Where TSMC has expanded abroad it used customer- and government-co-funded JVs: JASM (Japan, with Sony/Denso/Toyota; TSMC ~72.6%) and ESMC (Dresden, with Bosch/Infineon/NXP). Intelligent structuring — partners and subsidies absorb part of the capital, and the JV partners are also anchor demand, de-risking utilization. By avoiding premium-priced acquisitions, TSMC sidesteps the most common value-destruction channel.

Geographic diversification — a customer-funded premium, not a political tax. The hardest call. Overseas fabs dilute gross margin 2–3% early, widening to 3–4% later — a real cost. But the evidence says moat-reinforcing, not value-destroying: it is heavily subsidized (US$6.6B CHIPS grants + up to US$5B loans for Arizona; US$2.4B of grant proceeds through FY2025 investing cash flow alone); it is already working (the first Arizona fab swung to a ~NT$16.1B/US$514M profit in 2025 from a NT$14.3B 2024 loss, with Apple as anchor and Nvidia Blackwell running at Fab 21); and customers are paying up for supply-secure US capacity, making the dilution effectively a customer-funded insurance premium against the franchise’s single largest tail risk.

Incentives and ownership. Executive pay is strongly variable and profit-linked: CEO C.C. Wei’s FY2025 package totaled ~NT$2,423M (US$77.2M), of which base salary was under 1% — the rest bonus, stock awards and a large “All Other Compensation” line (the opacity of which warrants a diligence question). The structure ties pay to profit-sharing and performance rather than tenure. The caution: insider ownership is thin — all directors and officers as a group hold ~0.23%; the largest holder is the R.O.C. National Development Fund at 6.38%; founder Morris Chang is long retired and there is no controlling family. TSMC is widely-held and professionally-managed — governance is sound, but alignment rests on pay design rather than personal equity skin in the game.

Verdict (Capital Allocation): Decisively intelligent, and integral to the thesis. Reinvests a 35%-ROE earnings stream at returns far above cost of capital; self-funds a record capex program plus a growing dividend; avoids dilution, peak buybacks and overpriced M&A; uses subsidies and co-investment to extend the franchise geographically with first results validating the bet. Watch-items (capex/depreciation persistently >1.0x; 2–4% overseas margin dilution; thin insider ownership and an opaque “Other” comp line) are real but secondary.

8. Changes and Headwinds — Last Two Years

Strategic / capacity changes.

• US$165B US investment intention (2025-03-04 6-K): an additional US$100B on top of the prior ~US$65B — three new Arizona fabs, two advanced-packaging facilities, and an R&D center. Arizona Fab 1 reached high-volume production end-2024; Fab 2 under construction; Fab 3 commenced 2025.
• Japan/JASM (Kumamoto): Fab 1 reached volume production Dec-2024; a second fab began construction Oct-2025 with Sony/Denso/Toyota as minority partners.
• Germany/ESMC (Dresden): ground broken Aug-2024 with Bosch/Infineon/NXP; specialty/mature focus.
• Technology: N2 (2nm) entered volume production in 2025 — TSMC’s first gate-all-around node; A16 (16Å, backside power) risk production slated for 2026.
• CoWoS advanced packaging scaling ~4x (from ~35k to ~130k wafers/month by end-2026) to feed AI-accelerator demand.

Capital-return changes. Quarterly dividend raised twice (NT$4.50 → 5.00 → 6.00), a +29% FY2025 increase and a continued shift to a firmly quarterly cadence.

Regulatory / litigation developments.

• Export-control regime tightened (Dec-2024 Entity List, Jan-2025 AI Diffusion Framework / Foundry Due-Diligence Rule; Taiwan blacklisting Huawei/SMIC June-2025) — net advantage vs SMIC, but a live compliance exposure, including a potential >US$1B Commerce penalty over chips allegedly rerouted to Huawei via Sophgo (framed as a risk factor, not an accrued liability).
• Patent litigation: Feb-2025, Longitude/Marlin filed an ITC complaint and an E.D. Texas suit alleging infringement of five US patents; the ITC instituted an investigation 2025-03-21. Outcome not estimable; no accrual. The live legal item to monitor.

Headwinds. (1) Overseas-fab gross-margin dilution (2–3% in FY2025, guided 2–3% more in FY2026); (2) the FY2026 capex step-up (US$52–56B) compressing near-term FCF; (3) NT$ appreciation risk on USD-denominated revenue; (4) the AI-capex cycle’s eventual digestion; (5) the ever-present Taiwan geopolitical overhang.

Verdict (Changes): On net, the last two years strengthened the franchise — the AI mix shift, node leadership (N2), and packaging chokehold widened the moat — while adding two manageable financial headwinds (overseas dilution, capex step-up) and one unquantifiable but escalating tail (geopolitics/export-control compliance).

9. Risk Analysis

Catastrophic-loss assessment. A total or near-total permanent loss is plausible only through Risk #1 (a kinetic Taiwan conflict destroying or denying the asset base). Short of that, the realistic downside is a cyclical/valuation drawdown (Risks #2, #12) of the kind the −13%/yr bear scenario captures, not a permanent impairment of the franchise. The geopolitical tail is genuinely binary and genuinely unquantifiable — it is the reason a business this good carries a structurally lower multiple than its economics alone would warrant.

10. Valuation Discussion (Embedded Expectations)

Embedded-expectations and scenario analysis only. No price target, no recommendation. Per-share figures per ADS (1 ADS = 5 ordinary shares); market data 2026-06-05/06.

The setup. At ~US$415/ADS, TSMC carries a US$2,153B market cap and — on ~US$63B net cash — an enterprise value of ~US$2,090B. That EV capitalizes US$54B of FY2025 net income and just US$32B of FY2025 FCF. The entire question reduces to one tension: TSMC trades at the richest multiple in its own recorded history (99.5th percentile on P/E, P/B and P/S vs its own ~10-year history) while simultaneously looking cheap on forward earnings (~21x) and against every leading-edge peer. Both are true.

Current multiples (cleanly computed; some aggregator EV/EBITDA and P/S figures are corrupted by an ADR-vs-ordinary-share unit mismatch and are discarded):

The wedge between ~40x P/E and ~67x P/FCF is the cycle in one number: FCF is artificially depressed because FY2025–26 capex front-loads the AI build-out. Earnings are the cleaner lens today; FCF only “earns its multiple” once capex and depreciation reconverge.

Peer comparison (forward P/E, the cleanest cross-read given ADR unit issues):

The striking fact: TSMC carries the lowest forward P/E in the entire leading-edge complex despite the best ROE, a net-cash balance sheet, zero dilution and the strongest guided growth. On business quality it should arguably trade above GFS/UMC and near ASML, yet sits ~12 turns below ASML. The discount is explained by (a) the Taiwan geopolitical risk premium, (b) capital intensity and residual cyclicality, © a foreign-issuer/ADR discount, and (d) customer concentration.

Reverse-DCF — what the price embeds. Running EV (US$2,090B) backward through a two-stage DCF (10-yr stage 1, 3% terminal, 9.5% discount rate), the required FCF growth depends on the base:

• On FY2025 actual FCF (US$32B): the market requires ~22–23% FCF CAGR for a decade — a demanding bar.
• On normalized post-peak FCF (~US$55B) — the level FCF reaches if capex/revenue reverts from 33% toward the mid-20s% as the build matures — the market requires only ~14–15% FCF CAGR.

This is the embedded-expectations core. Bears anchor on actual FCF and (correctly) call ~22% a stretch; bulls anchor on normalized FCF and (correctly) call ~14% achievable if AI demand holds. The disagreement is not really about TSMC’s growth rate — it is about whether capex intensity normalizes. The market is underwriting roughly management’s base case (high-teens-to-20% revenue compounding, 50%+ operating margins); it is not pricing a blue-sky mania, but it is not discounting an air-pocket either. There is no margin of safety embedded.

Scenario analysis (4-year, to FY2029):

• Bear combines an AI-capex air-pocket and/or a Taiwan geopolitical de-rating, with overseas dilution dragging margins to 45% and the multiple compressing toward TSMC’s historical commodity-foundry range — a ~−13%/yr return (multiple and earnings impairment).
• Base assumes guidance roughly holds but decelerates, margins steady ~50%, multiple normalizes to a still-premium 24x. The base-case FY2029 EPS/ADS (~US$19.3) equals current consensus NTM — confirming the optical ~+85% jump from US$10.44 is a calendar-2026 estimate off the trailing base, not a structural break. Return: a thin ~+3%/yr — you earn the growth but the rich entry does the work against you.
• Bull requires a sustained AI super-cycle, N2/A16 pricing power fully absorbing overseas dilution, and the multiple holding at 30x — ~+16%/yr.

The asymmetry is informative: downside (−13%/yr) is wider than base upside, because the entry multiple starts at a 99.5th-percentile point with little cushion.

Resolving the paradox. How is TSMC both its richest-ever self and the cheapest leading-edge name on forward P/E? The two metrics measure different things. The own-history percentile compares TSMC to its own past, when it was priced as a cyclical commodity foundry at 12–18x; the AI super-cycle has structurally re-rated the franchise toward a secular-growth multiple, which by construction prints as “richest ever” — that percentile is the risk flag (no cushion; a re-rate back toward the mean is the largest downside). The forward 21.5x looks cheap only if the embedded EPS leap to ~US$19.3 materializes; the cheapness is entirely conditional on the forward E being correct. The honest synthesis: the forward multiple embeds the growth — it does not discount it.

An EPV / asset-value cross-check. A Greenwald earnings-power lens frames the floor differently from the growth lens. Strip growth entirely and capitalize FY2025 normalized operating earnings (~US$62B operating income, tax-effected to ~US$51B) at an 8–9% unlevered cost of capital: a no-growth EPV lands roughly US$570–640B of operating value, plus ~US$63B net cash — call it ~US$650–700B, or roughly one-third of the current ~US$2.09T EV. In other words, about two-thirds of today’s price is growth value — the franchise’s ability to reinvest at high returns — and only one-third is the capitalized current earnings power. That is the mathematical statement of “no margin of safety”: the EPV floor is far below the market price, so a growth disappointment has a long way to fall before valuation support engages. It also reframes the ~US$63B net cash — real, but ~3% of EV, not the cushion it would be in a smaller-cap. The bull’s rejoinder is that for a business compounding at 20%+ with a widening moat, paying ~3x EPV is unremarkable; the bear’s is that paying 3x EPV at a 99.5th-percentile multiple for a capital-intensive, geopolitically-exposed cyclical is precisely how permanent capital is lost when the cycle turns. Both are correct given their premises — which is why this is a price/judgment question, not a quality question.

Verdict (Valuation): Embedded expectations price approximately management’s base case. This is neither a bubble multiple nor a bargain — it is a full price for a high-quality compounder, with risk-reward gated on execution rather than on a depressed entry point. The reverse-DCF requires ~22% growth on today’s depressed FCF or ~14% on a normalized base; scenarios span roughly −13%/yr to +16%/yr over four years, with a thin ~+3%/yr base case. The defining feature is the absence of multiple-based margin of safety.

11. Variant Perception

Consensus view. TSMC is a wide-moat, AI-levered secular compounder — the indispensable manufacturer of the AI era — that deserves a premium-but-reasonable ~21x forward multiple because it will compound revenue at ~20%+ with 50%+ operating margins for years. Consensus is broadly correct on the business and is pricing roughly the base case.

Strongest bull case. The AI super-cycle is early, not late; accelerator demand compounds at the guided mid-to-high-50s% for years; TSMC’s leading-edge monopoly plus the CoWoS packaging chokehold let it raise prices (N2 “no discounts”) and expand margins through the overseas-fab dilution; capex intensity normalizes as the build matures, inflecting FCF sharply higher; and the forward multiple proves cheap in hindsight as EPS marches from US$10 toward US$25+/ADS. In this world the 99.5th-percentile valuation is justified by a structural earnings re-rating, and the stock compounds at ~15%/yr.

Strongest bear case. The AI-capex cycle is a textbook boom; hyperscaler digestion produces an air-pocket that craters leading-edge utilization on a now-58%-HPC revenue base; the overseas-fab margin drag compounds rather than offsets; and — at a 99.5th-percentile starting multiple with zero cushion — the market re-rates TSMC back toward its historical cyclical-foundry mean even on modest earnings disappointment. Layer in the unquantifiable Taiwan tail, and the stock delivers a ~−13%/yr permanent multiple-plus-earnings impairment.

The 3–5 assumptions that matter most.

1. AI-accelerator demand persistence — does HPC compound near guidance, or digest? (Drives both the E and the multiple.)
2. Capex normalization — does capex/revenue revert toward the mid-20s%, inflecting FCF, or stay elevated? (Drives the FCF-based valuation.)
3. Pricing power vs. overseas dilution — does N2/A16 pricing fully absorb the 2–4% overseas margin drag? (Drives margin durability.)
4. Competitive yield gap — do Intel 18A/14A or Samsung SF2 close the gap and win marquee customers by 2027? (Drives the moat’s width.)
5. Taiwan geopolitics — the binary tail that caps the multiple regardless of fundamentals.

Where variant perception could exist. Not in “is this a great business” (it is, and that is consensus) but in the path: a contrarian could argue either (a) the market under-appreciates the FCF inflection once capex normalizes (bull variant — the forward multiple is genuinely too cheap), or (b) the market is anchoring on a re-rated secular multiple that will not survive the first AI digestion cycle (bear variant — the 99.5th-percentile signal is the tell). The evidence supports neither extreme decisively; it supports the view that the business is mispriced by no one and the price leaves no room for error.

12. Fact vs. Interpretation Table

13. Open Questions

1. Capex trajectory beyond FY2026 — does the US$52–56B FY2026 guide mark a peak intensity, or a new plateau? The FCF-based valuation hinges on this.
2. Normalized FCF level — what is mid-cycle FCF once capex/depreciation reconverge? The ~US$55B normalization is an assumption, not a disclosure.
3. Nvidia/AI concentration durability — how much of the 17% Nvidia share is structural vs. front-loaded by the current accelerator cycle?
4. Overseas-fab terminal margins — does the dilution truly cap at 3–4%, and do subsidies/pricing offset it durably, or does it compound?
5. Export-control penalty (Sophgo/Huawei) — size and timing of any Commerce penalty; the ITC patent case outcome.
6. The “All Other Compensation” line — what comprises the largest component of CEO pay (NT$1,142M)? A governance opacity worth resolving.
7. Effective tax rate durability — does the ~17% ETR hold given the Taiwan unappropriated-earnings surtax and global minimum-tax (Pillar Two) developments?
8. Pricing power ceiling — how much further can N2/A16 ASPs rise before customers (Apple, Nvidia) push back or accelerate in-housing/dual-sourcing?

14. What Must Be True

For the bull case (sustained ~15%/yr compounding):

• AI-accelerator demand compounds near the guided mid-to-high-50s% CAGR for several more years (no durable digestion air-pocket).
• Capex intensity normalizes toward the mid-20s% of revenue, inflecting FCF from ~US$32B toward US$55B+.
• N2/A16 pricing power fully absorbs the 2–4% overseas-fab margin dilution, holding operating margin ≥50%.
• TSMC retains its >90% leading-edge share — Intel/Samsung do not close the yield gap with marquee wins.
• No Taiwan geopolitical escalation.
• Falsification test: a single demand-digestion year of flat-to-down HPC revenue, or capex/revenue staying ≥32% through FY2027, would break the FCF-inflection thesis and the bull IRR.

For the bear case (multiple-plus-earnings de-rating, ~−13%/yr):

• An AI-capex air-pocket cuts leading-edge utilization on the 58%-HPC revenue base.
• The 99.5th-percentile multiple re-rates toward the historical cyclical-foundry mean (12–18x) on the first earnings disappointment.
• Overseas-fab dilution compounds; pricing power proves insufficient to offset.
• Falsification test: continued ≥20% revenue growth with margin stability and a visible FCF inflection (capex normalizing) for two consecutive years would refute the de-rating thesis and re-establish the secular multiple.

The synthesis: the business case is not seriously contestable; the price case is entirely about whether AI demand and capex normalization deliver on schedule, with a binary geopolitical tail capping the multiple either way. There is no margin of safety in the multiple at ~US$415 — the analysis points to a great business whose stock requires execution to merely meet, not beat, embedded expectations.

15. Source Appendix

See Appendix B — Source Appendix below for the full, categorized source list (primary filings, company disclosures, third-party market data, and trade press), with URLs and access dates. Primary anchor: TSMC FY2025 Form 20-F (filed 2026-04-16, period ended 2025-12-31), supplemented by the trailing 6-K corpus, public XBRL company facts, and public market-data aggregators; all material quantitative figures are reconciled to the 20-F.

No price target and no BUY/SELL recommendation appears anywhere in Sections 1–15. The only position taken in this document is the clearly-labeled “Claude’s Take” block at the top, which is the author’s own subjective view.

Appendix A — Diligence Questionnaire — Taiwan Semiconductor Manufacturing (NYSE: TSM)

Supplemental to the research note. Fact/Interpretation/Assumption labels applied where material. No price target or recommendation.

General

What thoughtful questions have other investors asked about this company? The recurring serious questions are: (1) Is the AI demand real and durable, or a capex bubble? — given HPC is now 58% of revenue. (2) What is TSMC actually worth net of the Taiwan tail risk — i.e., how much of the apparent forward-P/E discount to ASML is a rational geopolitical/ADR haircut versus an opportunity. (3) Will the overseas-fab build (US$165B in Arizona, plus Japan/Germany) permanently impair the ~60% gross margin, or do subsidies and pricing offset it? (4) Does capex ever stop? — capex/depreciation has sat above 1.0x for years; investors debate when FCF inflects. (5) Is the leading-edge monopoly permanent, or do Intel 18A/14A and Samsung SF2 eventually break it? The quality of the business is not seriously debated; the debates are about price, geopolitics, and capex.

Cyclicality & Earnings Nature

Are earnings at a cyclical high or low? Interpretation: near a cyclical high on margins (GM 59.9%, OpM 50.8% are records or near-records) and on growth (+31.6%), driven by the AI/HPC up-cycle. The −4.5% revenue dip in FY2023 is the reminder that the franchise is still cyclical underneath the AI surge.

Driven by the external environment or internal actions? Both. The AI-accelerator demand wave is external (Nvidia/AMD/hyperscaler capex), but the ability to capture it at 60% margins — leading-edge yield, capacity, CoWoS packaging, pricing power — is internal and structural.

How stable are revenues? Recurring at the franchise level (multi-year design wins, capacity-reservation agreements), cyclical at the volume level (smartphone/PC/inventory cycles). Currently the AI cycle is masking the cyclicality.

Outlook for products/services? Strong leading-edge roadmap: N2 (2nm) volume production began 2025; A16 (16Å) risk production 2026. Management guides ~25% USD revenue CAGR 2024–2029 and mid-to-high-50s% AI-accelerator CAGR (Assumption — management hypothesis).

How big will this market be? Growing. Global semis ~US$772B (2025) heading toward ~US$975B (2026, WSTS); foundry is the structurally advantaged slice, and leading-edge logic (TSMC’s ~74%-of-revenue arena) is the fastest-growing sub-segment. Predominantly international demand (North America 75% of TSMC revenue, much of it for global end-products).

Business Quality & Competitive Moat

Is the industry getting more or less competitive? At the leading edge, less — it is consolidating toward a TSMC monopoly (Samsung losing share, Intel exited merchant leading-edge). At the trailing edge, more — Chinese state-subsidized overcapacity. TSMC sits overwhelmingly in the good half.

How profitable is the business (ROIC, ROE)? ROE ~35% (FY2025); ROIC tracks close to ROE given modest leverage, roughly 3–4x a ~10% WACC. Among the most profitable manufacturers globally.

How profitable is the industry — how many competitors, what barriers? At the leading edge: effectively three attempted competitors (TSMC, Samsung, Intel), one dominant. Barriers are among the highest in any industry — US$20–30B+ per fab, ASML EUV/High-NA rationing, cumulative-volume yield learning curves, process talent, ecosystem lock-in.

Can the business be easily understood? Yes at a high level (it makes chips for others), but the technology and the capital-cycle dynamics are genuinely complex; the moat mechanics require domain understanding.

Can it be undermined by foreign low-cost labor? No — this is a capital- and knowledge-intensive business, not labor-arbitrage. The threat is state-subsidized capital (China at mature nodes), not cheap labor.

Do brands matter? Not consumer brands, but TSMC’s reputation for yield, reliability and trustworthiness (no channel conflict) is a genuine reputational asset that anchors customer captivity.

Nature of competition? Technology/yield leadership and capacity availability, not price (TSMC is the premium-priced option and gaining share — the opposite of price competition).

Customers’ switching costs? High — re-design against a new PDK, IP re-qualification, multi-million-dollar mask sets per tape-out, yield-ramp risk. Demonstrated by customers returning to TSMC after dual-sourcing attempts (Qualcomm, Tesla’s top chip).

Financial Condition & Balance Sheet

Assets not fully recognized on the balance sheet? The process-technology know-how, customer relationships and ecosystem (the actual moat) are internally generated and largely unrecognized. R&D is expensed (NT$246B in 2025).

Off-balance-sheet liabilities? The material item is the parent guarantee of TSMC Arizona’s CHIPS-related obligations (up to US$6.6B funding + US$5B loans). Long-term equipment/materials/energy purchase commitments exist but are largely qualitative in disclosure.

How conservative is the accounting? Conservative. Deloitte audited since 1987; inventories flat and receivables +3% against +31.6% revenue (no channel-stuffing or pulled-forward revenue); trivial SBC; ~zero dilution; no one-time items distorting FY2025. Quality-of-earnings screen is clean.

How CapEx-hungry is the business? Extremely — capex ~34% of revenue (US$40.6B FY2025; US$52–56B guided FY2026), capex/depreciation 1.87x. This is the defining financial feature and the main pressure on near-term FCF. However, the reinvestment earns ~35% ROE, so it is value-creative, not value-destroying.

Capital Allocation & Management

How much FCF, and how is it used? FCF US$32.0B FY2025 (after US$40.6B capex). Priority hierarchy: capex first (the moat engine), progressive dividend second (NT$22/share, ~28% payout), with essentially no buybacks and no M&A. Philosophy: reinvest at high returns, return the residual, never dilute.

Significant acquisitions recently? No — growth is organic. Overseas expansion uses customer/government-co-funded JVs (JASM with Sony/Denso/Toyota; ESMC with Bosch/Infineon/NXP), avoiding premium-priced M&A.

Buying back shares? Essentially no (treasury activity is restricted-stock reclamation). A deliberate contrast with US semis peers.

Issuing large amounts of new shares to insiders? No — share count is flat (~25.9B); SBC is 0.03% of revenue.

Compensation policy of directors/management? Heavily variable and profit-linked (CEO base <1% of a ~US$77M package). Caveat: a large, opaque “All Other Compensation” line warrants scrutiny (Open Question).

Motivations of management? Professional management (no controlling family; Morris Chang retired). Pay is well-designed and performance-linked, but insider ownership is thin (~0.23% for all directors/officers), so alignment rests on incentive design rather than personal equity stakes.

Valuation & Market Data

Is the stock an ADR, MLP, or K-1 issuer? An ADR (NYSE: TSM; 1 ADS = 5 ordinary shares; ordinary shares trade as TWSE: 2330). Foreign private issuer filing 20-F/6-K under IFRS. No K-1; no MLP. ADR holders bear Taiwan dividend withholding and depositary service fees.

Dividend policy? Progressive, now quarterly: NT$22/share FY2025 (+29% y/y), ~0.9% yield, ~28% payout — a residual-return policy subordinate to capex.

How profitable is the business? Exceptionally — 44.5% net margin, ~35% ROE, 59.9% gross margin.

Is net income diverging from cash from operations? No — OCF/net income = 1.34x; earnings are over-covered by cash. (FCF is lower than net income only because of the heavy capex, which is investment, not an earnings-quality red flag.)

Risks & Downside

What factors would cause the stock to decline? An AI-capex air-pocket cutting utilization; a multiple de-rate from the 99.5th-percentile starting point on any earnings disappointment; overseas-fab margin compounding; export-control penalties; NT$ appreciation; and — the dominant tail — Taiwan-China geopolitical escalation.

Risk of a catastrophic loss? Yes, but concentrated in essentially one scenario: a kinetic Taiwan conflict destroying or denying the asset base. Short of that, downside is a cyclical/valuation drawdown, not a permanent franchise impairment.

Chance of a total loss? Low in probability but non-zero and genuinely binary, via the Taiwan tail. This unquantifiable risk is precisely why a business with these economics trades at a structurally lower multiple than peers like ASML.

Recent News & Events

Has the business environment changed recently? Yes, materially and favorably on demand: HPC/AI overtook smartphones as the #1 platform (58% of revenue), and Nvidia surged to the #2 customer (17%). N2 entered volume production in 2025.

Significant acquisitions? None — but a major capital event: the US$165B US investment intention (March 2025), expanding the Arizona program to three fabs + two packaging facilities + an R&D center.

Change in accounting policies? None material identified; reporting is consistent IFRS.

Recent changes — new markets, facilities, management? New facilities: Arizona (Fab 1 in volume production, Fabs 2–3 progressing), Japan/JASM (Fab 1 producing, Fab 2 under construction), Germany/ESMC (Dresden). Litigation to monitor: the Longitude/Marlin ITC/E.D. Texas patent case (instituted March 2025) and the potential Sophgo/Huawei export-control penalty.

Appendix B — Source Appendix — Taiwan Semiconductor Manufacturing (NYSE: TSM)

Primary sources first. All material quantitative figures are reconciled to TSMC’s SEC filings (20-F / 6-K). Access dates June 2026.

1. Primary — SEC filings (TSMC, CIK 0001046179)

2. Company disclosures / management commentary

• TSMC Q1 2026 and Q4 2025 earnings-call commentary — capex guidance (US$52–56B FY2026), overseas-fab margin-dilution guidance (2–3% FY25, 2–3% FY26), ~25% USD revenue CAGR and AI-accelerator mid-to-high-50s% CAGR targets. (Management commentary treated as hypothesis and validated against filings and external data.)
• TSMC investor relations / IR site (tsmc.com) — corporate structure, technology roadmap (N2, A16), CoWoS/3DFabric packaging.
• TSMC earnings-call transcripts (recent quarters) — company IR and public transcript sources.

3. Third-party quantitative data (reconciled, not primary)

• Public market-data aggregators — snapshot, multi-period statements, and own-history valuation percentiles (composite percentile 99.5 vs own ~10-yr history; PE 99.4 / PB 99.6 / PS 99.6), accessed 2026-06-05.
• Public market-data aggregators — price (US$415.17), market cap (~US$2.15T), 52-wk range (US$202–450), ROE, multiples, peer forward P/E table. Caveat: some aggregator EV/EBITDA (2.98x) and P/S (0.52x) figures are corrupted by an ADR-vs-ordinary-share unit mismatch and were discarded; clean multiples computed from the 20-F + net-cash position.

4. Industry / market data (third-party)

• WSTS — global semiconductor market forecast (~US$772B 2025 / ~US$975B 2026), wsts.org, accessed 2026-06-06.
• TrendForce / Counterpoint / Design-Reuse — foundry market-share data (TSMC ~69.9% 2025; Samsung 7.2%, SMIC 5.3%, UMC 4.35%, GFS 3.87%); 2026 capex estimates; node-yield and pricing commentary; overseas-fab profitability. Multiple articles, accessed 2026-06-06.
• VisualCapitalist / PatentPC / marklapedus.substack.com — leading-edge share estimates (>90%), accessed 2026-06-06.
• Tom’s Hardware / TechTimes / PhoneArena — N2 yield (~65–70%), N2 wafer pricing (~US$30,000), Tesla/Qualcomm sourcing, Intel 18A / Samsung SF2 yield comparisons, accessed 2026-06-06.
• DigiTimes / Astute / WccfTech — CoWoS capacity expansion (~35k→~130k wpm), Nvidia CoWoS booking (~60%), accessed 2026-06-06.
• CNBC — Taiwan blacklisting Huawei/SMIC (June 2025), accessed 2026-06-06.
• Trefis / Electronics Weekly — Intel Foundry external revenue (~US$174M Q1’26), accessed 2026-06-06.

5. Analytical frameworks

• Greenwald & Kahn, Competition Demystified (barriers-to-entry taxonomy, market-share-stability and ROIC tests, EPV) and Marathon, Capital Returns (supply-side capital-cycle analysis), applied in the Business Quality, Industry, Competitive Position, Capital Allocation and Valuation sections.

Distinguishing Fact / Interpretation / Assumption / Open Question is carried in the body (Fact-vs-Interpretation table) and the Open Questions section. Management commentary is labeled and treated as hypothesis throughout.