Taiwan Semiconductor Manufacturing Company Limited (NYSE: TSM) — The Widest Moat in Technology, Now Spending Its Cushion

Ticker: NYSE: TSM (ADR; ordinary shares TWSE: 2330) · Sector: Information Technology — Semiconductors (dedicated foundry) Date: August 7, 2026 Coverage status: UPDATE. Prior report dated June 6, 2026. Primary sources: FY2025 Form 20-F (filed 2026-04-16); Form 6-K of 2026-07-16 (Q2 2026 results and Q3 guidance); Q2 2026 earnings call transcript (2026-07-16); Form 6-K of 2026-07-13 (June revenue). IFRS/TIFRS, reported in New Taiwan Dollars (NT$). 1 ADS = 5 ordinary shares; 25,932 million ordinary shares outstanding. Balance-sheet items converted at the company’s own NT$32.00/US$1.00 guidance assumption unless noted.

⚡ Claude’s Take

This block is the author’s own subjective opinion. It is general information, not investment advice, and not a recommendation to buy or sell any security. The analysis that follows this block is deliberately position-free and carries no recommendation and no price target — that discipline is intact everywhere below this block.

Verdict: HOLD — unchanged in direction, but I am raising the accumulation zone and lowering my conviction on durability. Accumulate into weakness below roughly US$360 (~21x FY2026E EPS, and the level of the rising 200-day EMA); back up the truck below roughly US$300 (~17x). Not a short, and I would not sell a long-term position here. Two months ago I called this “great business, full price,” and named US$350 and US$280 as the zones where I would act. The stock is flat since — US$414 then, US$418 now — but that flatness conceals a genuine round trip: an all-time high of US$477.57 on 30 June, a 21.5% drawdown to US$374.67 on 29 July, and an 11.6% recovery. In the meantime the business delivered the single best quarter in its history. I am moving my zones up roughly US$10–20 because the earnings base underneath them moved up more than that.

Tag: “The moat is still the widest in technology. It is just no longer getting wider for free.”

What changed, and why it matters more than the price didn’t. Q2 was extraordinary: revenue US$40.20B (+33.7% in USD), gross margin 67.7% — up 9.1 points year-on-year — operating margin 60.3%, net income +77.4%, ROE 40.3%, and an incremental operating margin of 90%. Management raised FY2026 revenue growth to “slightly above 40%” and, on the same call, disclosed a 2027 price increase of 5–10% across advanced and mature nodes (Nikkei/Reuters). Foundry share rose again, to 72.3%. On the evidence, this is a business getting stronger, not weaker. And yet the market sold the print — down 2.3% on 16 July on the heaviest volume of the period, and another 2.8% the next day. That reaction is the most informative fact in this report, and it was right. Three things arrived alongside the beat. First, capex was raised a second time, to US$60–64B — 35–38% of revenue against 33% last year — with the CFO promising the next three years will be “even more significantly higher,” the CEO admitting part of the raise buys tool-price inflation rather than wafers, and an additional US$100B for Arizona announced with no schedule attached. Second, the CEO capped his own margin ambition: asked about memory’s 86% margins, he said “about 68%, I would be happy about that” — with Q2 at 67.7% and Q3 guided to 66%. The four-quarter margin expansion that generated the earnings surprise is at its self-declared ceiling. Third, and most importantly for the thesis, my stated bearish trigger was half-hit. In June I said I would turn negative if Intel 18A/14A closed the yield gap and won marquee customers. KeyBanc now puts 18A yields above 85%, up from ~65% in a quarter, against N2 at ~90% — a 10-15 point lead compressed to roughly five — and ASML has confirmed Intel is the first to ship high-volume logic on High-NA EUV, the technology TSMC has pointedly declined to adopt. The customer half is not met: the reported design wins are engagements, not disclosed wafer commitments, and Intel Foundry’s external revenue was ~US$174M a quarter last we checked against TSMC’s ~US$40B. Half a trigger is not a verdict change. It is a reason to stop paying for infinite durability.

The framing, evidence-based: a high-beta quality-growth name in an ordinary correction inside an intact uptrend — not a falling knife, and emphatically not a crowded momentum trade. The factor model is unusually clear here. Momentum loading is just +0.16 in the base model despite 12-month relative strength of +81.6%, so you are not being paid a crowding premium to sell. Price sits above a rising 200-day EMA at US$362, 12.4% off the high. But the same model contains the fact that should govern position sizing rather than timing: after the market factor, TSM’s largest systematic exposure is the Taiwan country factor at +0.93 — nearly as large as its market beta, and the empirical nearest-neighbours are Taiwan ETFs before they are foundry peers. The geopolitical concentration in this memo is not an abstraction; it is roughly half the stock’s systematic risk. On valuation, both true things are in tension: at 10.8x book TSMC has never been more expensive in its recorded history, above every prior year-end and every prior intra-year high; at ~24x FY2026E earnings it is cheaper than in June, because trailing EPS is up ~28% while the price is up 0.9%. The multiple did not get paid for — it got grown into. What I will not do is pay today’s price for a reverse-DCF that needs free cash flow to more than triple, from US$35.5B to roughly US$126B, on a capex normalisation management has explicitly refused to promise.

Conviction: medium, trimmed from medium. Flips me bullish: Intel Foundry’s external revenue line failing to inflect over the next two prints while TSMC’s 2027 price increases stick — that would restore the durability premium — or a de-rating toward US$360 with the earnings trajectory intact. Flips me bearish: a disclosed, multi-billion-dollar leading-edge wafer commitment from a top-five fabless customer to Intel or Samsung; a second consecutive quarter of inventory build without a revenue step; or the first guide that pairs rising capex with decelerating revenue — the classic late-capital-cycle signature, and the one thing this company has never yet done.

🔄 Changes Since the June 6, 2026 Report

This is a follow-up engagement. The prior memo’s baseline thesis — a near-monopoly leading-edge foundry with an exceptional but fully-priced franchise — survives. Eight things changed materially in nine weeks.

Falsification tests from the prior memo — status. The June report’s bull test was “capex intensity normalising into an FCF inflection while AI demand holds.” It has broken: capex intensity rose from 33.4% to 35–38% and FCF conversion fell to 51%. The bear test was “Intel 18A/14A or Samsung SF2 demonstrably closing the yield gap and winning marquee leading-edge customers.” It is half-hit — the yield half on credible third-party evidence, the customer half not at all. The Taiwan-escalation test has not been hit; if anything US–China chip policy loosened, with case-by-case export licensing replacing the presumption of denial. Net: the business outperformed the prior memo’s base case; the durability evidence deteriorated at the margin; the valuation question is unchanged in character and slightly improved in level.

📈 Stock Price Action — Five-Year Event Map

TSMC’s ADS has compounded roughly seven-fold in less than four years — from a five-year low of US$57.21 (2022-11-03) to an all-time high of US$477.57 (2026-06-30) — and last closed at US$418.20 (2026-08-06), 12.4% off that high. The 52-week range is US$225.26–US$477.57; the stock is +31.4% year-to-date against the 2025 close, and +0.9% against this report’s prior reference of US$414.41 (2026-06-05), having round-tripped +15% and then −21.5% in between. The 200-day EMA sits at US$362.11 and is rising; the 21- and 50-day EMAs are clustered near US$412–414. (Price levels are FACT, from the AZI adjusted series to 2026-08-06. Attributed drivers are INTERPRETATION.)

Cycle narrative. (1) The 2021–22 decline was macro, not company-specific: the fastest rate-tightening cycle in four decades compressed long-duration multiples while a genuine PC and smartphone inventory correction hit volumes. (2) The bottom in November 2022 coincided with two unrelated catalysts — a disclosed Berkshire Hathaway stake and China’s reopening — and was followed by the beginning of the AI narrative, crystallised by Nvidia’s May 2023 guidance, which lifted TSMC 12.0% in a session. (3)–(4) Through 2024 the AI story converted from narrative into foundry orders; the stock repriced on successive guidance raises, with the July 2024 setback a reminder that US–Taiwan policy headlines move this security independently of its P&L. (5) January 2025’s DeepSeek session — a 13.3% single-day fall, still the largest in five years — was the market briefly pricing the possibility that AI compute demand was less than assumed; the April 2025 tariff shock and its reversal two sessions later were pure policy volatility. (6) From April 2025 the super-cycle ran with few interruptions: HPC overtook smartphone as the dominant platform, N2 entered volume production, and the multiple expanded alongside the earnings. (7) The June–July 2026 drawdown was not a TSMC event. Bank of America’s “bubble risk” note took the whole complex down on 1 July; the 29 July session removed over US$1 trillion from chip market capitalisation on fears that AI infrastructure spending is “peaking faster than expected,” with Micron down as much as 13% and the KOSPI triggering circuit breakers. TSMC’s own record print landed inside that window and was sold on the capex raise and the N2 margin-dilution guide, not on the results. (8) The recovery since 30 July has been sector-wide and sentiment-driven, led by strong hyperscaler cloud commentary. Two facts are worth holding together: TSMC fell 21.5% peak-to-trough while SK hynix’s Seoul line fell 52% over a comparable window — the foundry is demonstrably the more defensive way to own this cycle.

1. Executive Summary

Taiwan Semiconductor Manufacturing is the world’s dominant dedicated semiconductor foundry: it manufactures logic chips to the designs of fabless customers — Apple, Nvidia, AMD, Qualcomm, Broadcom and, increasingly, the hyperscalers’ in-house silicon teams — and sells no chips of its own. In the quarter ended 30 June 2026 it generated US$40.20 billion of revenue (+33.7% year-on-year in US dollars, +36.0% in NT$), a 67.7% gross margin, a 60.3% operating margin, NT$706.6 billion of net income (+77.4%), a 55.6% net margin and a 40.3% return on equity. The single most revealing number is the incremental operating margin: 90% of every additional revenue dollar fell through to operating income. On a trailing twelve-month basis the company earned US$69.3 billion on US$138.8 billion of revenue and finished the quarter with US$67.2 billion of net cash.

The moat is unchanged in kind and confirmed in the period. It is economies of scale fused with customer captivity — Greenwald’s most durable category — and it produced three pieces of fresh evidence this quarter. Foundry share rose to 72.3% from 70.4%, which inverts the market-share-stability test: a dominant firm is not merely holding but gaining roughly two points a year at a 72% base. Pricing power was exercised across the entire portfolio, with reported 2027 base-price increases of 5–10% on advanced and mature nodes — the second of those being the structurally weak segment where TSMC has no monopoly. And the technology roadmap extended: N2 reached 3% of wafer revenue in its first meaningful quarter, A16 goes to mass production in 2H 2026, and A14 was confirmed for 2028 volume production with internal vehicles at roughly 90% device performance and 90% SRAM yield, plus new A13 and A12 derivatives for 2029.

Three things nevertheless cut the other way, and together they are why this update is more cautious than its predecessor rather than less. First, the process lead narrowed on credible third-party evidence. KeyBanc reports Intel’s 18A yields rising from ~65% to above 85% in one quarter against TSMC’s N2 at ~90%, and ASML has confirmed Intel as the first to ship high-volume logic on High-NA EUV — a technology TSMC has deliberately declined on cost and field-size grounds. TechInsights independently scores 18A ahead of N2 on performance while N2 leads decisively on density. Intel’s reported customer engagements remain unconfirmed design-ins rather than wafer commitments, so this is a narrowing of the lead, not a loss of it; but it is the first genuine narrowing in a decade. Second, capital intensity is rising into the boom. FY2026 capex was raised to US$60–64 billion — 35–38% of revenue, against 33.4% last year — the CFO has promised the following three years will be “even more significantly higher,” an additional US$100 billion was committed to Arizona without a schedule, and the CEO conceded that part of the increase buys tool-price inflation rather than capacity. Trailing free cash flow of US$35.5 billion converts only 51% of net income. Third, management capped its own margin ambition at “about 68%”, with Q2 at 67.7% and Q3 guided down to 66% on N2 dilution — meaning the four-quarter gross-margin expansion that drove the earnings surprise has, by the company’s own account, largely run its course.

The industry backdrop is favourable at the leading edge and newly complicated elsewhere. Section 232 tariffs of 25% on advanced logic took effect on 15 January 2026, and a simultaneous US–Taiwan agreement exempts or discounts Taiwanese chipmakers that expand US production — which reframes the Arizona programme from political tribute into a purchased tariff option and partly justifies the margin dilution it causes. Export policy toward China loosened, with case-by-case licensing replacing blanket denial. But memory has become a cost problem rather than a peer: DRAM prices are up more than 300% year-on-year as the big three shifted over 80% of capacity to HBM, and TSMC’s own CEO now cites “rising component prices” as the reason consumer and price-sensitive end markets are being challenged. TSMC’s response — a DRAM-wafer supply arrangement with Winbond feeding its wafer-on-wafer and SoIC stacks — is vertical insurance, not a memory entry.

Valuation is, as in June, the entire question, and it is more finely balanced than it was. At US$418.20 the ADS carries a US$2.17 trillion market capitalisation and a US$2.10 trillion enterprise value, trading at 31.3x trailing earnings, 10.8x book, 15.1x EV/sales, 21.1x EV/EBITDA and 61x free cash flow. On book value TSMC has never been more expensive — 10.8x exceeds every prior year-end and every prior intra-year high in its ten-year record. On forward earnings it is cheaper than in June, at roughly 24x FY2026 estimated EPS, because trailing earnings have risen about 28% while the price has risen 0.9%. A reverse-DCF at a 9% cost of capital and 3% terminal growth requires steady-state free cash flow near US$126 billion against US$35.5 billion today — which needs both a further doubling of revenue and a normalisation of capex intensity from 36% back toward 20%, the latter being precisely what management declined to promise. This memo takes no position on the security; it sets out the embedded expectations and the evidence on both sides.

2. Business Overview

The model, and why the structure matters. TSMC is the world’s largest pure-play (dedicated) foundry: it fabricates logic chips to its customers’ designs and designs or sells no chips of its own. That absence is the structurally decisive feature. An integrated device manufacturer such as Intel or Samsung both designs and manufactures, and therefore competes with the very fabless firms it would court as foundry customers. TSMC carries no such channel conflict — Apple, Nvidia, AMD, Qualcomm and Broadcom can hand it their most valuable intellectual property without fear that the foundry will build a competing product. Morris Chang’s founding insight in 1987, separating design from manufacturing, is both the reason the foundry market exists and the reason a pure play, not an IDM, dominates it. The company disclosed that in 2025 it deployed 305 distinct process technologies and manufactured 12,682 products for 534 customers.

How it makes money. TSMC sells wafer-fabrication capacity at defined process nodes, priced by node, design complexity and — increasingly — advanced-packaging content. Revenue is the product of wafer volume and average selling price per wafer, and the second term rises sharply as mix shifts to advanced nodes. The pricing mechanics are the engine of the current margin story: a leading-edge 300mm wafer at N3 sells for roughly US$18,000–20,000 and N2 for approximately US$30,000, against a few thousand dollars at mature nodes. Each node migration is therefore simultaneously a volume event and a steep ASP step-up. Two additional levers were confirmed this quarter. TSMC has completed customer discussions on 2027 base-price increases of 5–10% across both advanced and mature production (Nikkei Asia, reported by Reuters) — the first documented across-the-board list-price action in this cycle. And the company continues to collect capacity-reservation deposits and customer prepayments that de-risk the capital build, a structural working-capital advantage that funds part of the fab arms race on customers’ balance sheets.

The mix inversion is now complete and accelerating. In FY2023, High-Performance Computing was 43% of revenue and smartphone 38%. In FY2025 the ratio was 58% / 29%. In Q2 2026, HPC reached 66% of revenue — growing 20% sequentially — while smartphone fell 4% sequentially to 22%. IoT is 5%, automotive 4% (+15% sequentially), digital consumer electronics 1%. TSMC is no longer a company whose demand cycle is set by the smartphone replacement cycle; it is a company whose demand cycle is set by aggregate hyperscaler AI capital expenditure. That is a higher-growth demand base and a more capex-cycle-sensitive one, and it is the single most important structural fact about the business today.

Technology mix. In Q2 2026, 2nm contributed 3% of wafer revenue in its first meaningful quarter; 3nm was 30%, 5nm 33%, 7nm 11%. Advanced technologies — 7nm and below — were 77% of wafer revenue, up from 74% for FY2025 and 58% in FY2023. Mature nodes (16nm and above) are a shrinking-share but stable-absolute 23% tail, much of it differentiated specialty work. N2 monthly capacity is scaling from roughly 50,000–60,000 wafers in the first half of 2026 toward 100,000–140,000 by year-end; A16, which brings Super Power Rail backside power delivery, goes to mass production in the second half of 2026.

The customer base: deeper captivity inside rising concentration. North America was 75% of FY2025 revenue and remains the dominant geography; China is below 10%. Top-ten customers were 78% of FY2025 revenue, with Apple at 19% and Nvidia at 17%. The apparent paradox — concentration rising while the largest customer’s share falls — is explained by the custom-ASIC cohort: Google (TPU), Amazon (Trainium/Inferentia), Microsoft (Maia) and Meta (MTIA) now design their own AI silicon and fabricate it at TSMC. Asked directly on the Q2 call whether top-five concentration at a record level was a risk, C.C. Wei was dismissive — “that’s not our concern… there’s a lot of new player in the AI industry.” The honest reading is that concentration has intensified, and that the correct interpretation of it is captivity rather than fragility only for as long as no credible second source exists at the leading edge. A new demand vector was flagged on the call: agentic AI is driving a resurgence of CPU demand in AI datacentres, and whether the winning architecture is x86, Arm or RISC-V, “they are almost all TSMC’s customers” — a genuine architecture-agnostic call option.

Recurring versus 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 with switching costs measured in tens of millions of dollars per tape-out. At the volume level it remains cyclical — the FY2023 revenue decline of 4.5% is the reminder — and the current AI boom is masking that cyclicality rather than abolishing it. Management said as much: demand is strong “all the way to probably 2029, 2030… Whether in between there’s a dip or not, I’m not very sure.”

Verdict (Business Overview): A structurally recurring, genuinely differentiated franchise whose mix has inverted decisively toward its highest-value nodes and highest-growth end market, with a newly-broadened demand vector in agentic-AI CPUs. The offsetting facts are extreme and increasing customer and geographic concentration, and an end-market cyclicality that the AI cycle currently conceals.

3. Industry Dynamics

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 and IP licensors — Synopsys and Cadence, a design-software duopoly with over 70% share, and Arm, the dominant instruction-set architecture. Next, fabless designers capture brand and architectural value but own no fabs. They hand designs to foundries, which convert designs into silicon. Wafers pass to OSAT and advanced-packaging providers — ASE, Amkor, and critically TSMC’s own CoWoS and SoIC lines. Underpinning everything are the equipment makers: ASML in lithography, plus Applied Materials, Lam Research, Tokyo Electron and KLA.

Profit concentrates at two tollbooths. The first is ASML, sole supplier of EUV and High-NA EUV; no chip below roughly 7nm exists without its machines, which gives ASML monopoly pricing and gates total industry leading-edge supply. The second is TSMC. Everyone in between is captive to varying degrees: fabless designers cannot manufacture, and cannot cheaply switch foundries because re-taping-out a chip against a different process design kit is a multi-quarter, multi-million-dollar engineering effort carrying yield risk. TSMC’s 67.7% gross and 60.3% operating margins are the financial fingerprint of that position.

Market size, growth and share. The tailwind remains exceptional. The global foundry market reached a record US$47.95 billion in Q1 2026, and TSMC’s Q1 revenue of US$35.9 billion (+40.6% year-on-year) took its share to 72.3%, up from 70.4% in Q4 2025 (TrendForce). The rest of the table is a study in distance: Samsung 6.5%, SMIC 5.1%, UMC 3.9%, GlobalFoundries 3.3%, HuaHong 2.5%, then Tower, Nexchip, VIS and PSMC at 0.8% each. Intel Foundry does not appear in the top ten. The blended figure understates the leading-edge position, where TSMC’s share is estimated above 90%. Apply Greenwald’s market-share-stability test — a swing of under roughly two points over five to eight years signals formidable barriers — and TSMC does not merely pass it; the test is inverted, because share has risen by roughly two points in a single year at a 72% base. That is a dominant-firm regime, and the barriers are escalating rather than eroding: a single leading-edge fab now costs US$20–30 billion; EUV and High-NA access is rationed by ASML at roughly US$350 million per High-NA tool; the cumulative-volume yield learning curve compounds; process talent is scarce; and ecosystem lock-in deepens with every ported IP block.

The capital cycle: boom-phase deployment across the entire chain. Apply Marathon’s supply-side lens and the picture has changed materially since June, and not in TSMC’s favour. At the leading edge the normal capital cycle has historically been partly suspended — record capital chasing high returns, but mean-reversion blunted because only one rational actor could profitably build leading-edge capacity, ASML throughput capped total supply, and AI demand surprised to the upside. That structure is now being tested from several directions simultaneously. TSMC itself is spending US$60–64 billion in 2026 (35–38% of revenue) with “even more significantly higher” guided for the following three years. Intel is adding High-NA capacity and has, on third-party evidence, materially improved 18A yields. ASML announced an EUV capacity expansion for 2028. And across the memory chain, Samsung, SK hynix, Micron and CXMT have jointly put a large 2028–29 supply window on the construction calendar. The honest Marathon reading: capital is being deployed across the entire semiconductor supply chain at once, at the point of maximum demand confidence, and 2028–29 is where that supply arrives. TSMC’s protection remains that no rival converts capex into competitive leading-edge yield — but that protection is measurably thinner than it was in June.

A new industry mechanism: memory as a cost input, not a peer. DRAM prices are up more than 300% year-on-year as the big three shifted over 80% of capacity to higher-margin HBM, sharply contracting conventional DRAM supply. SK hynix printed a 76.3% operating margin in Q2 2026 — a figure C.C. Wei referenced on his own call with undisguised envy. The consequence for TSMC is not competitive but demand-side: bill-of-materials inflation is crowding out non-AI silicon budgets. Wei was explicit — “we observe consumer and the price-sensitive end market segment are being challenged due to the impact of rising component prices and macroeconomic uncertainties. As such, we are being prudent in our business planning.” This is the first genuinely new industry mechanism since the prior report, and it is a headwind to the roughly one-third of revenue that is not HPC. TSMC’s structural answer is the Winbond arrangement announced in late June: Winbond supplies DRAM wafers built on its CUBE 3D-cache architecture, which TSMC integrates into its wafer-on-wafer and SoIC hybrid-bonding stacks. This is not an entry into merchant DRAM; it is vertical insurance against the HBM oligopoly’s pricing power and an extension of the packaging franchise up the memory-logic interface.

Regulation and geopolitics — now the dominant structural variable, and newly two-sided. Two policy changes since June cut in opposite directions. Section 232 tariffs of 25% on a narrow set of advanced logic semiconductors took effect on 15 January 2026 — and on the same day the United States and Taiwan concluded a trade agreement giving Taiwanese chipmakers that expand US production a reduced tariff, or outright exclusion, on semiconductors and related equipment. That agreement is the economic mechanism that makes the Arizona programme rational: TSMC is purchasing tariff exemption and customer-mandated supply security with gross margin. It reframes the 2–4 percentage points of overseas-fab dilution from value destruction into the premium on an insurance policy — a materially more favourable reading than the prior report’s. Cutting the other way, the underlying structural risk is undiminished: the overwhelming majority of TSMC’s leading-edge capacity sits on one island within PRC missile range, and the factor model puts a +0.93 Taiwan country loading on the stock, nearly equal to its market beta. Separately, US export policy toward China loosened, with case-by-case licensing replacing the blanket presumption of denial for AI-chip exports, subject to end-use certification and an approximate one-million-unit H200 volume cap — a modest positive for volumes and a reduction in the tail risk framed in June.

Mature nodes remain structurally poor and bifurcated. Only two mature segments are tight, and both are AI-derived: power management ICs (0.18µm-class, required in volume by AI datacentres) and CMOS image sensors. Everything else, per Wei, is “not so much of, say, in a lot of shortage. Not at all.” Chinese state-subsidised mature capacity continues to expand with no market-clearing discipline, because subsidies and debt forgiveness break the capital cycle that would normally punish overbuilding. TSMC’s exposure is a minority of revenue and largely specialty, but the overseas specialty fabs — JASM in Japan for image sensors, ESMC in Germany for automotive and industrial — push it into the contested arena at an unattractive point in that sub-cycle.

Verdict (Industry): Structurally 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 — where TSMC sits almost uncontested across 77% of wafer revenue. Structurally poor at the trailing edge, where Chinese state capacity keeps the capital cycle broken. The change since June is that the leading edge’s capital-cycle protection is thinner: capital is now being deployed simultaneously across foundry, lithography and memory into a 2028–29 supply window, and the industry’s own cost structure — memory — has become a demand headwind for everything that is not AI.

4. Competitive Position

Define the relevant market first (Greenwald Step 1). Not “semiconductors” broadly but leading-edge (7nm-and-below) logic foundry — the arena where fixed costs stay fixed and scale compounds. Getting this right matters: TSMC’s dominance is overwhelming here and merely solid at mature nodes, and conflating the two produces a systematically wrong read of the moat.

The moat, named: economies of scale fused with customer captivity — Greenwald’s strongest and most durable category — reinforced by a genuine process advantage, an advanced-packaging chokehold, and an ecosystem network effect. Each layer, pressure-tested against the financial outcome it protects:

1. Economies of scale plus capital intensity (the load-bearing layer). A leading-edge fab costs US$20–30 billion; TSMC’s FY2026 capex alone is guided at US$60–64 billion. The fixed-cost base of R&D, EUV tooling and fab construction is amortised over a wafer volume no competitor approaches, so TSMC’s per-wafer leading-edge cost is structurally the lowest. The confirming evidence this quarter is the 90% incremental operating margin: nine-tenths of every incremental revenue dollar reached operating income, which is only possible when the incremental cost of serving demand is near zero relative to price. Falsification test: remove the scale and TSMC cannot fund the R&D and capex arms race. This is the moat that, if broken, breaks the 67.7% gross margin and the 40.3% ROE. It remains firmly intact.

2. Process-technology and yield leadership (genuine, secondary, and — this quarter — narrowing). This is where the update bites. As of the June report, N2 was running at roughly 65–70% yield against Intel 18A at 50–55% and Samsung SF2 in the mid-50s: a lead of 10–15 points. KeyBanc now reports Intel 18A yields above 85%, up from roughly 65% in a single quarter, against TSMC N2 at approximately 90% — a lead compressed to roughly five points. In the same window ASML confirmed Intel Foundry as the first company to ship a high-volume logic product using High-NA EUV (Panther Lake). TechInsights’ independent comparison scores Intel 18A highest on performance (2.53) ahead of TSMC N2 (2.27) and Samsung SF2 (2.19), while N2 leads decisively on density at roughly 313 million transistors per mm² in high-density cells against 18A’s 238 million. Samsung SF2 remains at 50–60% and far from merchant commercial scale. Three qualifications matter before drawing conclusions. First, the 85% figure is a sell-side estimate, not an Intel disclosure, and it arrived in a week when Intel’s stock needed good news; it is PLAUSIBLE, not CONFIRMED. Second, the reported customer list — Nvidia, Apple, AMD, Microsoft, Micron, Marvell, OpenAI — comprises unconfirmed design-ins and evaluations, with node, volume and timing all unstated; none of the named companies has confirmed a foundry contract. Third, density and performance are different currencies: an AI accelerator constrained by reticle area and power density values density, which is where N2 wins. Greenwald’s caution nevertheless applies — proprietary-technology advantages erode, because “in the long run everything is a toaster” — and the only thing that has kept TSMC’s technology lead from behaving that way is that it is fused to scale. Falsification test: lose the process lead and the ASP premium erodes, compressing margins; scale and packaging would still hold. Reinforcing, not load-bearing — but reinforcing less than it did in June.

3. Customer switching costs (real and financially significant). Moving a chip to another foundry means redoing the physical design against a different PDK, re-qualifying IP, paying for new advanced mask sets costing tens of millions per tape-out, and absorbing yield-ramp risk. C.C. Wei’s formulation of this on the Q2 call is the best available description of the mechanism: “choosing a technology, ramping it up, is not buying a milk from 7-Eleven… You need to understand the technology. You need to really utilize it using the test chip… and then prepare the capacity and ramp it up. That’s why I say it takes about five years.” The behavioural evidence continues to support him, but note the asymmetry: five-year switching lead times cut both ways. If Intel’s 18A/14A engagements are real design-ins today, the wafer revenue consequence would not appear until 2028–29 — which is exactly the window in which the industry’s capacity arrives.

4. Advanced packaging (CoWoS / SoIC / 3DFabric) — and an important change in its character. AI accelerators are gated not by the transistor alone but by TSMC-proprietary advanced packaging, and in June this looked like a second, hardening moat layer. The Q2 call reframed it. Wei: “our packaging capacity is so tight that now it’s limiting my customers’ growth. We welcome that additional flexibility in the market” — referring to Intel’s EMIB-T — “That will help TSMC’s front-end wafer business growth, which is a majority part of TSMC’s business… I welcome have this additional alternative.” This is candid and commercially sensible, and it is also the sound of a chokehold being voluntarily relaxed. Packaging has moved from a competitive weapon that denied rivals AI-accelerator volume to a bottleneck TSMC would rather share so that its higher-value front-end wafers can ship. Prospectively, packaging is a weaker barrier than the June memo credited. Wei’s structural defence — “the front end’s wafer business and the back end’s business are two different things… If they are the same, you can expect ASE become the front-end competitor also” — is correct as far as it goes, and is the reason this does not become a front-end threat.

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 the IP vendors so that designing for TSMC is the path of least resistance. More customers means more ported IP, which lowers the next customer’s design cost and raises rivals’ effective switching cost. Genuine, hard to quantify, and unchanged. It is worth noting that A13 was explicitly designed with design rules backward-compatible with A14 “to ensure smooth IP migration” — the ecosystem effect being deliberately engineered into the roadmap.

Pricing power: the strongest confirmation of the period, with a stated ceiling. Nikkei Asia reported on 21 July, corroborated by Reuters, that TSMC has completed customer discussions on 2027 base-price increases of 5–10% across both advanced and mature production; the ADS rose 5.55% that session. The advanced-node increase is unsurprising in a supply-constrained market. The mature-node increase is the more informative datum: it is a price rise in the segment where TSMC faces subsidised Chinese overcapacity and holds no monopoly, and it succeeded. That is pricing power in the strict sense. Set against this is the most important single sentence management uttered this quarter. Asked why foundry should not earn memory-like returns, Wei said: “We don’t suddenly increase our price by… a 4x or 5x… We earn our value, and we make sure that our profit, our gross margin, is enough for our long-term sustaining expansion… I’m really jealous about memory companies, 86% gross margin. 86%? About 68%, I would be happy about that.” With Q2 at 67.7% and Q3 guided to 66%, management has told the market that the margin is at its target, not on a path to somewhere higher. Two implications follow. The benign one: this is a supplier deliberately not extracting maximum rent from captive customers, which is exactly the behaviour that keeps a franchise durable for decades and keeps regulators and customers from funding alternatives. The unwelcome one: the gross-margin expansion that generated four consecutive quarters of earnings surprise is, by the company’s own account, essentially finished. Forward earnings growth must now come from volume and price, not from margin.

Competitive dynamics versus named rivals.

• Intel Foundry. The most credible threat and materially more credible than in June. 18A yields reportedly above 85%; first to high-volume High-NA logic; highest independent performance score at the 2nm class; 14A targeting risk production in 2027 with a claimed 15–20% performance or 25–35% power improvement over 18A. Against that: not in the top ten by foundry revenue; external revenue was roughly US$174 million a quarter as of the last observation against TSMC’s ~US$40 billion; every reported customer win is unconfirmed. The decisive metric to watch is Intel Foundry’s disclosed external revenue line, which no analyst yield estimate can substitute for.
• Samsung Foundry. 6.5% share; SF2 yields 50–60%, below merchant commercial scale. Enormously profitable in memory — which is precisely why its foundry ambitions are chronically under-prioritised. Not a near-term leading-edge threat.
• SMIC. 5.1% share, third overall, capped near 7nm by EUV export controls. A mature-node and geopolitical concern, not a leading-edge one.
• Rapidus and the state-backed entrants. Real capital, no scale, no ecosystem, no customer trust. Not a 2020s consideration.

Greenwald scorecard. Market-share stability: passes decisively and inverted — share rose from 70.4% to 72.3% in a quarter. ROIC and profitability: passes emphatically — 31.2% ROIC, 40.3% ROE, 60.3% operating margin, 90% incremental operating margin. Cost of entry versus 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; earnings power value sits far above reproduction cost, the signature of a barrier-protected business rather than a merely well-run one.

Verdict (Competitive Position): A durable, multi-layered advantage that remains one of the widest in global equities — but for the first time in this coverage the evidence points to a moat that has stopped widening. The load-bearing scale layer is intact and confirmed by the 90% incremental margin and by rising share. What deteriorated is everything above it: the process lead narrowed from 10–15 points to roughly five on credible third-party evidence; the packaging chokehold is being voluntarily relaxed because it constrains TSMC’s own growth; and management has capped its own margin ambition at roughly the level already achieved. Two honest qualifiers remain: customer concentration is extreme and rising, better read as evidence of captivity than of fragility only while no second leading-edge source exists; and capex at 36% of revenue is boom-phase deployment whose favourable supply side persists only while Intel and Samsung keep failing to convert capital into competitive yield — a condition that is, this quarter, less securely met than it was.

5. Growth History and Forward Opportunities

History. Revenue compounded at roughly 24.5% over the five years to FY2025 (NT$1,587 billion in 2021 to NT$3,809 billion in 2025), but not in a straight line: +42.6% in 2022, −4.5% in 2023 during the PC and smartphone inventory trough that coincided with the N3 ramp, +33.9% in 2024, +31.6% in 2025. The 2023 decline is the discipline in this section — it is the proof that the volume line is cyclical even when the franchise is not. Growth has been almost entirely organic, funded from reinvested cash flow rather than acquisition or equity issuance; the share count has been effectively unchanged for a decade.

The current run-rate is a step change on top of that. Quarterly revenue in NT$ has gone NT$839bn → NT$934bn → NT$990bn → NT$1,046bn → NT$1,134bn → NT$1,270bn across the last six quarters — a 51% increase in five quarters, with sequential growth of 12.0% in the most recent one. June 2026 monthly revenue was NT$442.68 billion, +67.9% year-on-year; first-half 2026 revenue was NT$2,404 billion, +35.6%. Management has raised FY2026 revenue growth guidance to “slightly above 40%” in US dollars, which on FY2025’s US$121.4 billion implies roughly US$170 billion for the year. The composition tells the story: HPC grew 20% sequentially in Q2 alone and now represents 66% of revenue.

Forward opportunity set — five vectors, in descending order of confidence.

1. Node migration (highest confidence). The roadmap is now visible to 2029 and each step is both a volume and an ASP event. N2 entered meaningful contribution at 3% of wafer revenue with monthly capacity scaling toward 100,000–140,000 wafers by end-2026; A16 reaches mass production in 2H 2026; A14 pre-production 2027, volume production 2028, delivering 10–15% speed at iso-power or 25–30% power at iso-speed plus ~20% density against N2; A13 and A12 in 2029, the former achieving a 97% optical shrink for over 6% die-area saving with backward-compatible design rules, the latter bringing Super Power Rail. Wei’s claim that “A14 and its derivative technologies… [will be] an even larger and long-lasting node for TSMC than N2” is a hypothesis, but it rests on disclosed internal metrics — roughly 90% device performance and roughly 90% 256Mb SRAM yield on internal vehicles, with customer tape-outs “ongoing and ahead of schedule.” That last phrase is the most checkable forward statement management made.
2. Advanced packaging (high confidence, capacity-constrained). CoWoS, SoIC and 3DFabric remain sold out to the point of limiting customer growth. The roadmap extends beyond 14x reticle-size CoWoS, and TSMC is building a pilot line for glass substrates and carriers, roughly a year from production maturity. Packaging is 10–20% of capex bundled with test and mask-making — management declined an analyst’s request to break it out separately, which is a small but real disclosure deficiency given how central it has become.
3. Agentic AI and the CPU resurgence (medium-high confidence, newly articulated). Management’s argument is that agentic workloads restore CPU relevance in AI datacentres alongside accelerators, and that TSMC wins regardless of architecture: “no matter what CPU approach is taken, whether it’s x86, Arm-based, or RISC-V, they are almost all TSMC’s customers.” This is a credible architecture-agnostic option on a workload shift, and it is incremental to — not a substitute for — accelerator demand.
4. Silicon photonics / COUPE (medium confidence, early). TSMC’s co-packaged optics platform is in production and ramping; Broadcom and Nvidia are named early adopters. Wei: “the AI data center need to lower down the power consumption and increase the bandwidth of the communication channel. I believe the COUPE will continue to increase the demand, and then will become a very important technology in the next few years.” No revenue quantification was offered. Treat as a real but unsized option.
5. Specialty mature nodes (low-to-medium confidence, narrow). Growth is confined to the AI-adjacent segments — power management ICs and CMOS image sensors — where JASM and ESMC capacity is being added. The commodity mature market is soft.

The honest quality assessment of this growth. It is high quality on every conventional test and concentrated on one. It is organic, funded from operating cash flow, undiluted, priced with power rather than volume-discounted, and it converts to profit at a 90% incremental operating margin. But it is overwhelmingly a single demand vector — AI datacentre capital expenditure by fewer than ten counterparties — and management’s own description of how it validates that demand deserves quoting in full, because it is unusually candid: “I believe every customer tell me the truth. Everyone. You put all the truths together, it’s not the truth… we are checking the AI data centers of progress, the building, the location, the demand, the rest. We checking all that to make sure that TSMC’s chips will not be put in inventory.” That is a CEO stating plainly that his order book is an aggregation of aggressive forecasts requiring judgement to discount. It is exactly the right posture, and it is also an admission that the visibility supporting a US$60–64 billion capital programme is softer than the programme’s size implies.

Verdict (Growth): High-quality growth in every structural respect, at the highest single-customer-cohort concentration in the company’s history. The node roadmap gives genuine visibility to 2029; the ASP and pricing levers are live and demonstrated; the conversion to profit is exceptional. The qualification is that essentially all incremental growth now depends on one capital-expenditure cycle whose participants management itself does not take at face value — and that the non-AI third of the business is being actively squeezed by memory-driven bill-of-materials inflation.

6. Financial Quality

The margin ladder. The clearest way to see what has happened is quarter by quarter (gross / operating / net margin):

Nine points of gross margin in four quarters, on a business that was already the most profitable large-scale manufacturer in the world. The drivers, per management, are utilisation, advanced-node mix, cost-improvement programmes and a favourable exchange rate, partially offset by overseas-fab dilution. The 90.1% incremental operating margin is the number to hold onto: it is what a genuine scale moat looks like on an income statement, and no amount of narrative can substitute for it.

Returns on capital. ROE reached 40.3% in Q2 2026, from 34.5% a year earlier; ROIC 31.2%; return on assets 27.1%; return on capital 34.0%. The sustainable growth rate implied by these returns and the ~24% payout is 33% — meaning TSMC can fund growth of roughly a third a year from retained earnings alone without leverage. These are not merely good numbers for a capital-intensive manufacturer; they are good numbers for a software company.

Cash generation, and the tension at the heart of the financials. Trailing twelve-month figures: revenue US$138.8 billion, EBITDA US$99.4 billion, EBIT US$77.8 billion, net income US$69.3 billion — and free cash flow of US$35.5 billion. That is a 51% conversion of net income into free cash flow, and it is the single most important financial fact in this update. In Q2 alone, operating cash flow of NT$783 billion was met by NT$496 billion of capital expenditure (US$15.7 billion), leaving NT$287 billion of free cash. The company is generating enormous cash and spending roughly two-thirds of it on capacity. The FY2026 capital budget of US$60–64 billion against roughly US$170 billion of revenue is 35–38% capex intensity, up from 33.4% in FY2025, with the CFO stating that the next three years will be “even more significantly higher” than the last three.

There are two defensible readings, and the memo should state both. The favourable one: this is a business earning a 31% ROIC reinvesting at incremental returns far above its cost of capital, and every previous episode of TSMC spending aggressively into a demand cycle has created substantial value. Depreciation, at NT$198.5 billion in Q2 against NT$496 billion of capex, will rise for years and will eventually carry the margin down mechanically — but the revenue that capacity supports is contracted years in advance. The unfavourable one: free cash flow is the number an owner ultimately receives, capex intensity is going the wrong way, part of the increase buys tool-price inflation rather than wafers, and management has declined to give any multi-year number that would let an outside investor test the plan. Both readings are true. The question is which one the price is underwriting, and the answer is unambiguously the first.

Balance sheet. Fortress-grade and improving. At 30 June 2026: cash and marketable securities NT$3,518 billion (approximately US$110 billion); total debt NT$982 billion; net cash NT$2,152 billion, roughly US$67 billion; total assets NT$9,376 billion; equity attributable to shareholders NT$6,433 billion (~US$201 billion, or approximately US$38.8 per ADS). Current ratio 2.46; net debt to equity −33.2%; tangible common equity 68.6% of assets. There is no financing risk in this company at any plausible point in the cycle, and the net cash position materially cushions a downturn — which matters, because the capital programme it is funding is the largest in industrial history.

Working capital — the first genuine yellow flag. Inventory rose to NT$385.5 billion from NT$311.5 billion at Q1 and NT$288.1 billion at Q4 2025 — up 34% in two quarters. Inventory days rose seven to 87; accounts receivable days rose three to 29; the cash conversion cycle lengthened from 26.1 to 36.3 days. Management’s explanation is the N2 ramp, and it is credible: a new node builds work-in-progress before it builds revenue, and N2 went from nothing to 3% of wafer revenue in the quarter. That is the benign reading and it is probably right. It is nonetheless the first inventory build of this cycle, and a second consecutive quarterly rise without a matching revenue step would be a different signal entirely. It belongs on the watch list.

Quality of earnings — clean, with two aggregator defects worth recording. The Q2 result required no adjustment: no one-time gains, no unusual tax benefit (effective rate 18.0%, in line), no capitalisation games. Notably, non-operating items made the quarter worse, not better — a NT$95.8 billion non-operating loss against NT$28.9 billion in Q1, reflecting FX and valuation movement on a NT$3.5 trillion investment portfolio — which means the pre-tax growth of +74.9% understates the underlying operating improvement of +65.4% at the EBIT line on a cleaner base. Stock-based compensation is a rounding error (NT$114 million in Q1 against NT$572 billion of net income), and the weighted-average share count moved from 25,929 million to 25,932 million year-on-year: effectively zero dilution. Two data-integrity findings, recorded because they would mislead anyone working from the aggregators. First, ROIC.ai maps TSM’s income-tax charge into an “extraordinary item” line (NT$156.1 billion in Q2 2026), producing a phantom one-time item and an inflated continuing-operations EPS of NT$184.45; reconciliation to the 6-K confirms there is no extraordinary item and reported EPS is NT$27.25. Second, ROIC’s share count for TSM (4,676 million) does not correspond to the 25,932 million ordinary shares actually outstanding, so every per-share figure in that feed is wrong. All per-share numbers in this memo are computed from reported net income divided by 25,932 million ordinary shares, multiplied by five for the ADS ratio, and cross-checked to the company’s own disclosed US$4.31 per ADR unit for Q2.

Verdict (Financial Quality): Exceptional, and improving on every operating measure — with the sole and significant exception of free-cash-flow conversion. Economics do improve with scale: the 90% incremental operating margin, the 40% ROE and the nine-point gross-margin expansion settle that question decisively. Earnings quality is high and dilution is nil. The qualifications are that free cash flow converts only 51% of net income because capital intensity is rising into the boom; that the gross margin is at management’s own stated ceiling; and that inventory has begun to build for the first time in this cycle.

7. Capital Allocation

The single question in this section is whether the largest capital programme in the history of manufacturing is being deployed intelligently. The evidence is mostly, but not entirely, favourable, and it is weaker than it was in June.

Capital expenditure. FY2026 capex has been raised twice inside seven months: US$52–56 billion (January) → “closer to US$56 billion” (April) → US$60–64 billion (July). Against FY2025’s US$40.6 billion actual, the midpoint of the current guide is +53% year-on-year. Allocation is 70–80% to advanced process technologies, ~10% to specialty, and 10–20% to advanced packaging, testing, mask-making and other. The CFO’s forward framing was explicit and unbounded: “Last time we said our CapEx in the next three years will be significantly higher than the CapEx in the past three years. The CapEx in the next three years will be even more significantly higher than the past three years.” The CEO added: “You bet that will continue to increase.”

The Arizona commitment. On the Q2 call, TSMC announced an additional US$100 billion of investment in Arizona, taking the announced total to US$265 billion — “probably additional four more fabs,” front- and back-end, for 2nm-and-below logic plus advanced packaging. Separately, 13 leading-edge and advanced-packaging fabs are planned in Taiwan over the next several years. Asked twice for a schedule, management gave none: “If you ask me to give you a firm schedule, no, we don’t have it today, but we do have a plan. We speed it up. We try to speed it up as fast as possible.” The prior report treated the overseas build as a subsidised, customer-funded supply-security investment. That reading is now strengthened by policy: with 25% Section 232 tariffs on advanced logic live since January and a US–Taiwan agreement that reduces or eliminates them for Taiwanese chipmakers expanding US production, the Arizona spend buys a tariff exemption on the majority-North-American revenue base. Guided dilution of 2–3% of gross margin now, widening to 3–4% in later stages, is therefore best understood as the premium on an insurance policy that protects roughly three-quarters of revenue — not as value destruction. That is a materially more favourable framing than in June, and it is the most important upgrade in this section.

Three genuine concerns, stated plainly.

• Part of the increase buys inflation, not capacity. Asked what was driving the raise, Wei answered: “the most important reason is because of demand continue to increase… The second reason is inflation. Now we buy the tools with inflation price.” The proportion was not quantified and cannot be derived from disclosure. Some fraction of the incremental US$8–10 billion therefore produces no incremental wafers. This is a real quality-of-deployment negative and it is invisible in the capex line.
• No multi-year number, and no schedule. Management declined a three-year capex figure (which it did provide in the 2021 cycle), declined to update the five-year AI CAGR beyond “stronger and stronger,” declined a 2027 revenue framing beyond “next few years is going to be a very good business for TSMC,” and declined a schedule for US$100 billion of Arizona spending. Each refusal is individually defensible. Collectively they mean an outside investor cannot test the plan against anything.
• The demand basis is explicitly judgemental. The CEO’s own framing — every customer tells the truth, and the sum of the truths is not the truth — is intellectually honest and operationally correct. It is also the clearest possible statement that a US$60–64 billion annual commitment rests on management judgement about counterparties’ aggressive forecasts.

Shareholder returns. Dividends are rising steadily and conservatively: NT$467 billion paid in 2025 (+28.6% year-on-year), NT$18 per share; NT$24 per share in 2026, +33%, with a further increase signalled for 2027. At NT$32/US$ that is approximately US$3.75 per ADS, a yield of roughly 0.9% at US$418. The trailing payout ratio is about 24%. There are no buybacks — which, at 10.8x book and an all-time-high valuation, is exactly correct; repurchasing stock here would be the single clearest capital-allocation error available to this management, and they are not making it. There is no dilution: the share count is flat to four decimal places, and stock-based compensation is immaterial. There has been no empire-building M&A; the Winbond arrangement is a supply agreement, not an acquisition.

A discipline signal worth weighting. Asked whether TSMC would invest in or finance its customers’ customers, as several peers in the AI complex now do, Wei was unambiguous: “So far, no, TSMC don’t do this kind of financial arrangement because of we think we are working with current customer with the current model smoothly and also successfully.” In a cycle where circular vendor financing is spreading — and where it materially degrades the quality of reported revenue at several large AI-complex companies — refusing it is a meaningful positive. Every dollar of TSMC’s revenue is paid for by a customer with its own money.

Incentive alignment. TSMC’s compensation structure remains conservative by US large-cap standards, with modest equity issuance and no evidence of option-driven behaviour. The insider record adds essentially nothing: across 143 Form 4 filings covering 161 transactions since August 2025, the entire “purchase” (code P) activity totals 31,884 shares across 153 systematic, plan-scale transactions; sales total 200,050 shares, essentially a single 200,000-share VP disposal in May 2026; and the largest gross figures are bona-fide gifts (code G) of 6.17 million disposed against 5.17 million acquired, largely offsetting estate and trust transfers. There is no signal here in either direction. It should also be noted that as a foreign private issuer TSMC’s more meaningful insider disclosure runs through Taiwanese law, so the EDGAR window is incomplete — a reason to weight this evidence lightly rather than to read it as reassurance.

Verdict (Capital Allocation): Still good, and still the right instinct — but resting on faith to a degree it did not a quarter ago. The positives are substantial and specific: no buybacks at a record multiple, no dilution, no empire-building, no circular financing, a rising and well-covered dividend, and an Arizona programme that policy has now converted into a rational purchase of tariff exemption and supply security. Against that, capital intensity is rising from 33% to 36%+ of revenue, an unquantified portion of the increase buys tool inflation rather than wafers, free cash flow converts only 51% of net income, and management has withdrawn every quantitative multi-year anchor an outside investor might use to check the plan. Management has earned the benefit of the doubt across three decades. This is the first quarter in this coverage in which they have asked for quite so much of it.

8. Changes and Headwinds — Last Two Years

The AI mix inversion (2024–2026). The defining change. HPC went from 43% of revenue in FY2023 to 51% in FY2024, 58% in FY2025 and 66% in Q2 2026, while smartphone fell from 38% to 22%. Advanced nodes (≤7nm) went from 58% to 77% of wafer revenue. Revenue grew from US$69 billion (FY2023) to roughly US$170 billion guided for FY2026 — a two-and-a-half-fold increase in three years. This is not a cyclical upswing layered on a stable business; it is a different business.

Margin transformation (2025–2026). Gross margin ran 54.4% (FY2023) → 56.1% (FY2024) → 59.9% (FY2025) → 67.7% in Q2 2026. Operating margin reached 60.3%. This is the change that most explains the share price, and — per management’s “about 68%” remark — it is the change least likely to repeat.

The capital-expenditure escalation (2025–2026). From US$40.6 billion (FY2025) to a guided US$60–64 billion (FY2026), with three further years guided “even more significantly higher.” Announced Arizona investment reached US$265 billion, plus 13 planned fabs in Taiwan. This is the largest sustained corporate capital programme in industrial history and it is being executed by a company with US$67 billion of net cash and no financing constraint.

Trade and industrial policy (2025–2026). The decisive regulatory development is Section 232: a 25% tariff on advanced logic semiconductors effective 15 January 2026, paired the same day with a US–Taiwan agreement reducing or eliminating that tariff for Taiwanese chipmakers expanding US production. Alongside it, US export controls toward China loosened, with case-by-case licensing replacing the blanket presumption of denial and an approximate one-million-unit H200 cap. Net: the environment is materially less hostile than the June report assumed on export controls, and materially more directive on where TSMC builds.

The competitive narrowing (Q3 2026). Intel’s reported 18A yield step from ~65% to above 85% in a single quarter, ASML’s confirmation of Intel as first to high-volume High-NA logic, and TechInsights’ performance ranking of 18A above N2 are, collectively, the first credible narrowing of TSMC’s process lead in a decade. All of it is third-party evidence and none of it has yet appeared in Intel Foundry’s revenue.

Memory-driven cost inflation (2026). DRAM prices up more than 300% year-on-year, with the big three shifting over 80% of capacity to HBM. The consequence for TSMC is that consumer and price-sensitive end markets — roughly a third of revenue — are being squeezed by bill-of-materials inflation, a headwind management named explicitly for the first time this quarter. The response is the Winbond DRAM-wafer arrangement for wafer-on-wafer and SoIC integration: insurance against the HBM oligopoly, not a memory business.

Litigation and operational events. The trade-secrets suit against former Senior Vice President Wei-Jen Lo, filed in Taiwan’s Intellectual Property and Commercial Court in November 2025 after his move to Intel, remains pending with no ruling reported; Intel denies the allegations. In late July 2026 a strong earthquake prompted a precautionary evacuation at the JASM plant in Kumamoto, Japan, with operations gradually resuming — a reminder that geographic diversification substitutes one natural-hazard concentration for another. A Sony sensor joint venture with TSMC was reported in early August 2026, extending the specialty-node relationship.

Verdict (Changes): On balance these developments strengthen the earnings power and weaken the durability premium. The mix inversion, margin transformation and policy resolution are large positives that are already in the numbers and largely in the price. The competitive narrowing, the memory cost channel and the capital escalation are the offsets, and all three point in the same direction: TSMC will earn more over the next two years than the June report assumed, and it will have to work harder for what it earns after that.

9. Risk Analysis

The risks that actually matter. Three of these dominate, and they are not the ones that generate headlines. First, the Taiwan tail. It is the only risk on this list capable of permanently impairing capital rather than merely deferring it, its probability is genuinely unknowable, and the factor model says it is roughly half the stock’s systematic risk. No amount of Arizona capacity changes this within a decade. Second, the capital-cycle risk in 2028–29. The entire semiconductor supply chain is building simultaneously at the point of maximum demand confidence — this is the textbook Marathon setup, and the fact that TSMC is the highest-quality participant does not exempt it from the utilisation consequences of an industry-wide supply arrival. Third, the free-cash-flow conversion risk, which is not a risk to the business at all but is the central risk to the valuation: if capex intensity does not normalise, the reverse-DCF simply cannot be satisfied at any plausible growth rate.

Conversely, two risks are widely discussed and, on the evidence, over-weighted. Customer concentration is a consequence of the moat rather than a threat to it, and will remain so until a second leading-edge source demonstrably exists. And mature-node Chinese competition, while structurally real, affects less than a quarter of wafer revenue in a segment where TSMC just raised prices.

10. Valuation Discussion

Where the security trades. At US$418.20 (2026-08-06), with 25,932 million ordinary shares (5,186 million ADS-equivalents), TSMC carries a market capitalisation of approximately US$2.17 trillion and, net of US$67.2 billion of net cash, an enterprise value of approximately US$2.10 trillion.

A necessary correction to the own-history screen. The AZI valuation index reports TSMC at the 97.9th percentile of its own multi-year valuation history (P/E 97.5th, P/B 98.1st, P/S 98.1st, as of 2026-07-24). Those percentiles are computed on stale fundamentals — a trailing EPS of US$10.73, book value of US$33.24 per ADS and sales of US$23.79 per ADS, against actual trailing figures of approximately US$13.4, US$38.8 and US$26.8. The effect is to overstate the trailing P/E by roughly six turns (37.6x screened against 31.3x actual) and book by roughly 1.3 turns. The direction survives the correction — TSMC does trade near the top of its own range — but the level does not, and the distinction matters precisely because earnings are compounding at 77% year-on-year, which is exactly when stale-denominator screens fail. All figures in the table above are computed from reported statements.

The two true statements, and the tension between them. On book value, TSMC has never been more expensive: 10.8x exceeds every year-end in the ten-year record and every intra-year high, including FY2025’s 8.52x peak, and compares with roughly 2.9x at the 2015 trough. On forward earnings, TSMC is cheaper than it was in June. FY2026 revenue of roughly US$170 billion (guided “slightly above 40%” growth; first-half actual US$76.1 billion, Q3 guide midpoint US$45.2 billion, implied Q4 around US$49 billion) at a blended net margin near 52–53% produces FY2026 estimated EPS of roughly US$17.0–17.5 per ADS, or about 24x. On a FY2027 growing at management’s roughly 25% target with a stable margin, EPS lands near US$21–23, or roughly 18–20x. Trailing EPS has risen approximately 28% since the June report while the price has risen 0.9%. The multiple was not paid for; it was grown into. That is a genuine improvement in the setup and it is the reason this update raises the accumulation zones rather than lowering them.

Embedded expectations — the reverse-DCF, stated as an identity rather than a forecast. At a 9% cost of capital and 3% terminal growth, a US$2.10 trillion enterprise value requires steady-state free cash flow of approximately US$126 billion. Trailing free cash flow is US$35.5 billion. Closing that gap requires two things to happen together, not one:

• Revenue must roughly double again from US$139 billion trailing toward US$280–300 billion; and
• Capex intensity must normalise from today’s 35–38% of revenue back toward roughly 20%. To size the second requirement: US$300 billion of revenue at a 50% operating-cash-flow margin generates US$150 billion of operating cash flow; at 20% capex intensity that leaves US$90 billion of free cash flow — still short of the US$126 billion the current price requires. The arithmetic only closes on a combination of higher revenue, sustained high-60s margins and a capex normalisation that management has explicitly refused to promise (“the CapEx in the next three years will be even more significantly higher”).

What the market is pricing correctly. The AI demand curve through roughly 2028 — this is contracted, visible and corroborated by the order book, the packaging constraint and hyperscaler capital plans. The node roadmap: N2 ramping, A16 to mass production in 2H26, A14 to volume in 2028 with disclosed internal yield metrics and tape-outs ahead of schedule, A13 and A12 in 2029. The pricing power, now documented at 5–10% across the portfolio for 2027. The near-monopoly share, rising to 72.3%. None of this is in dispute and none of it is where the risk lies.

What the market may be pricing incorrectly. Three things. That capex intensity ever normalises — the reverse-DCF requires it, management has guided the opposite for three years, and part of the spending buys tool inflation. That gross margin goes higher — the CEO named “about 68%” as the level he would be happy with, and Q3 is guided to 66%; the market has extrapolated a four-quarter margin expansion that its author has said is finished. That Taiwan concentration deserves no discount — the stock carries a +0.93 country loading and its nearest empirical peers are Taiwan index funds, yet it trades at a premium to every foundry comparable and, on forward earnings, at a discount only to ASML.

Scenario framework (four years, to FY2030; assumptions explicit; no price target).

The asymmetry is worth stating plainly: the bull case requires management to hit its own target and then do something it has explicitly declined to promise (normalise capex), while the bear case requires only that an industry-wide capacity build meets a demand pause. That is not an argument that the bear case is likely — the demand evidence is strong and the CEO is actively verifying datacentre construction to avoid channel inventory. It is an argument that the distribution is wider than a 24x forward multiple on a 40%-growth business intuitively suggests.

Comparables — read with a caveat. Within the AI-semiconductor complex, TSMC’s roughly 24x forward earnings remains the lowest multiple among the structural winners: Broadcom trades at approximately 49x EV/EBITDA and the 90th percentile of its own ten-year range; ASML at a materially higher forward P/E; Intel at the 99th percentile of its own price-to-book and price-to-sales history while still losing money on a GAAP basis. On that basis TSMC is the cheapest high-quality asset in its own value chain. The caveat is the one the factor model supplies: TSMC’s nearest empirical comparables are ASE Industrial (0.861 similarity), the iShares MSCI Taiwan ETF (0.849) and the Franklin FTSE Taiwan ETF (0.843) — Taiwan country exposure before foundry exposure. A comp set that ignores that is measuring the wrong thing.

Verdict (Valuation): No margin of safety in the multiple, but a better setup than in June — because the earnings, not the price, did the work. At 10.8x book TSMC has never been more expensive; at 24x FY2026 estimated earnings it is cheaper than it was two months ago. The reverse-DCF requires free cash flow to more than triple, which requires a capex normalisation management will not commit to. This memo takes no position; it observes that the entire investment question reduces to whether an investor is willing to underwrite a decade of compounding at a price that leaves no room for a mistake.

11. Variant Perception

What consensus believes. That TSMC is the indispensable, un-substitutable enabler of the AI build-out; that its share and pricing power are unassailable; that a roughly 24x forward multiple is undemanding for a business growing 40% with 60% operating margins; and that the June–July drawdown was a sector-sentiment event rather than a company event. The sell-side commentary through the period is remarkably homogeneous — variations on “buy the dip,” with the sceptical pieces framed as valuation discomfort rather than thesis disagreement. Consensus is, in short, correct about the business and unreflective about the price.

The strongest bull case. TSMC is the closest thing in global equities to a toll on computation itself, and the toll is rising. Share went up to 72.3% in a quarter. Gross margin went up nine points in four quarters. Pricing was raised 5–10% across the entire portfolio for 2027, including in the segment where it has no monopoly. Ninety per cent of incremental revenue reaches operating income. The roadmap is visible to 2029 with A14 tape-outs ahead of schedule. Demand exceeds supply by a gap the CEO will only describe as “very big.” A new demand vector — agentic-AI CPUs — arrived this quarter and TSMC wins it regardless of which architecture prevails. And the balance sheet holds US$67 billion of net cash against zero financing risk. Meanwhile the multiple is the lowest in the leading-edge complex and the stock trades 12% below its high with a Momentum factor loading of just +0.16, meaning no crowding premium is being paid. If AI capital expenditure compounds through 2029 as management believes, today’s price will look unremarkable.

The strongest bear case. The stock requires free cash flow to more than triple, from US$35.5 billion to roughly US$126 billion, while capital intensity is moving the wrong way — 33.4% to 36%+ of revenue, with three more years guided “even more significantly higher” and some unquantified portion of that buying tool inflation rather than wafers. The margin engine that produced the earnings surprise is at its own management’s stated ceiling. The process lead that justifies the ASP premium has narrowed from 10–15 points to roughly five on credible third-party evidence, and Intel has a first-mover position on High-NA that TSMC has declined. The packaging chokehold is being voluntarily relaxed. Inventory rose 34% in two quarters. The entire supply chain — foundry, lithography, memory — is building simultaneously into a 2028–29 window, which is the classic Marathon setup for a capital-cycle bust. And roughly half the systematic risk is a Taiwan country factor for which the price contains no discount. On book value, this is the most expensive TSMC has ever been.

The three to five assumptions that actually matter.

1. That AI datacentre capital expenditure compounds through 2029 rather than pausing in 2028. Everything else is second-order. Falsified by: hyperscaler capex guidance cuts, CoWoS order cancellations, or two consecutive quarters of TSMC inventory build without a revenue step.
2. That capex intensity eventually normalises toward 20% of revenue. The valuation requires it; management has guided against it. Falsified by: a fourth consecutive capex raise, or FY2027 capex guided above US$70 billion. Confirmed by: the first year in which capex grows slower than revenue.
3. That the process lead holds well enough to sustain high-60s gross margins. Falsified by: a disclosed, multi-billion-dollar leading-edge wafer commitment from a top-five fabless customer to Intel or Samsung — the revenue consequence, not another yield estimate. Confirmed by: Intel Foundry external revenue failing to inflect over the next two prints while TSMC’s 2027 price increases stick.
4. That the 2027 5–10% price increases actually hold. They were reported by trade press, not disclosed by the company, and they are the mechanism that offsets N2 and overseas dilution. Falsified by: a Q4 or Q1 gross-margin guide below 64% without a utilisation explanation.
5. That Taiwan-Strait risk stays in the tail. Unfalsifiable in advance, which is why it is a position-sizing question rather than a thesis question.

Where the tape says consensus may be offside. The factor-positioning evidence is genuinely useful here, and it points in two directions at once. On one hand it argues against the most common bearish framing — that TSMC is a crowded momentum trade due for a violent unwind. It is not: the Momentum loading is +0.16 in the base model despite +81.6% twelve-month relative strength, price sits above a rising 200-day EMA, and idiosyncratic volatility is a moderate 20.8%. This is a high-beta growth vehicle in an ordinary correction, not a positioning bubble. On the other hand it argues against the most common bullish framing — that TSMC is a diversified global technology franchise. The Taiwan country loading of +0.93 is nearly as large as the market beta, and the empirically nearest securities are Taiwan index funds before they are foundry or semicap peers. The market is, whether investors realise it or not, trading TSM substantially as a leveraged Taiwan asset. That is the variant perception: not that the business is misunderstood, but that the risk is mis-attributed. Investors underwriting a semiconductor thesis are, in factor terms, half-underwriting a geopolitical one — and being paid a premium multiple rather than a discount for it.

12. Fact vs. Interpretation

13. Open Questions

1. What is Intel Foundry’s external revenue? Every other question about the competitive narrowing is downstream of this one. Yield estimates and unconfirmed design-ins do not pay for fabs; disclosed wafer revenue does. This is the single highest-value unknown in the report.
2. What proportion of the US$8–10 billion capex raise buys tool-price inflation rather than incremental capacity? Management named inflation as a driver and did not size it. It is not derivable from disclosure, and it directly determines the return on the incremental spend.
3. What is the multi-year capex path? Management provided a three-year figure in the 2021 cycle and declined to this time. Without it, the capex-normalisation assumption on which the valuation depends cannot be tested at all.
4. Do the reported 2027 price increases of 5–10% actually hold, and on what volume base? The source is trade press, not the company. This is the mechanism that offsets N2 and overseas dilution, and its confirmation or failure will show up in the Q4 2026 and Q1 2027 gross-margin guides.
5. Is the inventory build entirely the N2 ramp? Management’s explanation is credible and probably right. A second consecutive quarterly increase without a matching revenue step would mean something else.
6. How much advanced-packaging capacity is TSMC willing to concede to Intel’s EMIB-T, and at what margin cost? The CEO welcomed the relief. Whether that is a temporary accommodation of a shortage or a durable narrowing of the packaging moat is unresolved.
7. What are the economics of the Winbond DRAM arrangement? Volumes, pricing, exclusivity and capital commitment are all undisclosed. Strategically sensible; financially unquantifiable.
8. How does the US–Taiwan tariff exemption behave under a change of administration or a deterioration in the relationship? The Arizona economics depend on a policy that can be re-tightened.
9. What is the resolution and exposure of the Wei-Jen Lo trade-secrets suit? Pending with no ruling. Financially immaterial in likely outcomes; informationally significant about the mobility of process knowledge.
10. What does A14 cost to build, and does TSMC ultimately adopt High-NA? Management is explicitly weighing tool cost against maturity. If High-NA proves decisive at A14 and below, Intel’s early adoption is a genuine option; if not, TSMC’s deliberation saves billions.

14. What Must Be True

For the bull case

1. AI datacentre capital expenditure must compound, not pause, through 2029. TSMC’s demand is now 66% HPC and rising, sourced from fewer than ten counterparties. Falsification test: two consecutive quarters in which aggregate hyperscaler capital-expenditure guidance is cut, or CoWoS/advanced-packaging order reductions, or two consecutive quarters of TSMC inventory build without a matching revenue step. Any one of these breaks it.
2. Gross margin must hold in the mid-to-high 60s despite N2 and overseas dilution. The 2027 price increases must offset the guided 3–4 points of N2 dilution and the widening 3–4 points of overseas dilution. Falsification test: a quarterly gross-margin guide below 64% that management does not attribute to utilisation. That would mean the pricing did not stick.
3. Capital intensity must eventually fall. The valuation requires free cash flow to more than triple; that arithmetic does not close at 36% capex intensity at any plausible revenue level. Falsification test: FY2027 capex guided above US$70 billion, or a fourth consecutive raise, or any year in which capex grows faster than revenue for a third consecutive time.
4. The process lead must remain sufficient to defend the ASP premium. N2 must stay ahead in the metrics customers actually buy, and A14 must arrive on the disclosed 2028 schedule. Falsification test: A14 volume production slipping beyond 2028, or a disclosed multi-billion-dollar leading-edge wafer commitment from a top-five fabless customer to a competitor.

For the bear case

1. The industry-wide 2028–29 capacity build must meet a demand pause. Foundry, lithography and memory are all constructing into the same window. Falsification test: leading-edge utilisation staying above 90% through 2028 with TSMC still supply-constrained. If the gap the CEO calls “very big” is still big in 2028, the bear case has no mechanism.
2. Intel or Samsung must convert process progress into disclosed leading-edge revenue. Yield estimates are not a competitive event; revenue is. Falsification test: Intel Foundry external revenue remaining below roughly US$1 billion a quarter through 2027 while TSMC’s share holds above 70%. That would confirm the narrowing is technical rather than commercial.
3. The gross margin must revert materially below the high 60s. The bear valuation requires margin compression, not merely growth deceleration. Falsification test: four consecutive quarters of gross margin at or above 66% with N2 at scale. That would demonstrate the dilution has been absorbed and the ceiling is a floor.
4. The Taiwan discount must eventually be demanded by the market. Today none is applied. Falsification test: unfalsifiable in advance — which is precisely why it belongs in position sizing rather than in the thesis.

15. Source Appendix

See Appendix B — Source Appendix below for the full, dated source list with URLs, and Appendix A — Standard Diligence Questionnaire for the supplemental diligence work.

APPENDIX A — Standard Diligence Questionnaire

Taiwan Semiconductor Manufacturing Company Limited (NYSE: TSM / TWSE: 2330) · As of August 7, 2026 Supplemental to the analysis above. Labels: FACT / INTERPRETATION / ASSUMPTION.

General

What thoughtful questions have other investors asked about this company?

The Q2 2026 call is the best available window into what sophisticated investors actually want to know, and the questions clustered into five themes — every one of which management declined to answer numerically.

• Capital-expenditure visibility (Sunny Lin, UBS; Haas Liu, BofA). “Back in the COVID super cycle, TSMC did provide a three-year CapEx outlook… will it be possible for you to share any color?” Answer: no number, only “even more significantly higher.” INTERPRETATION: the buy-side has correctly identified that the capex path, not the demand path, is now the swing variable in the valuation — and that management has withdrawn the disclosure it provided in the last cycle.
• Competitive threat from Samsung and Intel (Charlie Chan, Morgan Stanley). “Several U.S. companies are engaging with those industry peers… Would TSMC worry that your competitors take more slots?” Answer: the 7-Eleven analogy — “choosing a technology, ramping it up, is not buying a milk from 7-Eleven.”
• Customer concentration (Jim Fontanelli, Arete). “Your exposure to your top five customers is becoming meaningfully larger than at any point in your history.” Answer: “No, that’s not our concern.” INTERPRETATION: the question was better than the answer.
• Vendor/customer financing (Fontanelli again). “We’re seeing your direct customers put capital into both financing and investing in AI demand. Is that something TSMC is considering?” Answer: a clear no. INTERPRETATION: this is the most valuable exchange on the call — an investor probing for circular-financing contamination and getting an unambiguous denial.
• Pricing and the memory comparison (Gokul Hariharan). “You are no longer the most profitable semiconductor manufacturing company at this point in time.” This drew the “about 68%, I would be happy about that” answer — the most consequential sentence management spoke this year.

Two further questions recur in the published sell-side and buy-side commentary: whether the demand-supply gap is 30–50% or larger (Robert Sanders, Deutsche Bank — answer: “the gap is very big,” no number), and whether advanced-packaging capex should be broken out separately from testing and mask-making (Evelyn Yu, Goldman — answer: no).

Cyclicality & Earnings Nature

Are earnings at a cyclical high or low? FACT: on every conventional measure, a high. Gross margin 67.7% against a five-year range of 51.6%–67.7%; operating margin 60.3%; ROE 40.3%; utilisation at levels management describes as supply-constrained with a “very big” demand-supply gap. INTERPRETATION: these are peak-cycle economics. The nuance is that the structural level has genuinely shifted upward — the mix is 77% advanced nodes against 58% three years ago, and advanced nodes carry structurally higher margins — so the correct statement is “a cyclical high on a permanently higher base,” not “a cyclical high due to revert to the old mean.”

Driven by the external environment or internal actions? Both, and separably. External: the AI capital-expenditure cycle, which TSMC did not create. Internal: the node roadmap that captured it (N3 → N2 → A16 → A14), the CoWoS packaging investment that made accelerator volume possible, and the pricing decisions — 2027 base prices reportedly up 5–10%. INTERPRETATION: roughly, the external environment set the volume and the internal actions set the margin. The nine-point gross-margin expansion in four quarters is disproportionately an internal-actions story (mix, utilisation, cost programmes, FX), which is why it is more fragile than the volume story — mix improvement runs out when advanced nodes reach ~100% of wafer revenue.

How stable are revenues? FACT: highly stable at the franchise level, cyclical at the volume level. FY2023 revenue fell 4.5% during the inventory correction; monthly revenue is disclosed and currently running +67.9% year-on-year (June 2026). INTERPRETATION: design wins lock in for multi-year product cycles with tape-out switching costs in the tens of millions, so revenue sources are extremely sticky; revenue quantities follow end-market demand with a lag.

Outlook for products/services? FACT: N2 at 3% of wafer revenue and scaling toward 100,000–140,000 wafers per month by year-end 2026; A16 to mass production 2H 2026; A14 pre-production 2027 and volume 2028 (10–15% speed at iso-power or 25–30% power at iso-speed, ~20% density over N2); A13 and A12 in 2029. Advanced packaging sold out. COUPE silicon photonics in early production with Broadcom and Nvidia as named adopters.

How big will this market be — growing, shrinking, domestic or international? FACT: the global foundry market reached a record US$47.95 billion in Q1 2026, of which TSMC took 72.3%. Revenue is overwhelmingly international by customer domicile — North America 75%, China below 10% — and overwhelmingly Taiwanese by manufacturing location, though the Arizona (US$265 billion announced), Japan and Germany programmes are shifting that at the margin. INTERPRETATION: the addressable market is growing at rates without precedent for a capital-intensive industry, and the growth is concentrated precisely where TSMC’s share is highest.

Business Quality & Competitive Moat

Is the industry getting more or less competitive? INTERPRETATION: less competitive by share, marginally more competitive by technology. TSMC’s share rose from 70.4% to 72.3% in a quarter — the opposite of intensifying competition. But Intel’s reported 18A yield improvement from ~65% to above 85%, its first-mover position on High-NA EUV, and TechInsights scoring 18A above N2 on performance represent the first credible technical narrowing in a decade. The reconciliation: technical parity is necessary but not sufficient: converting it into wafer revenue takes five years and customer trust that Intel has not yet demonstrated.

How profitable is the business (ROIC, ROE)? FACT: ROE 40.3%, ROIC 31.2%, ROA 27.1%, return on capital 34.0% (Q2 2026). Operating margin 60.3%; incremental operating margin 90.1%. Implied sustainable growth rate 33%.

How profitable is the industry — how many competitors, what barriers to entry? FACT: the industry is bimodally profitable. TSMC earns 60% operating margins on 72.3% share; Samsung Foundry, SMIC, UMC and GlobalFoundries divide 18.8% at far lower margins; Intel Foundry loses billions per quarter. Barriers: US$20–30 billion per leading-edge fab; ASML-rationed EUV/High-NA access at ~US$350 million per High-NA tool; a cumulative-volume yield learning curve; scarce process talent; and a co-developed EDA/IP ecosystem. INTERPRETATION: this is a natural-monopoly structure at the leading edge and a commodity structure at the trailing edge, and TSMC has arranged its mix (77% advanced) to sit almost entirely in the first.

Can the business be easily understood? INTERPRETATION: yes, at the level that matters. Revenue equals wafers times price; price rises with node advancement; margin rises with utilisation and advanced-node mix and falls with new-node ramps and overseas fabs. Every one of those variables is disclosed quarterly. The difficulty is not comprehension but forecasting — specifically forecasting the capital-expenditure path and the AI demand curve, neither of which management will quantify beyond a year.

Can it be undermined by foreign low-cost labour? FACT/INTERPRETATION: no. Leading-edge fabrication is capital- and knowledge-intensive, not labour-intensive; labour is a small fraction of wafer cost. The relevant substitution threat is state-subsidised capital, not cheap labour — and it applies to mature nodes, where Chinese capacity expands without market-clearing discipline, not to the leading edge, where EUV export controls cap SMIC near 7nm.

Do brands matter? INTERPRETATION: not in the consumer sense — TSMC’s name appears on no product a consumer buys. But “TSMC” functions as a reliability and neutrality guarantee to a fabless customer betting a product cycle on a process, and C.C. Wei explicitly names “customer trust” as one of the three fundamentals alongside technology and manufacturing. That is a brand in the B2B sense, and it is the asset Intel most conspicuously lacks.

What is the nature of competition? INTERPRETATION: competition on process performance, yield, capacity availability and packaging — not on price. The evidence is that TSMC just raised prices 5–10% across its entire portfolio for 2027, including mature nodes where it holds no monopoly, and reportedly maintains a no-discount posture on N2 at roughly US$30,000 per wafer. A market in which the leader raises prices portfolio-wide is not price-competitive.

Customers’ switching costs? FACT: re-taping-out to a different foundry means redoing physical design against a different PDK, re-qualifying IP, buying new advanced mask sets at tens of millions of dollars each, and absorbing yield-ramp risk. C.C. Wei sizes the total process from evaluation to volume at about five years. INTERPRETATION: this is genuinely high and financially quantifiable, and it is the reason a technical narrowing at Intel does not translate into near-term revenue loss. It is also the reason that if Intel is winning design-ins today, the consequence would appear in 2028–29.

Financial Condition & Balance Sheet

Assets not fully recognised on the balance sheet? INTERPRETATION: substantially. The process-technology knowledge base, the accumulated yield learning curve, the customer-qualified PDK/IP ecosystem and the 534-customer relationship set are carried at approximately zero. R&D is expensed — NT$73.1 billion in Q2 2026 alone, roughly 5.8% of revenue — so the entire cumulative investment in process leadership sits nowhere on the balance sheet. In Greenwald’s terms, earnings power value materially exceeds asset reproduction value, which is the signature of a genuine barrier.

Off-balance-sheet liabilities? FACT: the primary off-balance-sheet commitment is contractual capital expenditure — a US$60–64 billion FY2026 programme, three further years guided “even more significantly higher,” and US$265 billion of announced Arizona investment with no schedule. INTERPRETATION: none of this is a liability in the accounting sense, and most is cancellable at management’s discretion. But it is a real economic commitment, and it is the single largest claim on future cash flow. There is no evidence of unusual operating-lease, pension or guarantee exposure.

How conservative is the accounting? INTERPRETATION: conservative. Q2 2026 required no adjustment: no one-time gains, an effective tax rate of 18.0% in line with trend, stock-based compensation of NT$114 million against NT$572 billion of quarterly net income (immaterial), and non-operating items that made the quarter worse rather than better. Depreciation is heavy and front-loaded (NT$198.5 billion in Q2), which understates rather than flatters current earnings. Two caveats: interim numbers are 6-K/press-release grade and explicitly “not approved by the Board of Directors,” and the FY2026 audited 20-F will not appear until roughly April 2027.

How CapEx-hungry is the business? FACT: extremely, and increasingly. FY2025 capex US$40.6 billion (33.4% of revenue); FY2026 guided US$60–64 billion (35–38%); Q2 2026 alone US$15.7 billion. Trailing free cash flow of US$35.5 billion converts just 51% of US$69.3 billion of net income. INTERPRETATION: this is the defining financial characteristic of the business and the central risk to the valuation. It is not a solvency issue — the company is self-funding with US$67 billion of net cash — but it means an owner’s cash return is roughly half of reported earnings for as long as the build continues.

Capital Allocation & Management

How much FCF does the business generate, how does management use it, what is the philosophy? FACT: TTM free cash flow US$35.5 billion. Uses, in order: reinvestment in capacity (US$60–64 billion planned in 2026, i.e. more than free cash flow, funded from operating cash flow and the cash pile), then dividends (NT$24 per share in 2026, +33%, roughly 24% payout, ~0.9% yield), then nothing. INTERPRETATION: the philosophy is unambiguous and stated: “a higher level of capital expenditures is always correlated to higher growth opportunities in the following years… As long as there are business opportunities, we will not hesitate to invest.” Shareholders receive a growing but deliberately modest cash return; the priority is capacity.

Significant acquisitions recently? FACT: none. The Winbond arrangement (announced late June 2026) is a DRAM-wafer supply agreement feeding TSMC’s wafer-on-wafer and SoIC stacks — not an acquisition. A Sony sensor venture was reported in early August 2026. INTERPRETATION: the absence of empire-building M&A across three decades is a significant and underappreciated capital-allocation positive.

Buying back shares? FACT: no. INTERPRETATION: correct. At 10.8x book — the highest in the company’s recorded history — repurchasing stock would be the clearest capital-allocation error available, and management is not making it.

Issuing large amounts of new shares to insiders? FACT: no. Weighted-average shares 25,932 million (Q2 2026) against 25,929 million (Q2 2025) — dilution of 0.01%. Stock-based compensation is immaterial.

Compensation policy of directors/management? INTERPRETATION: conservative by US large-cap standards, with modest equity issuance and no evidence of option-driven behaviour. The Form 4 record (143 filings, 161 transactions, August 2025 – August 2026) shows total insider “purchases” of just 31,884 shares across 153 systematic plan-scale transactions, sales of 200,050 shares (essentially one 200,000-share VP disposal in May 2026), and bona-fide gifts of 6.17 million disposed against 5.17 million acquired — largely offsetting estate/trust transfers. There is no signal in either direction. Note also that as a foreign private issuer TSMC’s more meaningful insider disclosure runs through Taiwanese law, so the EDGAR window is incomplete and this evidence should be weighted lightly rather than read as reassurance.

Motivations of management? INTERPRETATION: on the evidence of the Q2 call, long-horizon franchise stewardship rather than near-term share-price maximisation. Three specific data points support this: the explicit refusal to price at “4x or 5x” what customers can bear (“I don’t want to squeeze them out from the market”); the self-imposed gross-margin ceiling at “about 68%” when memory peers earn 86%; and the refusal to engage in vendor/customer financing that would flatter reported revenue. Each of these costs shareholders money in the short run and buys franchise durability in the long run. The offsetting observation is that the same long-horizon posture is being used to justify withholding every quantitative multi-year anchor an outside investor could use to test the capital plan.

Valuation & Market Data

Is the stock an ADR, MLP, or K-1 issuer? FACT: an ADR. NYSE-listed ADS at a ratio of 1 ADS = 5 ordinary shares (TWSE: 2330). No MLP or K-1 issues. As a foreign private issuer TSMC files Forms 20-F and 6-K, not 10-K/10-Q, and reports under TIFRS in New Taiwan Dollars — so US-centric screening tools frequently mis-handle the share count and per-share data. (This report found and corrected exactly that: a widely-used aggregator applies a 4,676-million share count against 25,932 million ordinary shares actually outstanding, making every per-share figure in that feed wrong.) The ADR structure also introduces NT$/US$ translation risk into reported ADS earnings independent of operating performance; management flagged FX as one of the six factors determining profitability and noted “we have no control over the foreign exchange rate.”

Dividend policy? FACT: quarterly, growing, conservatively covered. NT$467 billion paid in 2025 (NT$18 per share, +28.6%); NT$24 per share in 2026 (+33%); a further increase signalled for 2027. Payout ratio approximately 24%; yield approximately 0.9% at US$418 per ADS. Management’s stated commitment is to “a sustainable and steadily increased cash dividend per share on both an annual and quarterly basis.”

How profitable is the business? FACT: gross margin 67.7%, operating margin 60.3%, net margin 55.6%, ROE 40.3%, ROIC 31.2% (Q2 2026). Incremental operating margin 90.1%.

Is net income diverging from cash from operations? FACT: operating cash flow exceeds net income, but free cash flow does not. Q2 2026: net income NT$706.6 billion against operating cash flow NT$783.4 billion — a healthy 1.11x conversion, consistent with heavy non-cash depreciation. The divergence appears one line lower: NT$496.0 billion of capital expenditure reduces free cash flow to NT$287.4 billion. On a trailing basis, US$69.3 billion of net income becomes US$35.5 billion of free cash flow — a 51% conversion. INTERPRETATION: this is not an earnings-quality red flag; it is a capital-intensity fact. Net income is real and cash-backed at the operating line. The question is not whether the earnings exist but how much of them an owner receives while the build continues. Working capital is the one item to watch: inventory rose 34% in two quarters and the cash conversion cycle lengthened from 26.1 to 36.3 days, which management attributes to the N2 ramp.

Risks & Downside

What factors would cause the stock to decline? In descending order of expected impact: (1) evidence of an AI capital-expenditure pause — hyperscaler guidance cuts, CoWoS order reductions, or two consecutive quarters of TSMC inventory build without a revenue step; (2) a fourth consecutive capex raise or FY2027 capex above US$70 billion, confirming that intensity does not normalise and the reverse-DCF cannot close; (3) a disclosed multi-billion-dollar leading-edge wafer commitment from a top-five fabless customer to Intel or Samsung; (4) a gross-margin guide below 64% not explained by utilisation, indicating the 2027 pricing did not stick; (5) Taiwan-Strait escalation; (6) broad AI-complex de-rating of the kind seen on 1 July and 29 July 2026, when the sector shed over US$1 trillion on sentiment rather than fundamentals.

Risk of a catastrophic loss? FACT/INTERPRETATION: yes, from exactly one source. The overwhelming majority of leading-edge capacity sits on one island within PRC missile range. This is the only risk on the list capable of permanently impairing capital rather than deferring it, and the factor model quantifies how much of the security it represents: a Taiwan country factor loading of +0.93, nearly equal to the market beta of +0.94, with the nearest empirical peers being Taiwan index funds before foundry peers. Secondary catastrophic-adjacent risks — earthquake, typhoon, drought — are real (the July 2026 Kumamoto earthquake forced a precautionary JASM evacuation) but are insurable, recoverable and partially mitigated by geographic diversification.

Chance of a total loss? INTERPRETATION: negligible in every scenario short of armed conflict over Taiwan, and non-trivial in that one. The balance sheet cannot fail — US$67 billion of net cash, a 2.46x current ratio, negative net debt to equity. The business cannot be competed away within a decade on any current evidence. The franchise risk is geopolitical and binary, not operational or financial.

Recent News & Events

Has the business environment changed recently? FACT: yes, in five specific ways since the prior report of June 6, 2026. (1) Q2 2026 delivered record results — revenue US$40.20 billion (+33.7%), gross margin 67.7% (+9.1 points year-on-year), net income +77.4% — and FY2026 revenue growth guidance was raised to “slightly above 40%.” (2) FY2026 capital expenditure was raised to US$60–64 billion, and an additional US$100 billion was committed to Arizona (US$265 billion announced total). (3) TSMC reportedly completed customer discussions on 2027 base-price increases of 5–10% across advanced and mature nodes. (4) Intel’s 18A yields were reported to have risen from ~65% to above 85% in a single quarter, with ASML confirming Intel as first to ship high-volume logic on High-NA EUV. (5) Memory-driven component-price inflation began visibly crowding out consumer and price-sensitive silicon demand, and TSMC responded with a DRAM-wafer supply arrangement with Winbond.

Significant acquisitions? FACT: none. The Winbond DRAM arrangement is a supply agreement; a Sony sensor venture was reported in early August 2026.

Change in accounting policies? FACT: none disclosed. Reporting remains under TIFRS in NT$, with USD convenience translation. Note that Q2 2026 figures were disclosed as “not been approved by Board of Directors” — standard for TSMC’s interim releases.

Recent changes — new markets, facilities, management? FACT: Facilities — an additional US$100 billion for Arizona (roughly four more fabs, front- and back-end, for 2nm and below); 13 leading-edge and advanced-packaging fabs planned in Taiwan; three additional 3nm fabs (Taiwan, Arizona, Japan) plus conversion of 5nm tools to 3nm; JASM Fab 1 expansion in Japan for CMOS image sensors; ESMC in Germany for automotive and industrial. Markets — entry into memory-adjacent integration via Winbond CUBE DRAM wafers for wafer-on-wafer and SoIC stacking; silicon photonics (COUPE) in early production with Broadcom and Nvidia as named adopters; a newly-articulated agentic-AI CPU demand vector. Management — no changes; C.C. Wei remains Chairman and CEO, Wendell Huang CFO. The one management-adjacent event is the pending trade-secrets suit against former Senior Vice President Wei-Jen Lo, filed November 2025 in Taiwan’s Intellectual Property and Commercial Court following his move to Intel; no ruling has been reported and Intel denies the allegations. Regulatory — Section 232 tariffs of 25% on advanced logic effective 15 January 2026, paired with a US–Taiwan agreement reducing or eliminating them for Taiwanese chipmakers expanding US production; and a loosening of US export controls toward China, with case-by-case licensing replacing the blanket presumption of denial.

APPENDIX B — Source Appendix

Taiwan Semiconductor Manufacturing Company Limited (NYSE: TSM / TWSE: 2330) · Report date August 7, 2026 All sources accessed 2026-08-07 unless otherwise stated. Primary sources are listed first.

1. Primary — Company filings and disclosures

2. Primary — Market and quantitative data

3. Industry data and third-party research

4. Trade press and financial media

5. Analytical frameworks

A prior report on TSMC by the same author, dated 6 June 2026, is the baseline against which this update measures change; the specific changes are set out in the “Changes Since the June 6, 2026 Report” section above. Contemporaneous reports by the same author on SK hynix, Broadcom, Intel and ASML were used for peer cross-reads; the underlying figures cited from them are the companies’ own public disclosures, listed above and in those companies’ filings.

6. Evidence-quality notes

Four disclosures about the evidence base, recorded so the reader can weight the conclusions correctly:

1. The two most thesis-relevant new facts are third-party reports, not disclosures. Intel’s 18A yield (KeyBanc estimate) and TSMC’s 2027 price increases (Nikkei Asia, unnamed sources) are both trade-press items. Both are labelled as such throughout and neither is treated as confirmed. The named Intel customer list is explicitly unconfirmed by any of the parties.
2. Interim financials are unaudited. Q2 2026 figures are 6-K/press-release grade and were disclosed as “not been approved by Board of Directors.” The audited FY2026 Form 20-F will not be available until approximately April 2027.
3. Three data feeds were found to be defective or stale and were corrected. ROIC.ai maps TSM’s tax charge into an extraordinary-item line and applies a wrong share count (all its per-share data is unusable for this issuer); the AZI valuation percentiles are computed on stale fundamentals during a period of 77% year-on-year earnings growth; the FactorsToday leaderboard is dated 2026-06-11 and therefore predates the June–July drawdown it would otherwise be used to describe. Corrected figures are used throughout and the corrections are stated in the memo body.
4. The EDGAR insider window is incomplete. As a foreign private issuer, TSMC’s more meaningful insider disclosure runs through Taiwanese law. The Form 4 record was swept in full and found to carry no signal, but its absence of signal should be read as uninformative, not as reassurance.