TSMC: Earnings Have Outrun the Share Price to a 28.1x Trailing Multiple, but a Limit-Up Close at NT$2,425 Leaves Only a Limited Margin of Safety

The quarterly EPS figure included a roughly NT$63 billion gain related to the disposal and valuation of Vanguard International Semiconductor shares, contributing about NT$2.24 per share. Adjusted for that disclosed item, Q2 EPS was approximately NT$25.01. The core result was still exceptionally strong: normalized earnings grew faster than revenue, utilization improved, product mix shifted toward HPC, and manufacturing-cost improvements offset part of the overseas-factory drag.

HPC produced 66% of Q2 revenue and grew about 20% sequentially. Smartphones contributed 22%, IoT 5%, automotive 4% and digital consumer electronics 1%. That mix is a source of both quality and risk. AI accelerators use expensive leading-edge wafers, large die sizes and advanced packaging, supporting revenue per unit of installed capacity. But economic dependence has shifted toward a relatively small number of accelerator developers and hyperscalers whose investment budgets are expanding at extraordinary rates. Alphabet raised its 2026 capital-expenditure plan to US$195–205 billion and indicated another significant increase in 2027. Microsoft reported quarterly capital spending of roughly US$41 billion and continued to describe compute capacity as constrained.

The filing data confirm the concentration. TSMC’s ten largest customers generated 78% of 2025 revenue, up from 70% in 2023 and 76% in 2024. Its largest customer accounted for 19% and its second-largest for 17%. The company does not identify those customers in its audited disclosure, so attribution of those percentages to named companies remains an external estimate, not a reported fact. The top-two share of 36% means that customer-specific product cycles can matter even when the overall customer count exceeds 500.

The strongest change since May is the gross-margin base. TSMC’s 2025 gross margin was 59.9%; Q1 2026 reached 66.2% and Q2 reached 67.7%. That does not mean a permanent eight-point reset has already occurred. Current margins benefit from high utilization, a favorable product mix, leading-edge scarcity, cost improvement and foreign-exchange conditions. Management explicitly expects the N2 ramp to reduce gross margin by three to four percentage points during the second half of 2026. Overseas factories initially dilute margin by two to three points, with the drag potentially widening to three to four points as the overseas network expands.

N2 nevertheless appears to be progressing as a commercial node rather than merely a technology demonstration. It entered volume production in Q4 2025 with yields described by management as good, generated 3% of Q2 2026 wafer revenue, and is ramping across Hsinchu and Kaohsiung. TSMC expects the N2 family, including N2P and A16 derivatives, to serve both smartphone and HPC customers over a long product life. The early margin cost is an investment in a future revenue pool, but investors should not assume that every point of dilution will be recovered immediately through pricing.

One item supplied in the research brief does not survive verification. Current official management guidance is not NT$520–560 billion of annual capital expenditure. TSMC raised 2026 capital-expenditure guidance to US$60–64 billion. At the July 31 exchange rate, that is approximately NT$1.945–2.074 trillion. About 70–80% is assigned to advanced process technology, around 10% to specialty technologies, and 10–20% to advanced packaging, testing, mask-making and other assets. The lower TWD figure appears to be a unit or period mismatch and should not be used in valuation.

Full-year guidance also moved further than the brief’s “above 30%” starting point. Management’s July outlook called for 2026 revenue growth slightly above 40% in US-dollar terms. Q3 guidance was US$44.6–45.8 billion, a 65–67% gross margin and a 56–58% operating margin. At the guidance exchange rate of NT$32 per dollar, the midpoint implies quarterly revenue of approximately NT$1.446 trillion.

This earnings trajectory changes the valuation argument. Trailing P/E remains elevated relative to TSMC’s approximately 19-times ten-year ADR average, but the ADR ratio does not distort the P/E itself because both price and earnings scale by the same five-to-one factor. More relevant to the present decision, an independently modeled 2026 EPS range of NT$102–105 places the ordinary share at roughly 23.1–23.8 times current-year earnings. That forward multiple is not cheap in an absolute sense, but it is much less demanding than the stale 29.5-times framing, particularly for a company growing revenue above 40% with net cash and a foundry share above 70%.

The bull and bear disagreement now centers on the duration of exceptional economics. Bulls see a company whose capacity is booked by customers with expanding AI budgets, whose market share has continued to rise, and whose gross margin may remain above 60% even after N2 and overseas dilution. Bears see an earnings peak produced by unusually concentrated AI spending, scarce capacity, favorable mix and a period in which customers have little bargaining leverage. They also see annual capex approaching NT$2 trillion, an overseas buildout that duplicates parts of Taiwan’s efficient manufacturing cluster, and a geopolitical tail that ordinary earnings models cannot fully price.

Qualitatively, this is high-quality compounding growth with cyclical and geopolitical tail risk. TSMC has compounded revenue and earnings over decades because its competitive advantage has repeatedly strengthened during industry transitions. Yet it remains a semiconductor manufacturer: utilization, customer inventories, technology cycles and capital intensity still matter. Its Taiwan concentration adds a low-frequency, extreme-impact risk that separates the stock from an otherwise comparable global compounder.

The old report’s “above fair value” conclusion cannot be carried forward. Earnings have risen too quickly, margins have reset too far upward and guidance has changed too materially. The company is no longer priced on the stale earnings base used in May. The remaining question is narrower: whether a forward multiple in the low-to-mid 20s provides enough compensation for margin normalization, customer concentration and Taiwan-related tail risk. The evidence points to fair-to-moderately favorable pricing, rather than deep undervaluation.

The supplied description of the previous report also contains a numerical inconsistency: the old NT$2,235 anchor was inside the stated NT$1,900–2,600 bull band, not above its NT$2,600 upper bound. The prior bands were in any case too detached from the company’s then-current earnings power to serve as useful starting points. The valuation ranges in this refresh are derived independently and bear no mechanical relationship to them.

Company evolution and financial history

TSMC began as an institutional solution to a structural problem. By the mid-1980s, semiconductor design and manufacturing were normally housed inside integrated device manufacturers. New design companies needed billions of dollars of fabrication assets before they could sell at scale, while an independent manufacturer serving many designers had yet to be proven as a durable business model. Morris Chang brought a long operating career at Texas Instruments and General Instrument, followed by leadership of Taiwan’s Industrial Technology Research Institute. His experience suggested that rising process complexity would eventually separate chip design from manufacturing.

The company was incorporated in Taiwan on February 21, 1987 as a venture involving the Republic of China government, Philips and other private investors. The government supplied institutional sponsorship and patient capital; Philips supplied capital, production knowledge and access to semiconductor technology; Chang supplied the model and operating discipline. TSMC’s original decision was unusually narrow: it would manufacture customers’ chips while refraining from selling competing designs under its own brand. That constraint became the foundation of customer trust.

The earliest strategic competitor was less a single foundry than the vertically integrated semiconductor model itself. Texas Instruments, Motorola, NEC, Intel, IBM and other integrated manufacturers owned fabrication plants and used external capacity selectively. Dedicated foundries such as United Microelectronics later competed directly, while Chartered Semiconductor emerged in Singapore. Chartered was ultimately acquired by GlobalFoundries, and UMC remains active, but neither matched TSMC’s combination of leading-node investment and pure-play customer neutrality.

TSMC listed its ordinary shares on the Taiwan Stock Exchange on September 5, 1994 under code 2330. Its American depositary shares began trading on the New York Stock Exchange on October 8, 1997. Accessible primary records confirm the dates and structure but do not provide a sufficiently reliable, directly comparable Taiwan IPO offer price and capital-raised figure; this report excludes those numbers rather than importing an unverified secondary estimate.

The listing story was not an internet-style promise of a new end market. It was a claim that manufacturing specialization could become an industry of its own. The capital market initially viewed TSMC as a cyclical contractor with unusually good utilization and customer growth. Over time, the label changed: first to dominant foundry, then to mobile-computing beneficiary, then to leading-edge technology platform, and finally to essential AI infrastructure.

The first stage, from 1987 through the late 1990s, validated the model. TSMC aggregated orders from multiple fabless customers, spread fixed manufacturing costs across those designs and reinvested cash into progressively better process technology. The company’s existence allowed designers to enter the market without owning fabs, reinforcing the rise of the fabless sector. The 1994 Taiwan listing supplied capital for a business in which equipment spending preceded revenue and where scale reduced unit costs.

The second stage, through the 2000s, converted a viable model into a scale system. The dot-com collapse and 2001 semiconductor downturn tested whether a pure-play foundry could remain profitable when utilization fell. TSMC reported that, with the exception of 1997, it generated positive free cash flow in every year following its 1994 IPO through 2005. Its estimated pure-play foundry share reached about 51% by 2008, showing that the downturns eliminated weaker capacity providers faster than they eroded TSMC’s position.

During this stage, TSMC’s manufacturing network, customer-service infrastructure and process-design ecosystem became mutually reinforcing. More customers generated more tape-outs and yield-learning data. More volume justified more advanced equipment. Better yields reduced customer cost per good die, bringing further designs. The cycle was capital-intensive, but each turn raised the hurdle for a new entrant.

The third stage, roughly 2010 through 2017, was built around mobile computing. Smartphone application processors required high performance under severe power constraints, while annual device launches rewarded foundries able to execute new nodes on time. Apple’s adoption of TSMC manufacturing beginning with the A-series transition in the middle of the decade gave TSMC a large, predictable volume customer capable of underwriting expensive leading-edge ramps. The relationship helped move TSMC from a diversified contractor toward the default producer for premium fabless processors, although it also increased customer concentration.

Capital expenditure became a competitive weapon. A leading node requires coordinated spending on lithography, deposition, etch, metrology, clean rooms, mask infrastructure, process design kits and customer qualification. TSMC was willing to spend before final demand was visible because missing a node could cause customers to design around the company for several product generations. A balance sheet that stayed conservative relative to the scale of its cash generation supported that willingness.

The fourth stage began around the 2018 launch of 7-nanometer volume production. TSMC moved 7 nanometers into production in 2018, 5 nanometers in 2020 and 3 nanometers in 2022. During the same period, Intel suffered repeated process delays and GlobalFoundries withdrew from leading-edge development. Samsung remained the only broad external challenger, but its foundry business had to coexist with Samsung’s own chip and device products. TSMC’s customer-neutral model acquired greater strategic value precisely when alternatives narrowed.

Management succession did not interrupt this path. Morris Chang retired in June 2018, with Mark Liu becoming chairman and C.C. Wei chief executive. Liu served as chairman until June 2024, after which Wei combined the chairman and CEO roles. The transition retained an internally developed engineering and operations leadership team rather than replacing it with an outside restructuring mandate.

The pandemic period exaggerated both the strength and cyclicality of the model. Remote work, cloud investment, device purchases and automotive shortages drove semiconductor demand sharply upward. Customers double-ordered and carried excess inventories. TSMC’s earnings expanded, but 2022–2023 brought a correction in smartphones and consumer electronics. Revenue declined in TWD during 2023, gross margin fell to 54.4%, and net income contracted. The company nevertheless continued investing and preserved its lead, enabling it to respond when AI accelerator demand arrived.

The fifth stage, beginning in 2024, is an AI-led reacceleration accompanied by geographic duplication. AI training and inference require large processors, advanced packaging and high-bandwidth memory interfaces. TSMC supplies the leading-edge logic and much of the advanced integration infrastructure, while customers supply processor architecture and memory vendors supply HBM. Revenue rose 31.6% in TWD during 2025, net income increased 46.4%, and advanced technologies reached 74% of wafer revenue.

Governments and customers are also asking TSMC to build outside Taiwan. Arizona’s first N4 facility entered volume production in Q4 2024. A second facility targets N3 production in the second half of 2027, while additional N2-and-below fabs, packaging facilities and an R&D center are planned. TSMC’s Arizona investment has expanded from an initial US$12 billion commitment to a stated US$265 billion plan. Japan has entered production, a second Japanese fab is planned, and a specialty-node facility is under construction in Dresden.

This global footprint creates customer proximity, political support and some geographic redundancy. It also breaks part of the economic logic that made Taiwan so efficient: dense suppliers, short engineering loops, experienced labor and large shared infrastructure. Management’s disclosed overseas margin drag confirms that the diversification is economically costly before subsidies and strategic benefits are considered.

Selected financial history illustrates how each strategic turn changed the earnings base.

† H1 2026 gross margin is revenue-weighted from Q1 and Q2. ‡ Q2 2026 figure.

The data come from TSMC’s official annual-report archive, the 2025 annual report and 2026 quarterly releases. The pattern matters more than any single year: margins rose as the company moved from a broad foundry mix toward a high share of differentiated advanced nodes, but they fell during utilization and inventory corrections. The 2026 margin level is evidence of strong economics, not proof that cyclicality has disappeared.

Revenue growth has come from three sources. Wafer volume expanded as the fabless industry grew; revenue per wafer increased as customers moved to more expensive nodes; and product mix shifted toward large HPC die and advanced packaging. TSMC shipped 15.0 million 12-inch-equivalent wafers in 2025, up from 12.9 million in 2024, while advanced technology rose five percentage points as a share of wafer revenue. This combination explains why revenue grew faster than simple wafer volume.

Gross margin improved through higher utilization, better yields, richer node mix and pricing that reflects the increasing value and cost of advanced technology. The reverse also holds. When customer inventory rises or a new node starts below mature utilization, depreciation is spread across fewer good wafers and margin falls. N2 and overseas fabs will apply that mechanism in 2026–2028, even if consolidated demand remains strong.

The long-term relationship between earnings and operating cash flow is healthy but must be interpreted alongside capex. From 2021 through 2025, aggregate operating cash flow was roughly 1.53 times aggregate net income. Depreciation and amortization are large non-cash charges, so operating cash flow normally exceeds accounting profit. Yet annual capex absorbs much of that cash, particularly during leading-node ramps.

The five-year cash-conversion ratio looks excellent, but free cash flow after all capex was only about 58% of net income in 2025. In H1 2026, operating cash flow totaled approximately NT$1.482 trillion, capex NT$847 billion and free cash flow NT$636 billion, equivalent to about half of H1 net income. That gap comes mostly from growth investment, not from weak collections or aggressive accounting.

The balance sheet can carry that program. At June 30, 2026, cash and marketable securities totaled NT$3.518 trillion against NT$864 billion of long-term debt. Accounts-receivable days were 29 and inventory days 87. Net property, plant and equipment reached NT$4.303 trillion, illustrating both the asset base and the cost of staying ahead.

Returns on capital have risen because advanced-node margins and utilization expanded faster than the asset base. TSMC’s own long-term framework targets through-cycle gross margin of at least 56% and return on equity in the high 20s; current returns run above that. Part of the gap is a favorable cycle and constrained capacity. The durable advantage is the ability to earn above-industry returns across cycles, not to hold Q2 2026 margins every quarter.

The capital-market narrative has changed in four broad waves. Before the mid-2010s, investors primarily treated TSMC as a high-quality semiconductor cyclical. Mobile computing then earned it a growth premium. The 2020–2021 shortage period added scarcity value, followed by a 2022 derating as rates rose and inventories corrected. Since 2023, AI has expanded both earnings estimates and the valuation center. The share’s 52-week range through July 2026 was NT$1,125–2,535, showing how quickly the market has repriced the AI outlook and geopolitical risk within the same year.

The current trailing multiple of 28.1 times is about 48% above the reported ten-year average of approximately 19 times. It belongs in the upper part of TSMC’s historical distribution. The business is also different from its ten-year average: foundry share is higher, advanced-node alternatives are fewer, AI has increased compute intensity, and margins are stronger. A permanently higher valuation center is justified, but a return to the extreme scarcity premiums of an unconstrained AI boom should not be treated as the default.

Business model, moat, industry, and governance

TSMC reports one principal foundry business rather than a portfolio of unrelated divisions. Economic segmentation is better understood by application platform, process node and service layer. The same fab can serve several end markets, while a customer can use leading-edge logic, specialty processes and advanced packaging in a single system.

HPC has become the profit center because it combines strong volume growth with leading-edge nodes, large dies and packaging demand. Smartphones remain economically important, especially as initial node ramps historically depend on high-volume premium devices. Automotive and IoT diversify utilization across mature and specialty nodes, but neither currently determines consolidated growth.

The concentration in N3 and N5 explains much of the margin strength. These nodes have moved beyond their earliest ramp costs but remain essential to premium processors. N2 begins at low utilization and carries elevated depreciation, which is why an otherwise strong demand ramp can reduce consolidated margin initially.

Fixed costs dominate TSMC’s economics. Fabs, lithography systems, process development, clean-room infrastructure and engineering organizations must be funded before the related wafer revenue arrives. Variable costs include wafers, chemicals, gases, power, direct production labor and some outsourced services. Depreciation behaves economically like a fixed cost over the useful life of the equipment. Small changes in utilization swing profit hard.

This creates strong operating leverage in both directions. During Q2, higher utilization and cost improvement helped lift gross margin to 67.7%. In a downturn, the company cannot quickly remove depreciation, R&D teams or maintenance spending without damaging future competitiveness. A five-point decline in consolidated gross margin at the current revenue run rate would reduce annual gross profit by approximately NT$270–300 billion before secondary expense effects.

R&D and capex are continuing costs of the moat, not optional growth projects that can be switched off indefinitely. TSMC spent decades building a cadence: one node ramping, another qualifying, the next in research. Cut that cadence to protect one year of free cash flow and the company could surrender customer designs two or three years later.

The moat is a system of process learning, scale, neutrality and customer integration; none of the four works as well alone. Process technology supplies power, performance and density. Manufacturing scale converts equipment and engineering spending into lower cost per good die. Customer neutrality reduces the strategic concern that proprietary designs will strengthen a product competitor. Design libraries, electronic-design-automation flows, intellectual property blocks and packaging create switching costs before a wafer enters the fab.

Yield is the most important hidden variable. Nominal node labels do not tell customers how many usable chips come from a wafer, how consistent performance is across production lots, or whether volume can meet a launch date. TSMC’s accumulated production data and engineering response reduce those risks. A rival can announce similar transistor dimensions, but customers select a process based on total cost per functioning chip, schedule confidence and ecosystem readiness.

Scale is particularly valuable at the leading edge. TSMC manufactured 12,682 products across 305 technologies for 534 customers during 2025 and operated more than 17 million 12-inch-equivalent wafers of annual capacity. This breadth spreads technology and infrastructure costs across many customers while preserving the option to shift capacity among products as individual cycles change.

The pure-play structure is a real moat because it has survived decades and adverse conditions. Samsung can offer memory, foundry and packaging together, but its device and semiconductor-design businesses may compete with customers. Intel is trying to build an external foundry while manufacturing its own processors. TSMC’s organizational promise is simpler: customer products are the output, not an input into a competing branded platform.

Advanced packaging extends the moat beyond transistor manufacturing. AI accelerators increasingly connect multiple compute dies, high-bandwidth memory and input-output components. CoWoS, SoIC and related technologies let TSMC capture more value per system and coordinate design decisions between logic and packaging. A customer that qualifies both process and packaging with TSMC faces greater switching cost than one purchasing a commodity wafer.

Capital itself is a barrier. TSMC can fund a US$60–64 billion annual capex budget substantially from operating cash flow and still maintain a net-cash balance sheet. A new entrant would need similar spending without TSMC’s existing customer commitments, yield data or utilization. Governments can subsidize buildings, but subsidies do not instantly create the engineering learning required to produce good dies economically.

The moat has limits. Semiconductor process leadership must be renewed at each node. A sustained TSMC delay, a major Samsung yield breakthrough, Intel’s successful externalization of 18A or 14A, or a shift toward architectures produced efficiently on mature nodes could weaken pricing power. Customers also have incentives to cultivate alternatives because dependence on one supplier is strategically uncomfortable.

The industry remains cyclical even as secular demand grows. TSMC is exposed to the semiconductor inventory cycle, customer capex cycle, technology-iteration cycle and global macro cycle. AI currently places the company in a strong upcycle: leading-edge capacity and packaging are constrained, hyperscaler investment is rising and revenue is growing above 40% in USD. Smartphones and consumer electronics are recovering more slowly, which reduces but does not eliminate dependence on AI.

Foundry industry structure is unusually concentrated. TrendForce data placed TSMC’s conventional foundry share at approximately 72.3% in Q1 2026, followed by Samsung at 6.5%, SMIC at 5.1%, UMC at 3.9% and GlobalFoundries at 3.3%. TSMC’s share had already reached about 70.4% in Q4 2025. The continued gain indicates that AI demand is widening the scale gap rather than distributing evenly across suppliers.

The industry’s profit pool sits at the intersection of advanced logic, scarce process capacity and complex packaging. Equipment suppliers such as ASML retain strong bargaining power because extreme-ultraviolet lithography alternatives are limited. Leading foundries retain power over customers when capacity is scarce. Customers regain leverage when utilization falls or when competing processes become technically interchangeable. Mature-node foundries operate with lower barriers and consequently lower margins.

Entry barriers are higher than physical construction costs suggest. A new fab needs process recipes, equipment tuning, defect control, qualified suppliers, libraries, design tools and customer confidence. Each new node shortens the useful life of an entrant’s learning. State-supported Chinese foundries can become formidable at mature and selected advanced nodes, but export controls on advanced equipment restrict the speed and economics of their progress.

Policy and geopolitics shape both demand and cost. The United States awarded TSMC Arizona up to US$6.6 billion in direct CHIPS incentives under the 2024 agreement, while customer and government pressure encouraged increasingly advanced US production. Subsidies reduce the accounting cost but do not fully offset duplicated supply chains and lower initial productivity.

US export controls restrict shipments of advanced semiconductor technology and equipment to China and can change TSMC’s addressable customer set. In 2025, the Bureau of Industry and Security removed a license-free equipment-treatment mechanism for certain foreign-owned fabs in China, while indicating that licenses could support existing operations but not expansion or technology upgrades. This makes TSMC’s Nanjing operations structurally different from capacity in Taiwan, Japan or the United States.

Taiwan-Strait risk cannot be reduced to a modest quarterly earnings adjustment. A blockade, military conflict or interruption of power, water, ports or communications could stop production, disrupt customers globally and impair normal price discovery in the shares. Overseas fabs add partial redundancy, but the most advanced R&D, the largest engineering cluster and much of the leading-node capacity remain in Taiwan. This is a low-frequency risk with potentially catastrophic impact.

The valuation later in this report reflects that risk in three ways: a higher required return than would be assigned to a similar business without Taiwan exposure, conservative scenario multiples below those of an unconstrained global technology monopoly, and wide valuation bands. The model does not assign a precise probability to armed conflict because doing so would create false precision. A full conflict outcome falls outside ordinary going-concern valuation.

Governance is comparatively straightforward. C.C. Wei has served as CEO since 2018 and chairman since June 2024. The National Development Fund owns 6.38%, while no operating founder, family or private controller possesses voting control. ADR holdings represent 20.49% of ordinary shares in custody, but that line is a depositary structure rather than a single beneficial owner.

TSMC has no dual-class founder-control structure. Executive compensation includes substantial variable compensation and restricted stock, with performance measures tied partly to relative shareholder return and ESG objectives. The company adopted a clawback policy in 2023 and introduced a long-term incentive plan effective in 2025. The combined chairman-CEO role reduces separation between management and board leadership, although independent directors and committee structures provide oversight.

Capital allocation has favored technology investment first, balance-sheet resilience second and dividends third. TSMC did not repurchase common shares in 2025. It paid NT$18 per share for that year and management expects shareholders to receive NT$24 per share during 2026, with a higher dividend anticipated in 2027. The dividend is supported by earnings but remains a small part of total expected return at the current price.

The main governance risk is not extraction by a controlling shareholder. It is the possibility that geopolitical objectives imposed by governments and customers lead management to accept structurally lower returns on overseas capital. TSMC can negotiate subsidies and pricing, but it cannot fully control the political terms under which globally diversified capacity is demanded.

Competition and current fundamentals

TSMC operates in a market with several nominal foundry competitors but only two plausible leading-edge challengers. Samsung competes in advanced logic while also operating memory, consumer electronics and branded-device businesses. Intel is rebuilding process credibility and trying to convert an internal manufacturing system into a customer-facing foundry. UMC and GlobalFoundries concentrate on mature, specialty and differentiated nodes. SMIC is the largest Chinese foundry and a strategic competitor, but export controls constrain access to the most advanced production tools.

The concentration says more than a conventional league table. TSMC is larger than the next four competitors combined. Its lead also widened during the AI upcycle, when technical and packaging requirements became more difficult. The industry is closer to a dominant platform with specialized alternatives than to a balanced five-player market.

Samsung is the integrated alternative. It can combine foundry, advanced packaging and memory, an attractive proposition for some AI systems. Its enormous consolidated balance sheet lets it fund new nodes through memory cycles. Customers may also use Samsung to diversify supply or gain negotiating leverage with TSMC. The weakness is structural complexity: Samsung’s foundry must compete for capital inside a conglomerate and convince customers that their product road maps will remain confidential from Samsung’s own device and chip units.

Samsung’s foundry share below 7% suggests that technology announcements have not yet converted into broad customer volume comparable with TSMC. Recent large customer agreements could improve utilization and establish sub-2-nanometer references, making Samsung the most credible source of medium-term share loss for TSMC. Yet a contract win does not erase the accumulated difference in customer ecosystem, yield history or production scale.

Intel is the state-supported turnaround, and its strongest asset is an installed manufacturing base and deep process-engineering knowledge. Its strategic appeal to US customers is domestic advanced capacity. The company is also investing in advanced packaging and attempting to make its internal process road map useful to external customers. Q2 2026 consolidated revenue rose 25% to US$16.1 billion, but Intel’s GAAP earnings remained distorted and external foundry revenue was only about US$293 million, while the foundry segment continued to carry large losses.

Intel’s challenge is organizational as much as technical. An internal product group can tolerate process trade-offs that an external customer will not. A foundry customer expects neutral intellectual-property treatment, predictable design rules, service responsiveness and capacity commitments. Intel may become a useful second source before it becomes a financially comparable foundry.

UMC has settled into disciplined mature-node production, avoiding the spending race at the smallest geometries and focusing on 22/28 nanometers and other specialty processes. Q2 2026 revenue reached NT$68.73 billion, gross margin 32.5%, operating margin 21.8% and utilization 85%. Customers choose UMC for cost-effective, long-lived processes rather than for the highest transistor density. Its economics are less exposed to leading-edge capex, but it cannot take the principal AI accelerator profit pool from TSMC without a fundamental change in strategy.

GlobalFoundries is a Western specialty foundry, focused on radio frequency, communications, automotive, industrial, silicon photonics and other differentiated technologies rather than matching each leading node. Q1 2026 revenue was US$1.634 billion, equivalent to roughly NT$53.0 billion at the report-date exchange rate, with a 27.6% gross margin and 11.0% operating margin. Customers choose it for geographic supply, specialty intellectual property and long product lives.

SMIC fills the Chinese strategic-supply niche. Domestic customers and policy support create demand, while restrictions on advanced lithography and other tools limit its production economics at the leading edge. It can pressure mature-node pricing and progressively substitute imports, but the immediate risk to TSMC is greater through lost Chinese addressable market and policy fragmentation than through a globally competitive N2-equivalent process.

The market-share gap is evidence that customers currently value production certainty more than nominal second-source diversification. Customers would prefer more capable suppliers, but their highest-value products cannot be moved merely because another foundry offers a lower wafer quote. Requalification requires redesign, new libraries, yield validation and packaging changes. TSMC’s primary ecological niche is the trusted manufacturing platform for products whose time-to-market and power efficiency matter more than lowest nominal wafer cost.

Peer valuation comparisons require caution because the businesses are different.

† GlobalFoundries Q1 2026; its Q2 result was scheduled after the research date. ‡ Intel’s consolidated margin is not a clean measure of external foundry economics.

TSMC’s P/E premium to Samsung is justified by cleaner exposure to foundry growth, higher margins and freedom from memory and consumer-device cycles. GlobalFoundries trades at a higher P/E despite lower growth and margins because its current earnings base is smaller and more sensitive to incremental recovery. Intel’s trailing P/E is unusable because restructuring, segment losses and non-operating items dominate reported earnings.

The last four quarters show how rapidly TSMC’s earnings base moved.

Trailing revenue is approximately NT$4.44 trillion and trailing EPS NT$86.27. The progression reflects stronger advanced-node utilization, AI-driven HPC mix and improving cost absorption. Q2’s investment gain raised reported EPS, but the operating-margin trend predates that item.

H1 2026 revenue reached NT$2.404 trillion, an increase of 35.6% from the prior year. June alone grew 67.9%, partly reflecting timing and comparison effects, but the six-month total confirms that growth is broad enough to support management’s raised annual outlook.

Q2’s beat consisted of more than revenue. Gross margin exceeded the midpoint of guidance as cost improvement and utilization offset the initial overseas drag. Operating expenses did not grow as rapidly as gross profit, producing operating leverage. Cash flow remained positive after a sharp increase in capex, although the free-cash-flow conversion rate fell relative to accounting earnings.

Management’s Q3 guidance implies another sequential revenue increase but lower margins. The gross-margin midpoint falls 1.7 points from Q2, and the operating-margin midpoint falls 3.3 points. The expected decline is consistent with N2 ramp cost, overseas dilution, exchange-rate movements and the normalization of unusually favorable Q2 conditions. A margin decline within guidance would not itself invalidate the thesis; a decline materially below 60% without a corresponding acceleration in N2 revenue would.

Analyst expectations have moved upward following the beat-and-raise quarter. As one visible example, Morningstar raised its TWD fair-value estimate to NT$3,440 from NT$2,700, citing hyperscaler investment and the larger capex program. That is not a market-consensus figure, but it illustrates the direction and magnitude of post-result estimate revisions.

The market is trading real earnings growth rather than AI terminology alone. Revenue, margin, EPS, cash balances, advanced-node mix and customer spending all support the narrative. The narrative component lies in the assumed duration: current prices implicitly require that AI infrastructure spending remains strong enough to absorb TSMC’s new capacity and that a cyclical correction does not arrive before N2 and overseas fabs reach efficient utilization.

The bull case rests on five observable facts. Foundry share exceeded 72% in Q1; HPC reached 66% of Q2 revenue; N2 entered volume production with good reported yields; full-year USD revenue guidance moved slightly above 40%; and large cloud customers continued to raise capital spending. Together, these suggest that TSMC’s near-term growth is demand-constrained less than it is capacity-constrained.

The bear case also rests on observable facts. The top two customers supplied 36% of 2025 revenue; annual capex is approaching NT$2 trillion; N2 and overseas facilities carry an explicit combined margin burden; current P/E remains well above the ten-year average; and much of the company’s leading-edge manufacturing remains concentrated in Taiwan.

A normal AI-capex slowdown would not need to eliminate demand to hurt earnings. If accelerator customers moved from urgent capacity reservations to inventory digestion, leading-edge utilization could decline while depreciation from N2 and overseas fabs increased. A five-to-eight-point margin decline combined with slower revenue growth could reduce EPS by 20–30% from the base forecast. The stock would then face both earnings decline and multiple compression.

Pricing power, though, may offset more dilution than bears assume. A leading-edge wafer represents a small portion of the economic value of an AI server or premium smartphone, while failure to secure capacity can delay billions of dollars of product revenue. TSMC has room to price for increasing process and geographic costs as long as customers lack qualified alternatives. The key evidence will be consolidated margin after N2 exceeds a mid-single-digit revenue share and Arizona’s second fab begins production.

Valuation, risks, and tracking framework

The valuation starts with the ordinary share at NT$2,425, not the ADR. The share trades at 28.1 times trailing EPS of NT$86.27 and yields about 1.0% on the indicated NT$24 annual dividend. Using independently estimated 2026 EPS of NT$102–105, the current-year multiple is approximately 23.1–23.8 times. The forward compression is substantial because H1 2026 alone produced NT$49.33 per share.

Historical comparison points in two directions. A 28-times trailing multiple is expensive relative to the reported ten-year average of roughly 19 times and likely sits in the upper historical quintile. A low-to-mid-20s forward multiple is less extreme given TSMC’s market-share gains, high margins, net cash and expected revenue growth above 40%. The stock is priced for continued quality, but not for the full continuation of Q2’s 67.7% gross margin.

Peer comparison does not make TSMC automatically cheap. Samsung’s lower multiple reflects conglomerate exposure and cyclicality; GlobalFoundries’ higher multiple rests on a smaller recovering earnings base; Intel’s reported multiple is unusable. TSMC deserves a premium to diversified or loss-making foundries, but its absolute return must still stand on owner earnings and future cash flow.

Cash-flow passthrough is the first absolute-valuation check. Aggregate 2021–2025 operating cash flow was about 1.53 times aggregate net income. This confirms strong accounting quality. The apparent surplus comes primarily from depreciation, which must be compared with continuing capital requirements rather than treated as freely distributable cash.

TSMC does not disclose maintenance capex separately. This report estimates 2026 maintenance and sustaining capex at NT$0.85–1.00 trillion, or roughly 41–51% of the official NT$1.945–2.074 trillion total. The estimate includes replacement equipment, sustaining technology, environmental infrastructure and capacity required to preserve current earnings. The remaining NT$1.0 trillion or more is treated as growth and strategic capex for N2, advanced packaging and geographic expansion. This is an analytical assumption, not company guidance.

On an annualized operating-cash-flow base of roughly NT$3.0–3.2 trillion, subtracting estimated maintenance capex gives owner earnings of approximately NT$2.0–2.3 trillion, or NT$77–89 per share. The resulting owner-earnings multiple is about 27–32 times. It is moderately higher than the forecast accounting P/E but not more than 30% higher in the central estimate, so the scenario model uses both normalized EPS and owner earnings rather than discarding earnings-based valuation altogether.

Headline free cash flow after all capex is much lower because it treats every new fab as maintenance. That is too conservative for a company whose revenue is expanding above 40%, but it is a useful warning: the current equity value depends on new capacity generating future earnings. If the growth capex fails to earn adequate returns, the distinction between maintenance and growth disappears economically.

The valuation uses three methods. A forward P/E framework captures the market’s treatment of TSMC as a high-quality semiconductor compounder. An owner-earnings yield checks whether accounting profit converts into distributable cash. A five-year discounted-cash-flow cross-check uses required equity returns of approximately 9–11.5%, terminal growth of 3–4%, and a net-cash adjustment near NT$100 per share. The higher discount rates and broad ranges explicitly reflect Taiwan and export-control risk.

† Includes estimated dividends and an assumed normalization toward the scenario valuation range. This is valuation-scenario analysis within a research framework, not investment advice.

The conservative range assumes that AI spending grows much more slowly after 2026, N2 and overseas depreciation push gross margin toward 60%, and TSMC’s multiple contracts toward a quality-adjusted but historically restrained level. The base case assumes that AI demand broadens from a few accelerator programs, N2 pricing and yield recovery offset much of the initial drag, and gross margin remains structurally above the old 56% through-cycle objective. The optimistic case requires strong AI accelerator growth, packaging scarcity, successful pricing and no material escalation in export controls or Taiwan risk.

The P/E and DCF methods diverge most in the conservative case. A conventional multiple recognizes TSMC’s franchise value and market-share durability; the DCF penalizes high sustaining capital and a high risk-adjusted discount rate. That divergence is precisely why a price near NT$1,900–2,000 is materially more attractive than the current quote: it is supported by both methods rather than only by continued premium valuation.

The market expectation embedded in NT$2,425 is approximately that 2026 EPS exceeds NT$100, gross margin remains above 60% after the N2 and overseas drag, and 2027 growth remains at least in the low-to-mid teens. It does not require the Q2 gross margin to persist at 67.7%, but it does require management to prevent the combined dilution from becoming additive and prolonged.

The most important expectation-gap variables are gross margin, N2 revenue share, advanced-packaging capacity and hyperscaler capex. A Q3 gross margin near 66% would support the view that pricing and cost control are absorbing the ramp. A result near or below 62% would suggest that the explicit three-to-four-point N2 drag and overseas burden are arriving faster than the market expects.

The next result will also test whether Q2’s HPC growth was a capacity release or the beginning of an even steeper shipment curve. Revenue within Q3 guidance but with weaker margins could disappoint because the market already understands the demand story. Revenue above guidance and a gross margin at or above 67% would force another round of EPS upgrades.

The base case’s most fragile assumption is a sustainable gross margin of at least 61%. Reducing the assumed benefit of the margin structure to 70% of the modeled improvement lowers normalized 2027 EPS from approximately NT$124 to roughly NT$109–113. Applying a 21–22-times multiple reduces base value from around NT$2,750 to approximately NT$2,300–2,480. The current price would then offer little upside.

The independent margin-of-safety test is less favorable than the growth-adjusted P/E. The current price is near the midpoint of the conservative blended fair-value range rather than at a discount to it. The margin of safety against ordinary cyclical disappointment is limited, and the price provides no protection against the extreme Taiwan tail.

If EPS remained flat at NT$86.27 for three years and the P/E stayed at 28.1 times, the investor’s return would be approximately the 1.0% dividend yield. That is below Taiwan’s 1.90% ten-year government-bond yield as of July 31. If the P/E instead normalized to 20 times, three years of NT$24 annual dividends would still produce an annualized total return of approximately negative 9.5%. Under the flat-earnings test, there is no margin of safety at this buy price.

Margin-of-safety verdict: not obvious. TSMC is no longer a clear “good company, bad price” case because earnings have caught up substantially. It remains a good company at a price that depends on continued growth, leaving a full-sized new position better served by a lower entry point.

The first permanent-loss risk is AI customer normalization. Probability is medium and impact is high. The observable indicators are hyperscaler capex guidance, accelerator lead times, advanced-packaging utilization, HPC revenue growth and customer inventories. The transmission path runs from slower customer orders to lower leading-edge utilization, then to gross-margin contraction, EPS cuts and a lower P/E as TSMC is relabeled from AI compounder to semiconductor cyclical.

Customer concentration intensifies that path. The two largest customers already contribute 36% of revenue. A product delay, internal accelerator transition or deliberate second sourcing by either customer can affect capacity planning before aggregate AI demand changes. The alert is a sustained increase in top-two concentration above 45%, or evidence that a major customer is moving a leading product family to a rival process, not a one-quarter movement in customer mix.

The second risk is margin dilution from N2 and overseas fabs. Probability is high because management has already quantified it; impact is medium to high. The observable indicator is gross margin after adjusting for FX and utilization. The loss path is persistent depreciation and labor inefficiency, followed by lower owner earnings and a reassessment of whether overseas capex earns TSMC’s historic return on capital.

The danger is greater if N2 and overseas dilution overlap rather than replace one another. Management’s comments imply three-to-four points from N2 and an overseas burden that can widen to three-to-four points. These effects should not simply be added to predict a seven-to-eight-point decline because pricing, mix and cost improvements operate concurrently. Gross margin below 58% for two consecutive quarters would nevertheless indicate that the offsets are inadequate.

The third risk is capital misallocation through geographic duplication. Probability is medium and impact is medium. Arizona, Japan and Germany offer strategic insurance and customer proximity, but their economics may remain below Taiwan’s. The observable indicators are overseas utilization, subsidy receipts, customer pricing, consolidated return on equity and capex as a percentage of revenue. If capex remains above 45% of revenue while free cash flow stays below 40% of net income, the market may conclude that growth is consuming rather than creating owner value.

The fourth risk is geopolitics and export controls. The probability of incremental restrictions is medium; the probability of a severe Taiwan event is low but unknowable; the impact ranges from medium to catastrophic. Export-control tightening can remove Chinese customers or restrict factory equipment. A blockade or conflict can stop production and impair market liquidity. Overseas capacity reduces selected customer exposures, but it will not replicate Taiwan’s entire manufacturing and R&D cluster during the investment horizon.

The fifth risk is technology or execution failure. Probability is low to medium and impact is high. A significant delay in N2P, A16 or the following A14 generation would give Samsung or Intel an opening. The indicators are yield commentary, volume-production dates, customer tape-outs and whether advanced-node share continues rising. A one-quarter ramp cost is normal; a schedule slip that causes a major customer to redesign elsewhere can affect several years of revenue.

The sixth risk is currency. Most customer contracts and guidance are economically linked to USD, while a large part of Taiwan’s cost base is in TWD. A stronger TWD reduces translated revenue and margin, while a weaker TWD supports them. The Q3 guidance uses NT$32 per US dollar, compared with a July 31 market rate around NT$32.41. A move below NT$30 without pricing offsets would create a visible earnings headwind.

Positive catalysts over the coming year include N2 rising from 3% to a high-single-digit share without consolidated gross margin falling below 60%; advanced-packaging capacity additions that release currently constrained shipments; another increase in hyperscaler capex; pricing that offsets overseas costs; and 2027 dividend growth supported by free cash flow.

Negative catalysts include Q3 or Q4 gross margin below guidance, a customer inventory correction, capex rising again without corresponding revenue guidance, a leading product moving to Samsung or Intel, stricter export controls, or evidence that Arizona’s cost gap is larger and more persistent than management’s stated consolidated drag.

The gross-margin and N2 indicators distinguish healthy ramp dilution from structural cost erosion. Capex and free-cash-flow conversion test whether demand is arriving quickly enough to justify new capacity. Customer concentration and HPC mix capture the dependence on AI spending. The official financial calendar lists October 8, 2026 for the next quarterly conference, subject to change during the year.

Monthly revenue is the earliest operating signal, but it should be read in rolling three-month groups because shipment timing, exchange rates and calendar effects can distort an individual month. Gross margin is the more decisive quarterly indicator because it combines utilization, pricing, yield, node mix and overseas costs into one observable outcome.

Cross-synthesis and final research conclusion

Looking vertically, TSMC has proven an ability to turn each increase in semiconductor complexity into a stronger competitive position. The company did more than benefit from the growth of semiconductors. It created a business structure that allowed other companies to specialize in design, then used their combined demand to fund manufacturing investments none could justify alone. This is the capability that survived the dot-com crash, the global financial crisis, mobile transitions, the smartphone inventory cycle, the pandemic shortage and the 2023 correction.

Its past success came from a combination of era tailwinds and deliberate choices. The rise of fabless design supplied the initial tailwind. Mobile computing created predictable leading-edge volume. Cloud and AI increased the economic value of energy-efficient computation. TSMC’s contribution was to preserve customer neutrality, spend ahead of demand, institutionalize yield learning and maintain a balance sheet capable of supporting that spending through downturns.

Luck mattered in the sense that semiconductor demand became much larger than it appeared in 1987 and that several competitors made severe execution errors. Intel’s process delays and GlobalFoundries’ withdrawal from leading-edge development reduced alternatives. Yet TSMC repeatedly placed itself in position to capture those errors. A rival’s failure creates only a temporary opening unless another company has qualified technology, available capacity and customer trust ready at the same time.

Those success factors remain present. N2 entered volume production on schedule, advanced-node share continues to rise, customers are increasing AI infrastructure budgets and the balance sheet holds substantial net cash. The company’s foundry share has moved above 70%, rather than falling under the pressure of larger competitors.

The factors are also becoming more expensive to preserve. TSMC must fund an N2 ramp, advanced packaging and a multi-country manufacturing network at the same time. The US$60–64 billion capex plan is more than twice 2024’s approximate US$29 billion spending. More capital is being committed before overseas facilities have demonstrated Taiwan-equivalent returns. The business remains financially sound, but incremental returns are now the central question.

Horizontally, TSMC’s advantage over Samsung runs deeper than a better node: a cleaner customer proposition and a much larger learning base. Its advantage over Intel is foundry culture and external-customer scale. Its advantage over UMC and GlobalFoundries is access to the highest-value process and packaging profit pools. Its advantage over SMIC is equipment access, ecosystem breadth and global customer qualification.

Samsung is the competitor most likely to take meaningful profit share because it can combine foundry, packaging and memory and is willing to spend through cycles. Intel is the competitor most likely to receive policy-driven customer trials. Neither currently matches TSMC’s external leading-edge volume. TSMC’s weakness is temporary in operating terms but structural in geography: overseas costs can improve with learning, while Taiwan concentration cannot be eliminated without duplicating much of the cluster that created its returns.

The current valuation rewards both proven quality and part of future AI success. A 28.1-times trailing P/E pays for more than past performance. It assumes that current earnings continue rising and that margins do not revert toward historical semiconductor levels. The forward multiple near 23–24 times is less demanding because the 2026 earnings base has moved sharply upward. The share is neither priced like an ordinary cyclical nor valued as an unconstrained software monopoly.

The market’s most likely misjudgment is still using the 2025 margin base to describe the earnings power. Gross margin moved from 59.9% in 2025 to 67.7% in Q2 2026, while revenue guidance moved above 40% growth. Even after allowing for Q2’s investment gain, trailing and forward earnings are materially higher than the prior report assumed. The old valuation objection weakened through earnings growth, exactly as the refresh brief suggested might occur.

The opposite possible misjudgment is extrapolating the new margin base. Q2 combined high utilization, leading-edge mix, favorable cost absorption and strong customer urgency. N2 and overseas depreciation will become more visible over the next several quarters. A stable margin above 60% would confirm a structural improvement. Treating 67–68% as the new floor would be aggressive.

For the next 12 months, the critical variables are Q3 and Q4 gross margin, N2 revenue share, AI accelerator shipments, advanced-packaging capacity and the 2027 capex signal. Revenue is likely to remain strong enough that margin and cash conversion determine the expectation gap.

For the next three years, overseas-fab economics become decisive. Arizona’s N3 ramp, N2 capacity in Taiwan, customer pricing and advanced-packaging expansion will determine whether TSMC can keep return on equity near its historic level while diversifying production. The company needs overseas facilities to become commercially integrated plants, rather than politically required insurance assets.

For the next five years, the question is whether TSMC preserves its process cadence through N2P, A16 and A14 while AI compute broadens from a handful of customers and training clusters into inference, enterprise systems, consumer devices and sovereign infrastructure. A broader customer set would improve growth quality. Dependence on a few accelerator programs would increase cyclicality even if absolute demand remained large.

The company becomes a better investment when price and operating evidence align. A decline toward NT$1,850–2,050, with gross margin still above 60%, N2 on schedule and no deterioration in geopolitical conditions, would provide a conservative margin of safety. It could also become more attractive without a price decline if owner earnings rise enough to bring forward P/E below 20 times while free-cash-flow conversion improves.

The judgment should be overturned negatively if gross margin falls below 58% for two consecutive quarters, advanced-node share drops below 70%, a major leading-edge customer moves a flagship product to another foundry, or capex remains above 45% of revenue while free cash flow stays below 40% of net income. A material escalation in Taiwan-Strait or export-control risk would require immediate reassessment regardless of quarterly fundamentals.

Bull reasons:

• Q2 2026 revenue grew 36% year over year while gross margin reached 67.7%, showing that advanced-node demand is producing operating leverage rather than revenue without profit.
• TSMC’s foundry share rose to approximately 72.3% in Q1 2026, leaving it larger than the next four competitors combined.
• HPC reached 66% of Q2 revenue and grew 20% sequentially as hyperscalers continued to raise infrastructure spending.
• N2 entered volume production in Q4 2025 with good reported yields and already contributed 3% of Q2 wafer revenue.
• Trailing EPS rose roughly 16% from the prior report’s comparable base while the ordinary share rose only 8.5%, compressing the multiple through earnings growth.

Bear reasons:

• The ten largest customers contribute 78% of revenue, with the top two at 36%, exposing the earnings base to a small number of AI and premium-device programs.
• Management expects three-to-four points of N2 margin dilution and an overseas-fab drag that may also widen to three-to-four points.
• The US$60–64 billion capex plan is equivalent to approximately NT$1.95–2.07 trillion, placing pressure on free-cash-flow conversion and future returns on capital.
• The trailing P/E remains about 48% above its reported ten-year average, leaving room for multiple contraction if AI growth normalizes.
• Taiwan concentration creates a low-probability loss path whose impact could exceed the downside represented by any conventional earnings scenario.

The first pre-mortem begins in 2027. Hyperscaler AI capex stops growing after two years of extraordinary expansion, customers digest accelerator inventories, and advanced-packaging lead times normalize. Leading-edge utilization falls from near-full levels toward 80–85% just as N2 and overseas depreciation rise. Gross margin drops from the mid-60s to 54–56%, EPS falls toward NT$80–85, and the market applies a 15-times cyclical P/E. The implied price is NT$1,200–1,275, approximately 47–51% below the current close.

The second pre-mortem begins with a technology and policy combination. Samsung establishes commercially credible sub-2-nanometer yields for one or two large customers while US policy requires more advanced production outside Taiwan at prices that do not fully compensate TSMC. Overseas costs and duplicated R&D pull gross margin toward 50%, free cash flow falls sharply and a major customer directs 15–20% of its new volume to a second source. A 14–16-times multiple on NT$75–85 EPS produces NT$1,050–1,360. A severe Taiwan event could cause a larger loss or interrupt tradability entirely; that tail is not bounded by this price calculation.

The final judgment is that TSMC remains the strongest manufacturing franchise in global semiconductors, and the evidence since May materially improves the valuation case. The old conclusion that a roughly 29.5-times multiple was plainly above fair value no longer holds because the underlying earnings denominator changed. Current trailing P/E is about 28.1 times, while the internally estimated current-year multiple has compressed to approximately 23–24 times.

The current quote nevertheless sits near conservative fair value rather than below it. Q2 margins are unlikely to be the permanent floor, customer concentration is rising, annual capex has moved close to NT$2 trillion and Taiwan risk deserves a real discount. The share offers a credible base-case return over three to five years, but a limited margin of safety under flat earnings or an ordinary AI-capex correction.

【Company-profile scores】

• Fundamental quality: high
• Growth: high
• Moat: strong
• Financial soundness: strong
• Management credibility: high
• Valuation attractiveness: medium
• Risk level: high
• Suitable investor type: long-term growth investors able to tolerate semiconductor cycles and geopolitical tail risk

【Investment rating】

• Rating: Hold
• One-line thesis: Earnings have outrun the price, but the current quote lacks a conservative margin of safety against margin and geopolitical tail risk.
• Current-price classification: acceptable hold
• Whether to wait for a better price: yes for a full-sized new position
• Buy trigger: price at or below NT$2,050, consolidated gross margin at least 60%, N2 ramp on schedule and no material escalation in export controls or Taiwan-Strait conditions
• Opportunity cost of waiting: further AI estimate upgrades or successful N2 pricing could move fair value upward before the share reaches the preferred entry range
• Target holding horizon: three to five years
• Expected annualized return, conservative scenario: 0–3%
• Expected annualized return, base scenario: 8–12%
• Expected annualized return, optimistic scenario: 15–20%
• Max-loss risk: approximately 50–60% in an AI inventory correction combined with gross margin near 55% and a P/E contraction to 14–15 times; a severe Taiwan event could exceed that range
• Reassessment trigger: gross margin below 58% for two consecutive quarters
• Reassessment trigger: advanced-node share below 70% or N2 below 5% of wafer revenue by Q4 2026
• Reassessment trigger: capex above 45% of revenue while free cash flow remains below 40% of net income
• Reassessment trigger: top-two customer concentration above 45% or a flagship process migration to a competitor
• Reassessment trigger: a material new export restriction, blockade risk or interruption to Taiwan production

【Ideal Buy Price】1,850–2,050 TWD

The range sits approximately 20% below the midpoint of the NT$2,300–2,550 conservative fair-value scenario and is supported by both the conservative DCF and forward-P/E approaches.

Acceptable hold price: NT$2,350–3,150. This range is approximately within 15% of the central base value and contains the current NT$2,425 close.

Clearly overvalued price: NT$4,250–4,700. The lower boundary is just over 10% above the NT$3,850 upper end of the optimistic fair-value scenario.

【Valuation Range】

• current: 2,425 TWD (close as of 2026-07-31)
• bear (conservative · ideal buy zone): [1,850, 2,050]
• base (fair · acceptable hold zone): [2,350, 3,150]
• bull (optimistic · above the clearly-overvalued line): [4,250, 4,700]

Research uncertainties remain in five areas. TSMC does not disclose maintenance capex, so the owner-earnings calculation relies on an explicit sustaining-capital estimate. Customer names behind the reported concentration figures are not disclosed. Leading-edge pricing is negotiated and cannot be observed directly. A precise Taiwan-conflict probability would be spurious, so the model uses wider ranges and higher discount rates instead. The original Taiwan IPO price and proceeds could not be verified from an accessible primary disclosure and were deliberately omitted.

The core source base consists of TSMC’s 2025 annual report and Form 20-F, its Q1 and Q2 2026 financial statements, presentations and call transcripts, official monthly-revenue data, the company’s financial calendar and technology pages, market-price reports for the Taiwan ordinary share, Taiwan central-bank currency data, competitor filings, US government policy disclosures and TrendForce foundry-share estimates.

Other tickers mentioned

• 005930.KS: Samsung Electronics is the principal integrated leading-edge foundry challenger.
• INTC.US: Intel is rebuilding its process road map and external foundry business with US policy support.
• GFS.US: GlobalFoundries is the main listed Western specialty-foundry comparison.
• 2303.TW: United Microelectronics competes in mature and specialty process technologies.
• 00981.HK: SMIC is China’s largest foundry and a policy-supported mature-node and selected advanced-node competitor.
• NVDA.US: NVIDIA is a major source of AI accelerator demand and an important inferred customer exposure.
• AAPL.US: Apple historically supplied large leading-edge smartphone volumes that helped underwrite TSMC node ramps.
• MSFT.US: Microsoft’s infrastructure investment is a key indicator of AI semiconductor demand.
• GOOGL.US: Alphabet’s raised 2026 capex guidance supports the near-term AI demand outlook.
• AMZN.US: Amazon and AWS form part of the hyperscaler capital-spending cycle tied to advanced compute.
• AMD.US: AMD relies on advanced foundry capacity for high-performance processors and accelerators.
• ASML.US: ASML’s EUV systems illustrate the bargaining power and scarcity of leading-edge equipment suppliers.

This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.