Report contents
Industry landscape
Where the company competes: leading-edge foundry, one firm at 72.5%
Taiwan Semiconductor Manufacturing is a pure-play foundry. It manufactures chips that other companies design, selling capacity by the wafer at prices that vary by node, contract and customer. It does not sell chips under its own brand at scale, which separates it from Samsung, which builds memory and logic and also sells its own devices, and from Intel, which sells its own processors. Upstream sits a narrow equipment and materials layer: ASML for EUV lithography (the scanners that print the smallest circuit features), Applied Materials, Lam Research and Tokyo Electron. Downstream sit fabless designers such as Apple, Nvidia, AMD, Qualcomm and MediaTek, who take finished wafers to outsourced assembly and test. The company's pricing leverage therefore runs through a handful of very large customers who need the most advanced nodes and have few qualified alternatives.
The arena that matters is leading-edge logic foundry: high-volume production at 7nm and below, sold to fabless designers and to a smaller set of systems companies that design their own accelerators. That market is global. Designs originate in the US, Taiwan, China, Korea, Japan and Europe, and wafers ship worldwide. Where the fabs sit is becoming a political question, with subsidized capacity in Arizona, Kumamoto and Dresden, but customers still qualify a supplier on node, yield and schedule, not on the flag on the fab. So far the evidence only shows geography changing where capacity is built, not who wins the design.
Market structure: concentrated at the leading edge, fragmented behind it
TrendForce's September 2026 tally put combined second-quarter revenue for the world's ten largest foundries at $53.49 billion, up 11.5% from the prior quarter. TSMC took 72.5% of that pool, with revenue near $40.2 billion, up 12.1% quarter over quarter. TrendForce attributed the gain to fully booked 5/4nm and 3nm capacity for AI server GPUs and accelerators, iPhone inventory build, and the first revenue contribution from 2nm. The firm also reported that 3nm became TSMC's largest single revenue node.
| Foundry, 2Q26 | Revenue (USD bn) | QoQ | Share of top-10 |
|---|---|---|---|
| TSMC | 40.2 | +12.1% | 72.5% |
| Samsung Foundry | 3.26 | +1.8% | 5.9% |
| SMIC | ~3.0 | +20% | 5.4% |
| UMC | 2.18 | +12.7% | 3.9% |
| GlobalFoundries | 1.79 | +9.3% | 3.2% |
The concentration is starker than the table suggests. Only three firms have shipped volume production at 5nm or below: TSMC, Samsung and Intel. Intel does not appear in TrendForce's top-ten ranking for the quarter, and Samsung, after more than a decade of foundry investment, holds 5.9%. That gap is the cleanest evidence on entry barriers available. Leading-edge fabs cost tens of billions per site, depend on a single EUV scanner supplier, and require years of yield learning before a design wins volume. Just as binding is the design ecosystem: process design kits, verified IP libraries and two-to-three-year customer design cycles lock a chipmaker into a node before a single wafer is priced. Barriers to entry at the leading edge are high, and the number of credible suppliers is small enough to count on one hand.
Behind the leading edge the structure inverts. Mature-node foundry, roughly 28nm and above, counts UMC, SMIC, GlobalFoundries, HuaHong, Tower, VIS, Nexchip and PSMC among its participants, and their shares sit in the low single digits. TrendForce notes tight mature-node capacity supported pricing in 2026, with advance procurement across PCs, notebooks and consumer electronics tightening supply further. That is a cyclical read, not a structural one: eight or more meaningful suppliers competing for commodity process work is a recipe for weak pricing over a full cycle. TSMC participates there, but its revenue mix has migrated upward, and the mature business is not where its advantage sits.
One definitional caveat matters for the moat work that follows. The 72.5% is TSMC's share of revenue among the ten foundries TrendForce ranks, not of all logic manufacturing. Intel's internal fabs, Samsung's captive logic capacity and Chinese state-backed capacity that falls outside the ranking all produce logic chips. Measured against total logic wafer demand, TSMC's share would be lower, though the direction of the leading-edge number is not in doubt. The geography of competition is global; the boundary of the measured market is a choice, and it is worth keeping that choice visible when judging durability.
Competitive advantage
Does the company have a moat? Yes, at the leading edge
Moat: Strong. TSMC took 72.5% of the top-ten foundry pool in the second quarter of 2026, and it has earned after-tax returns on invested capital of roughly 20% to 27% in each of the last five fiscal years. A concentrated market whose leader keeps earning well above its cost of capital is the best evidence available that entry barriers are real, because those are the two things competition normally wears down. The scope is narrower than the share number suggests: this moat covers leading-edge logic, about 7nm and below. In mature nodes TSMC competes against eight or more suppliers and earns ordinary, cyclical returns.
Market position: 72.5% of the ranked pool and still gaining
TrendForce put second-quarter 2026 revenue for the ten largest foundries at $53.49 billion. TSMC's $40.2 billion works out to 72.5%, up 12.1% quarter over quarter against an 11.5% gain for the pool, so it took a little share in a quarter when everyone grew. Samsung Foundry, after more than a decade of dedicated investment, held 5.9%. SMIC held 5.4%, UMC 3.9% and GlobalFoundries 3.2%. TSMC's quarterly revenue is roughly twelve times Samsung's foundry revenue, and Intel, which has both the process technology and the capital to compete at the leading edge, did not make the top ten at all.
The five-year stability test is harder to document than the snapshot. The quarterly series in hand shows 72.3% in the prior quarter and 72.5% now, a move well inside the 2-percentage-point band that would point to a very strong barrier and nowhere near the 5-point swing that would suggest no barrier. But two quarters is not five years, and a full annual TrendForce series back to 2021 is not publicly compiled in the documents reviewed. That gap matters for the rating and is worth stating plainly.
What the structure does establish is persistence in the ranking. Samsung has been the named second source at the leading edge for years and still sits below 6% of the pool. Intel's foundry push has not produced a top-ten position. A rival cannot be that large, spend that much and stay that small unless something other than capital decides the outcome. So the trend evidence points to a stable or improving position, with the caveat that the measurement window is short and the share base is the ten ranked foundries, not all logic manufacturing.
Market position: Strong.
Profitability: returns widened through the last two upcycles
| Fiscal year | Revenue (TWD bn) | Gross margin | Net margin | ROE | ROIC (est.) |
|---|---|---|---|---|---|
| 2021 | 1,587.4 | 51.6% | 37.3% | 29.5% | 21% |
| 2022 | 2,263.9 | 59.6% | 43.9% | 38.7% | 27% |
| 2023 | 2,161.7 | 54.4% | 39.4% | 26.4% | 20% |
| 2024 | 2,894.3 | 56.1% | 40.0% | 29.7% | 21% |
| 2025 | 3,809.1 | 59.9% | 44.6% | 35.1% | 27% |
ROIC is an estimate built from reported figures: EBIT multiplied by one minus the effective tax rate, divided by total equity plus total debt. It excludes the cash offset, which would push the figure higher. On that basis the five-year average is about 23%, at the top of the 15% to 25% band that marks a clear competitive advantage; the 6% to 8% band that signals none is not close.
The record is not one good year. Margins rose in the 2021-22 upcycle, held through the 2023 downturn at a 54.4% gross margin and a roughly 20% ROIC, then rose again. The first nine months of 2026 produced TWD 3,897.9 billion of revenue, already more than all of 2025, with a 66.4% gross margin and a 51.1% net margin in the September quarter. The second quarter's 67.7% gross margin was up 9.1 percentage points year over year, which is the number that separates pricing power from a cyclical bounce.
Peer margin comparison is thin. Samsung reports foundry inside a much larger device business and does not disclose its foundry margin, and the income statements for UMC and GlobalFoundries are not in the material reviewed. The documented gap is in scale and in TSMC's own return trend; the margin gap against rivals is inferred from that scale, not measured directly.
Profitability: Strong.
Overall rating: Strong, with mature nodes excluded
Overall rating: Strong. Both tests pass, and profitability carries the extra weight. Returns have stayed in the high-teens to high-twenties for five straight years while capital intensity rose, which is what a protected position looks like. Three things would change the assessment: ROIC sliding toward 10% through a downcycle, leading-edge share falling below 60% for several quarters, or a second supplier qualifying high-volume 2nm production at a flagship customer.
Where the advantage comes from and how long it can last
The advantage is a fixed-cost machine that gets cheaper per wafer as volume grows, bolted to customers who cannot move a design without losing a year. Process technology is the visible output of that machine, not an independent source of protection; taken alone it is the weakest of the three. That ordering matters for durability, because it explains why well-funded rivals have failed to close the gap with capital alone.
Scale: fixed costs spread over a base no rival can match — Strong
The company raised 2026 capex guidance to $60-64 billion, or 35% to 38% of revenue, and told investors on its July call that spending over the next three years will be significantly higher again. A leading-edge fab's cost is set by tool count, cleanroom area and process R&D, almost none of which scales down with wafer volume. Two firms that spend the same on a fab carry per-wafer depreciation inversely proportional to the wafers they run through it. With roughly twelve times Samsung's foundry revenue, TSMC spreads those fixed costs over a base its nearest rival cannot approach, which lowers unit cost and lets it fund the next node earlier.
That starts a loop: lower unit cost funds earlier capacity, earlier capacity wins the design, the design brings volume, and volume lowers unit cost again. The loop is self-reinforcing only as long as the leader keeps spending, which is why the capex number is a moat indicator and not just a cash drain. The evidence that it is working is the margin: a 67.7% gross margin in the second quarter of 2026, up 9.1 points year over year, while depreciation was climbing. If scale were merely a size claim, margins would compress as the asset base grows.
The machine has a price. Operating cash flow was TWD 2,275 billion in fiscal 2025 against free cash flow of TWD 992 billion; in fiscal 2024 the pair was TWD 1,826 billion and TWD 870 billion. Capex absorbs close to half of what the business throws off, and the 2026 step-up widens that gap: TWD 951.8 billion of operating cash flow and TWD 510.1 billion of free cash flow in the September 2026 quarter alone.
The limit is political. Scale deters a commercial entrant that has to earn a return; it does not deter a government willing to absorb a decade of sub-scale losses. Intel and SMIC can both do that. Scale raises the cost of entry, and on its own it would not be enough.
Customer captivity: moving a design costs about a year — Fairly Strong
A fabless designer commits to a process 18 to 36 months before volume production. The process design kit, the verified IP library, EDA sign-off, yield characterization and advanced packaging are all qualified against one vendor's process. Moving that design to a second source means a re-spin, a fresh qualification cycle and a year of schedule. For a customer shipping AI accelerators on an annual cadence, a year of delay is lost share at the customer, not just an internal cost. Search costs are also real: only three firms can quote a 5nm-or-below process at volume, so there is not much of a market to shop.
Pricing is where captivity gets tested, and TSMC has been raising price. Nikkei and Reuters reported in July 2026 that the company disclosed a 5% to 10% increase for 2027 across advanced and mature nodes. Extracting that from the largest customers without losing allocations is the practical proof that switching is expensive for them.
The rating stops at Fairly Strong because captivity is per design, not per company. The buyer list is short and sophisticated, every one of them has a supply-security reason to qualify a second source, and a customer can move a mature-node part without touching its flagship processor. Bargaining power at renewal is real.
Process technology: a lead measured in nodes, not a wall — Moderate
Yield learning is genuine proprietary know-how and it compounds, which is why 3nm became the largest single revenue node in the second quarter of 2026 and 2nm shipped its first revenue in the same quarter. But the tools are not proprietary: ASML sells the same EUV scanners to Samsung and Intel. Recipes travel with engineers, and patents cover little of what actually determines yield. Samsung has shipped 3nm since 2022 and still holds 5.9% of the pool, which tells you the gap sits in yield and execution. Those gaps close when a competitor is willing to burn capital for long enough. The lead is probably one to two nodes, and it is the most contestable part of the story.
Durability: the open question is a second source at 2nm
The substitute that matters is not a different product. It is a credible second supplier at the leading edge. Chiplets and advanced packaging change how compute is assembled, but the packaging itself increasingly runs through TSMC, so the shift has not yet diluted its pricing power. Chinese capacity is a mature-node problem: SMIC grew 20% quarter over quarter in the same period, which pressures commodity process pricing rather than the leading edge.
The 2027 price increase is the live test of cross-price sensitivity. If customers accept it and stay, the moat holds through the 2nm generation. If a hyperscaler moves a flagship design to Samsung or to Intel's advanced process at volume, that would be the first real crack, and it would show up in share before it showed up in margins. The concentration of leading-edge capacity in Taiwan is a risk to output, not to pricing power against rivals, and the Arizona expansion has been announced without a schedule attached, so it does not yet change the competitive arithmetic on either side.
So far the evidence supports durability over the next two to three years. Whether it extends past the 2nm generation depends on a question the public record does not settle: whether any rival can convert state funding and purchased tools into volume yield at the leading edge.
Generated by DeepVal AI. Data and models may contain errors or be outdated. For research only; not personalized investment advice.
Continue reading the full report
- Comparison and summaryHow Taiwan Semiconductor Manufacturing compares with its closest rivals, and our moat rating
Free to sign up · 20 research points · unlocking this report uses 10
20 points cover 2 full reports. No card needed.
Research reflects its generation date. Preview the first two modules.
More on Taiwan Semiconductor Manufacturing
Taiwan Semiconductor Manufacturing
Business, moat and key questions
Open company page