The Harder Monopoly Has the Worse Margin: Where the Money Actually Sticks in the Semiconductor Chain

Taiwan Semiconductor Manufacturing kept 59.9 cents of gross margin on every dollar of revenue in 2025. ASML, the only company on the planet that builds EUV lithography machines, kept 52.8. Put those two numbers side by side and the foundry looks like the better business than the monopoly. It is not. The gross margin line simply stops before the interesting part.

In that same year TSMC pushed 40.9 billion dollars into new fabs — roughly a third of revenue. ASML spent 1.57 billion euros on property, plant and equipment, or 4.8 percent. Subtract that from the gross margin and ASML is left with 48 points against TSMC’s 26. The ranking flips, and it does not flip narrowly. It flips by a factor of 1.8.

This is not an accounting quibble. It is the question an investor should actually be asking of a value chain: not who reports the highest margin, but who gets to keep it without handing it straight back to concrete, cleanrooms and tooling. The semiconductor market did somewhere between 791.7 and 795.6 billion dollars in 2025 depending on whether you follow the Semiconductor Industry Association or World Semiconductor Trade Statistics — up around 26 percent, one of the strongest years the industry has ever recorded. Whether money was made is not in doubt. Where along the chain it stopped moving is.

Eight links, wildly unequal economics

A chip is made along a chain that breaks cleanly into links, and every link has completely different economics. It starts with architecture: Arm licenses instruction sets and design blocks and collects a fee on every chip shipped. Behind that sit the design tools without which nobody draws a modern chip at all — Synopsys and Cadence carve up that market between them, with three vendors accounting for more than 90 percent of revenue. Then come the fabless designers, Nvidia foremost. Then the equipment makers that fill the fabs: ASML for lithography, Applied Materials, Lam Research and Tokyo Electron for deposition, etch and clean. Then the foundries, dominated by TSMC. Then memory. And at the very end, packaging and test — the business of ASE and Amkor.

Line up the gross margins of those links for each company’s most recently reported fiscal year and you get a picture that has very little to do with the way the sector is usually described.

Link in the chain	Company	Revenue	Gross margin
Architecture / IP	Arm Holdings (FY Mar 2025)	about $4.0bn	97.0%
Design tools	Synopsys (FY Oct 2025)	$7.05bn	81.4%
Fabless design	Nvidia (FY Jan 2026)	—	71.1%
Foundry	TSMC (2025)	about $122bn	59.9%
Lithography	ASML (2025)	€32.7bn	52.8%
Other fab equipment	Lam Research (FY Jun 2025)	$20.6bn	49.9%
Memory	Micron (FY Aug 2025)	$37.4bn	about 40%
Packaging and test	Amkor (2025)	$6.71bn	14.0%

Eighty-three percentage points separate the top of that table from the bottom. Both companies work on the same physical object. Arm never touches it; Amkor holds it in its hands. That is precisely the point.

Gross margin measures cost structure, not power

The obvious reading is that the further up the chain you sit, the more pricing power you have. Directionally that is true, but as an explanation it is lazy, and in one crucial place it is simply wrong.

ASML holds 100 percent share of the EUV equipment market. Not 80, not 95 — every single machine anywhere on earth that patterns a chip at 5, 3 or 2 nanometers comes out of Veldhoven. Across lithography as a whole the share is around 94 percent. Modern industrial history offers very few monopolies that tight. And ASML still reports a lower gross margin, at 52.8 percent, than its customer TSMC at 59.9 percent, and than Nvidia at 71.1 percent.

The reason is mundane and constantly overlooked. Gross margin tells you how much of revenue survives direct cost of goods. It is therefore first and foremost a statement about a business’s cost structure, not about its bargaining position. Monopoly or not, ASML is a machine builder. Every EUV tool contains tens of thousands of bought-in components, including the most precise optics in existence. Those purchases land in cost of goods and compress the gross margin whether or not ASML can dictate the final price. Software carries no such cost, which is why Synopsys sits at 81 percent — and Arm, which ships not even software but rights, at 97.

Reading pricing power off the gross margin line means confusing “how expensive is this to make?” with “who is able to say no?” Those are different questions, and for returns the second one matters more.

The calculation that actually sorts the sector

One simple adjustment straightens the picture out. Take the gross margin and subtract the share of revenue the company carried back out of the door as capital expenditure in the same year. What remains is a rough answer to a better question: how much of each revenue dollar does this business get to keep once you account for what it must spend simply to hold its position?

Company	Gross margin	Capex 2025	as % of revenue	Difference
ASML	52.8%	€1.57bn	4.8%	48.0 points
TSMC	59.9%	$40.9bn	33.5%	26.4 points
Amkor	14.0%	$0.90bn	13.5%	0.5 points
Micron (FY Aug 2025)	about 40%	$15.86bn	42.4%	minus 2.4 points

That fourth row deserves a pause. Fiscal 2025 was Micron’s best memory year in over a decade: revenue up roughly half to 37.4 billion dollars, gross margin up 17 percentage points, data center at 56 percent of revenue, high-bandwidth memory alone approaching 2 billion dollars in the fourth quarter. And even in that exceptional year, capex of 15.86 billion dollars exceeded the entire gross margin. Guidance for the current fiscal year is around 20 billion.

Amkor is starker still, just on a delay. In 2025, a 14.0 percent gross margin sat against a 13.5 percent capital intensity — half a point of daylight. For 2026 the company has guided 2.5 to 3.0 billion dollars of capex, mostly for the new Arizona facility. Against a revenue base of just under seven billion, that is north of 40 percent — on a 14 percent gross margin.

What this calculation is not

An objection belongs here, and it is correct. In accounting terms you cannot simply subtract capex from gross margin. Depreciation on prior investment already sits inside cost of goods and therefore inside the gross margin itself. Subtracting current capex on top of that double counts.

So the measure is not an earnings figure and does not pretend to be one. It is a cash figure, and it answers a different question: how much money leaves the business again just so it can stay where it is? It has two known weaknesses. It flatters companies at the end of an investment cycle and punishes those in the middle of building — Amkor’s Arizona plant is growth, not maintenance. And it ignores that part of any capex budget is genuine capacity expansion that will carry future revenue.

Neither weakness changes the ranking here. TSMC has invested between 25 and 35 percent of revenue for more than a decade and plans to do so again in 2026 on a record budget of 52 to 56 billion dollars. That is not an outlier year, it is the business model. Memory makers structurally invest above their gross margin as soon as the cycle turns. And ASML has not had to come anywhere near double-digit capital intensity in any recent year. The gap is durable, not incidental.

Where ASML’s power does show up

If the gross margin does not capture the monopoly, what does? Three things, all disclosed in the same filings.

First, the order book. ASML closed 2025 with a backlog of 38.8 billion euros — more than a full year of revenue, pre-committed. In the fourth quarter alone it took 13.2 billion euros of new bookings, of which 7.4 billion were EUV systems. No company facing serious competition gets customers to commit more than twelve months of revenue in advance.

Second, the price list. A low-NA NXE:3800E runs around 180 million dollars. The new high-NA EXE:5200 generation sits at 350 to 400 million. ASML is charging better than double — and customers pay, because the arithmetic still works for them: an exposure on the high-NA tool costs roughly 2.50 dollars against 1.10 on the previous generation, but it removes process steps that would otherwise have to be run several times over. Intel is deploying the tool for its 14A node; Samsung took its first unit in October and has a second due in the first half of 2026.

Third, the research budget. ASML spent just under 4.7 billion euros on research and development in 2025, with more than 16,000 of its 44,000-plus employees working there. That is the real moat — not the patent, but the distance a competitor would have to close while ASML keeps moving.

The bottleneck you cannot buy

Here is the part of the story that should bother anyone who likes owning bottlenecks. The tightest constraint in the entire chain is not ASML. It sits one level above ASML — and it is not investable.

ASML’s own annual report describes Carl Zeiss SMT as its sole supplier of lenses, mirrors, illuminators, collectors and other critical optical components under an exclusive arrangement. The language in the filing is unusually blunt: if Zeiss were unable to supply over a prolonged period, ASML would effectively cease to be able to conduct its business. ASML further concedes that the number of systems it can build is capped by Zeiss SMT’s production capacity — not by its own, and not by demand.

Those optics are made in exactly two places on earth: Oberkochen in Baden-Württemberg and Wetzlar in Hesse, both in Germany. The mirrors are finished to tolerances smaller than the diameter of a single silicon atom. Carl Zeiss SMT belongs to Carl Zeiss AG, which in turn belongs to the Carl Zeiss Foundation. There is no share class. ASML itself paid one billion euros in 2016 for a 24.9 percent stake in Carl Zeiss SMT, closing in June 2017 — that was all that was available, and all that is going to be.

The second critical supplier is also German. The carbon dioxide laser that vaporizes the tin droplet from which the EUV light is generated comes from Trumpf in Ditzingen, near Stuttgart. Trumpf is a family business owned by the Leibinger family, likewise unlisted.

So two of the three most critical components in the most important tool in the digital economy are built in southwestern Germany, and a public-market investor can own neither. That is more than a curiosity. It means ASML stock is the only liquid instrument through which this bottleneck can be expressed at all — with the caveat baked in that ASML hands part of the rent to suppliers you cannot buy.

The objection worth taking seriously: memory

Against the thesis that margin lives at the ends of the chain, 2025 offers one very large counterexample. SK Hynix — a memory maker, and therefore squarely in the capital-intensive middle — booked 47.21 trillion won of operating profit on 97.15 trillion won of revenue for the calendar year. That is a 49 percent operating margin, rising to 58 percent in the fourth quarter. SK Hynix therefore earned considerably more per unit of revenue at the operating line than ASML, whose 2025 net margin was 29.4 percent.

Waving that away as an anomaly would be too convenient. High-bandwidth memory is not a commodity memory product. It is technically demanding, capacity is scarce, the customers are few and well funded, and it accounted for an estimated 40 percent of DRAM revenue in 2025. For as long as that scarcity holds, part of the memory market behaves like a bottleneck business and gets paid like one.

The difference is durability. ASML’s position rests on two decades of research, an exclusive optics supplier and a customer base that orders a year ahead. SK Hynix’s position rests on a capacity lead of roughly two years, and the competitive response to a capacity lead is to build capacity. Micron has guided to 20 billion dollars. That is the exact mechanism that ends memory cycles. It is possible this time is different, because AI infrastructure demand is compounding faster than supply can answer. But that is a bet on an imbalance, not a bet on a structure — and portfolios should price the two differently.

Three scenarios through 2029

The probabilities below are a reasoned view, not a calculation. Their main purpose is to show what would tell you which world you are in.

Scenario	What happens	How you would know
The bottleneck holds (about 45%)	High-NA lands at Intel 14A, TSMC A14 and Samsung SF1.4. ASML grows ahead of the market and the gross margin works its way above the 51 to 53 percent range guided for 2026 once the new generation is past its ramp costs.	Backlog stays above €35bn. EUV holds above 50% of quarterly bookings. No slippage in the A14 and 14A roadmaps.
The middle catches up (about 35%)	The binding constraint migrates from patterning to advanced packaging and memory. Whoever can stack die and feed them bandwidth sets the pace. Amkor and ASE earn well above their historical margins through the cycle; ASML grows merely in line with the market.	Amkor gross margin sustained above 18%. Packaging capex budgets growing faster than total industry equipment spending.
The cycle turns (about 20%)	Digestion of the AI build-out begins. Industry equipment spending falls well below the roughly $135bn expected for 2026, memory pricing breaks, TSMC trims. ASML’s backlog absorbs perhaps twelve months before Veldhoven feels it — the middle of the chain goes into the red first.	Two consecutive quarters of declining ASML bookings. TSMC cutting its capex plan mid-year. Falling memory prices alongside rising inventories.

What it means for a portfolio

None of this is a recommendation, but it does settle three practical questions.

The first concerns the sector ETF. Buying semiconductors through an index fund means buying monopoly rents and cyclical commodities bundled into a single security. Arm and Amkor are 83 gross margin points apart and respond to the same macro impulse in completely different ways. An index blurs exactly the distinction this piece is about. That is not an argument against index investing; it is an argument for knowing what you own when you deliberately overweight the sector.

The second concerns valuation. A business that keeps 48 points per unit of revenue deserves a different multiple from one that keeps 26, and both deserve a different multiple from one that keeps nothing. A surface-level price-to-earnings comparison between ASML and TSMC therefore says very little until you state how much capital stands behind each company’s earnings.

The third concerns how you actually hold the position, and this is where investors routinely lose money for reasons unrelated to the analysis. Dividends from Dutch companies carry 15 percent Dutch withholding tax; for a US taxable account that is generally creditable against US tax via the foreign tax credit, and holding ASML in an IRA typically forfeits that credit entirely — the withholding still comes off, but there is no US tax to offset. TSMC is worse: on distributions to holders of its American depositary shares, Taiwan withholds 21 percent, and because there is no US-Taiwan tax treaty there is no reduced treaty rate to claim. The gross yield you see quoted on a TSMC ADR is therefore about a fifth larger than the yield you receive. For a low-yield, capital-appreciation holding this is a rounding error; for anyone buying the chain for income, it is not.

The counterarguments

Three objections deserve to be stated plainly, because each could damage the thesis.

The first is political. ASML’s monopoly is profitable partly because export controls largely close the Chinese market to its most advanced machines — which simultaneously maximizes China’s incentive to build its own. A state-funded competitor does not have to earn a return on capital. It only has to work. The timeline is long, but it is not infinite, and the day a working Chinese EUV tool is demonstrated would be a valuation event.

The second is concentration. TSMC took roughly 70 percent of the pure-play foundry market in 2025, with advanced nodes at 7 nanometers and below accounting for 74 percent of wafer revenue. A customer base that concentrated is both a support and a risk for ASML. If TSMC cuts, half the market cuts.

The third is a valuation objection, and it is the most uncomfortable. Everything in this article is public. The annual reports are freely available, the market shares are widely cited, the Zeiss dependency sits in the risk section of a mandatory filing. It would be naive to assume the market has not read them. The question is therefore never whether ASML is a better business than Amkor — that is obvious — but whether the price gap between them already fully reflects the quality gap. No value chain analysis answers that. Only a valuation run at the day’s price does.

The bottom line

The semiconductor chain turned over close to 800 billion dollars in 2025 and is still growing. But it is not a sector. It is a sequence of very different businesses that happen to touch the same product. At the front end — rights and tools, and lithography just behind them — sit businesses that keep most of their margin. In capital-intensive manufacturing sit businesses that hand most of their margin back every year just to stay in the race. And right at the end, in packaging and test, a good year leaves half a percentage point.

Treat the chain as one bet and you are betting on growth. Treat it as eight bets and you get to choose which ones you want. Gross margin alone will not tell you which those should be. The question of how much capital stands behind each point of margin comes a lot closer.