Every AI chip passes through one company's machines

TL;DR. ASML holds 100% of the extreme ultraviolet lithography market, and roughly 94% of lithography overall on one count. Nikon and Canon exited EUV more than a decade ago and no competing machine exists. No chip below about 7nm can be manufactured without ASML's scanners, which means every Nvidia accelerator, every custom hyperscaler ASIC and every HBM stack in every AI data centre was made on them. The company reported €32.7 billion in 2025 net sales with a €38.8 billion backlog. And it has held this position for fourteen years without pricing like a monopolist, with pricing that tracks engineering cost rather than scarcity. The concentration is usually discussed as a country risk. It is more precisely one firm, in one country, dependent on one optics supplier.

---

Status: established. Market position, financial figures and product details are from company reporting and industry analysis. This article is descriptive and is not investment advice; nothing here evaluates any security. The characterisation of pricing restraint is an interpretation and is presented as one.

---

The position

ASML is the sole supplier of EUV lithography systems worldwide. Not the leader. The only one.

Nikon and Canon exited the technology more than a decade ago, leaving a market with one participant. Estimates of ASML's share of lithography overall run from 83% to 94% depending on the count, and its share of EUV specifically is 100%.

EUV uses 13.5-nanometre light to print circuit patterns, and without it transistors below roughly 7nm cannot be manufactured. That threshold is the entire relevant range for AI hardware.

So every Nvidia data centre GPU, every custom accelerator from Google, Amazon or Microsoft, every high-bandwidth memory stack feeding them, was produced on machines from one company.

The hyperscalers have announced capital expenditure in the hundreds of billions for 2025 and 2026. All of it flows through that chokepoint, and there is no alternative supplier to switch to, at any price, on any timeline.

The scale of the machines

A High-NA EUV system, the EXE:5200B, shipped to Intel in Q4 2025 at a price reported between $350 and $400 million per unit.

That price is itself a filter. TSMC said in April 2026 that it will skip High-NA through 2029, citing the cost. SK Hynix installed its first unit in late 2025 for HBM and advanced DRAM. Samsung is integrating it into a 2nm plan. Intel has committed to it for its 14A node.

Which produces an unusual situation: the customer choosing not to buy the most advanced tool is the largest and most successful foundry in the world, and its reason is price rather than capability.

ASML reported €32.7 billion in net sales for 2025, up 15.6%, with net income of €9.6 billion at a 52.8% gross margin, closing the year with a €38.8 billion backlog and record Q4 bookings of €13.2 billion, of which €7.4 billion was EUV.

The part worth pausing on

ASML has held an absolute monopoly on the most strategically critical machine in the global economy for fourteen years, and has not priced like one.

Industry analysis describes its pricing as tracking engineering costs rather than scarcity, and its conduct as closer to a research consortium than to a firm extracting rent from an unsubstitutable position.

That is genuinely unusual and it deserves explanation rather than admiration.

The likeliest explanation is symmetric dependence. ASML's top two customers account for around 38% of revenue, with TSMC alone at roughly 24% of net sales. Those are the only firms with the technical capability and capital to buy and operate these systems. A monopolist with three viable customers is not in the position the word usually implies.

A second is that extraction accelerates substitution. Nikon and Canon left because the engineering was brutal, not because it is impossible. Sustained rent extraction funds the research that ends the monopoly, and a firm with a fourteen-year lead has more to lose from inviting that than to gain from a price rise.

And a third is regulatory. A supplier at this chokepoint operating visibly as a monopolist invites intervention from every government that depends on it, which is currently all of them.

Where the concentration actually sits

Public discussion of AI hardware concentration is usually about Taiwan, and the geographic risk is real. But the more precise chokepoint is upstream of it.

A fab in Taiwan, Arizona, Japan or Germany still needs the same machines from the same company. Relocating fabrication changes the geography of the second stage and not the first.

And the concentration continues upstream. ASML's optics come from a single supplier, Zeiss, whose mirrors for these systems are among the most precisely manufactured objects ever made. The chain narrows rather than widens as you follow it back.

Which means the standard framing understates the problem in one way and overstates it in another. Understates, because moving fabs does not diversify the actual bottleneck. Overstates, because a chokepoint at a supplier in the Netherlands has a very different risk profile from one in a contested strait.

Three things this establishes

Diversification at the visible layer can leave the real dependency untouched. Building fabs in multiple countries is genuinely useful for several risks and does nothing about a single-supplier tool. Anyone assessing supply concentration has to follow the chain to where it is narrowest, not stop at the most discussed layer.

Monopoly position and monopoly conduct are separable. Fourteen years of unsubstitutable supply without scarcity pricing shows that the second does not follow from the first, and the reasons, symmetric dependence, substitution risk and regulatory exposure, are structural rather than a matter of character.

And the most advanced tool is not always adopted first by the most advanced customer. TSMC skipping High-NA through 2029 on cost, while Intel commits to it, inverts the pattern of the previous decade. Access to better lithography is necessary and not sufficient for process leadership, which is worth remembering whenever a capability is treated as a determinant.

What it does not establish

That ASML's position is unassailable. Nikon and Canon exited a hard engineering problem, not an impossible one, and state-funded programmes elsewhere are attempting it. Fourteen years is a long lead and not a permanent one.

That the restraint continues. The interpretation offered here is structural, and structures change. Nothing in the record guarantees the next fourteen years resemble the last.

That geography does not matter. Fab location bears on many risks, including tariffs, labour, energy and conflict exposure. The claim is narrower: it does not diversify this particular dependency.

And nothing about any company's valuation. This article describes a supply structure. It makes no claim about whether any security is fairly priced.

What is unresolved

Whether a competing EUV programme succeeds. Several are reportedly underway. None has produced a production system, and the engineering barrier is very high.

Whether High-NA becomes necessary or stays optional. If TSMC reaches competitive density at A14 without it, the tool's economics change substantially. If not, the largest foundry has a gap.

How exposed the chain is at the optics layer. Zeiss is a private company and its capacity constraints are not publicly detailed, which makes the narrowest point in the chain also the least visible.

And what happens to the backlog if capex pauses. A €38.8 billion backlog provides visibility and does not guarantee delivery timing if customers defer, which is the acknowledged sensitivity in the position.

The counter-argument

Calling this a chokepoint overstates fragility. ASML is a well-capitalised firm in a stable jurisdiction with a fourteen-year record of supplying its customers. A single supplier is only a risk if it fails or refuses, and there is no evidence of either. Sole supply and fragile supply are different things, and this article slides between them.

The restraint may be less remarkable than presented. Pricing at a level that sustains a very high margin, 52.8% gross in 2025, is not obviously restraint. It may simply be the profit-maximising price given customer concentration, in which case the behaviour needs no special explanation at all.

Following the chain upstream has no natural stopping point. Zeiss depends on suppliers, who depend on materials, who depend on mines. Every supply chain narrows somewhere, and singling out one layer as the real dependency is a choice rather than a finding.

And the geographic framing is not wrong, just incomplete. Advanced packaging, HBM and the workforce for leading-edge fabrication are all concentrated in East Asia, and those are genuine constraints that relocating fabs partially addresses. Dismissing the country framing understates what fab diversification does achieve.

The short version

ASML holds 100% of the EUV lithography market. Nikon and Canon exited more than a decade ago and no competing machine exists. Below roughly 7nm, nothing can be manufactured without these scanners, which is the entire relevant range for AI hardware.

So every Nvidia accelerator, every custom hyperscaler ASIC and every HBM stack in every AI data centre was made on machines from one company, and the hundreds of billions of announced capex all flows through that point with no alternative at any price.

The company reported €32.7 billion in 2025 net sales with a €38.8 billion backlog, and a High-NA system ships at $350 to $400 million per unit, a price at which TSMC has chosen to wait until 2029 while Intel commits.

And for fourteen years it has not priced like a monopolist, with pricing that tracks engineering cost rather than scarcity. The likeliest explanations are structural rather than admirable: its top two customers are 38% of revenue, extraction would fund the research that ends the lead, and visible rent-seeking at this position invites intervention from every government that depends on it.

The concentration is usually discussed as a country risk, and that framing points at the wrong layer. A fab in Arizona, Japan or Germany needs the same machines from the same firm. Relocating fabrication changes the geography of the second stage and not the first, and the chain narrows further upstream still, to a single optics supplier whose capacity nobody outside it can see.

Common questions

Does ASML really have a monopoly on EUV? Yes, in the literal sense. It is the only company that produces EUV lithography scanners, after Nikon and Canon exited the technology more than a decade ago. Its share of the EUV market is 100%, and its share of lithography overall is estimated between 83% and 94% depending on the count.

Why does that matter for AI specifically? Because EUV uses 13.5-nanometre light to print circuit patterns, and without it transistors below roughly 7nm cannot be manufactured. That threshold covers the entire relevant range for AI hardware. Every Nvidia data centre GPU, every custom accelerator from Google, Amazon or Microsoft, and every high-bandwidth memory stack feeding them was produced on ASML machines. The hyperscalers' announced capital expenditure flows through that single point with no alternative supplier at any price or timeline.

How expensive are these machines? A High-NA EUV system, the EXE:5200B, shipped to Intel in Q4 2025 at a reported price between $350 and $400 million per unit. That price is itself a filter on who adopts: TSMC said in April 2026 it will skip High-NA through 2029 citing cost, while SK Hynix installed its first unit in late 2025 for HBM and advanced DRAM, Samsung is integrating it into a 2nm plan, and Intel has committed to it for its 14A node.

Has ASML used its position to raise prices? Industry analysis describes its pricing as tracking engineering costs rather than scarcity, across fourteen years of unsubstitutable supply. That is an interpretation rather than a measurement, and the likeliest explanations are structural: its top two customers account for around 38% of revenue with TSMC alone at roughly 24%, so a monopolist with three viable customers is not in the position the word implies; sustained extraction would fund the research that ends the lead; and visible rent-seeking at this chokepoint would invite intervention from every government that depends on it.

Is the real risk Taiwan or the Netherlands? Both, at different layers, and the usual framing points at the wrong one for this particular dependency. A fab in Taiwan, Arizona, Japan or Germany needs the same machines from the same company, so relocating fabrication changes the geography of the second stage and not the first. The chain narrows further upstream still, to a single optics supplier whose capacity constraints are not publicly detailed.

Does building fabs elsewhere help at all? Yes, for other risks. Advanced packaging, high-bandwidth memory and the workforce for leading-edge fabrication are concentrated in East Asia, and fab diversification bears genuinely on tariff, labour, energy and conflict exposure. The narrow claim here is that it does not diversify the lithography dependency, because there is only one supplier regardless of where the fab stands.

Is a single supplier necessarily a fragile supplier? No, and conflating the two is the strongest objection to this article's framing. ASML is well capitalised, operates in a stable jurisdiction, and has supplied its customers for fourteen years without interruption. Sole supply is a risk only if the supplier fails or refuses, and there is no evidence of either. What sole supply does mean is that no amount of spending buys an alternative, which is a different property from fragility and matters for different reasons.

Could a competitor emerge? Possibly, and not quickly. Nikon and Canon exited a brutally hard engineering problem rather than an impossible one, and state-funded programmes elsewhere are reportedly attempting it. None has produced a production system. Fourteen years is a long lead and not a permanent one, and the barrier is high enough that the timeline for any credible alternative is measured in many years rather than product cycles.