The Single Supplier

One Dutch company holds one hundred per cent of the market for extreme ultraviolet lithography and around eighty-three per cent of all lithography sales worldwide. There is no second supplier. There has never been a second supplier.

Every statement this newspaper has published about the construction of the artificial-intelligence economy terminates here. The accelerators, the memory stacked beside them, the processors in every telephone and the custom silicon designed by the companies trying to escape their suppliers, are all printed by machines that one firm in one town makes. In 2025 it recognised revenue on forty-eight of them.

What follows applies the same method used in these pages for the companies downstream: what its own accounts disclose, who owns it, what it owns, what each link around it is trying to do, how they cohabit, what it would mean if it stopped, and what it would take to build another one.

The internal skeleton

In its 2025 financial year the company reported net sales of €32.7 billion and net income of €9.6 billion. It closed the year with an order book of €38.8 billion, which is more than a year of revenue sold before it is built, and in the final quarter alone it took €13.2 billion of new orders, of which €7.4 billion were for extreme ultraviolet systems. Net bookings for the year were €28.0 billion.

Those sales were delivered on three hundred and twenty-seven lithography systems, of which forty-eight were extreme ultraviolet: four of the newest high-numerical-aperture type and forty-four of the preceding generation. The comparable figures for 2024 were four hundred and eighteen systems and forty-four extreme ultraviolet units. Fewer machines and more money is the arithmetic of this business, and the forty-eight are the ones the rest of this article is about.

For 2026 it guided, with its first-quarter results in April, net sales of between €36 billion and €40 billion at a gross margin between 51 and 53 per cent, raised from the €34 billion to €39 billion it had given in January.

Three things in those numbers are worth separating from the rest.

The first is the gross margin, which at roughly half is remarkable in the opposite direction from what a monopoly implies. The company that has no competitor earns a lower gross margin than its own customer, the foundry described in these pages at 67.7 per cent, and far less than the designers at the top of the chain. A monopoly on the hardest machine in the world does not produce the highest margin in the chain, because the machine itself is enormously expensive to build.

The second is the concentration. The two largest customers account for 38 per cent of revenue. This is a company with one hundred per cent of a market that is nonetheless dependent on a handful of buyers, which is the structural condition of every link in this chain and the reason none of them can press its advantage.

And then there is the political line in the accounts. Sales to China were 36 per cent of net system sales in the fourth quarter of 2025 and fell to 19 per cent in the first quarter of 2026, with the company guiding China from 33 per cent of sales towards 20 across the year. That decline is not commercial. It is the effect of export controls administered by governments, and it is the clearest illustration available that this company's revenue map is drawn in capitals rather than in sales offices.

In which direction the money runs

Upstream, towards its own suppliers, the company does not principally pay with purchase orders. It pays with ownership and with development funding, because the components it needs do not exist as products and have to be created. Its five most important acquisitions, set out below, cost it around €6 billion of disclosed consideration, and every one of them closed a gap in its own machine.

Downstream, towards its customers, the money arrives before the product leaves. An order book of €38.8 billion against annual sales of €32.7 billion means customers are committing more than a year of production in advance, for machines that will be delivered into buildings that do not yet exist. In this industry the customer finances the manufacturer, and then the manufacturer finances its suppliers, and the chip arrives at the end of a chain of promises made years earlier.

Who invests in it, and what it owns

The three enemies who paid for the machine

In 2012 this company did something that has no obvious parallel in modern industry, and it is the reason the machine exists at all.

Developing extreme ultraviolet lithography had already consumed decades and was consuming more money than the company could justify from its own balance sheet. Its answer was to ask its customers to pay for it. Three of them agreed: an American processor manufacturer, a Taiwanese foundry and a Korean conglomerate, which is to say three companies that compete directly and bitterly with one another in several markets.

Together they committed €1.38 billion to research and development over five years, split between the new lithography and a larger wafer format, and they bought an aggregate 23 per cent of the company for €3.85 billion in cash. The American firm took up to 15 per cent and funded €829 million of research; the foundry took 5 per cent and funded €276 million; the Korean group took 3 per cent.

Two conditions made it work, and both are instructive.

The shares issued were non-voting except in exceptional circumstances. None of the three could direct the company, influence which customer received machines first, or learn what the others were ordering. The structure was designed so that participating did not confer control, because without that nobody would have participated.

And the entire cash proceeds were returned to the existing shareholders through a buy-back. The customers' money went into the technology and the shareholders were not diluted in value, only in proportion.

That arrangement is the clearest documented case this newspaper has found of the proposition examined elsewhere in these pages: that companies with no affection for each other will fund a common objective when each needs the result and the structure prevents any of them from capturing it. Nobody had to trust anybody. The instrument did the work.

Who owns it now, and the end of the alignment

The sequel to that arrangement is the part nobody tells, and it is more instructive than the arrangement itself.

All three customers sold. The foundry disposed of its entire holding in 2015. The processor manufacturer began selling in 2016 and was below three per cent by October 2018. The Korean group reduced its stake progressively and disposed of the last 0.4 per cent in December 2023, having made a reported seventeenfold return on what it put in.

Nothing went wrong. The machine worked. And once it worked, holding the shares served no purpose that could not be served better by holding the money, because the thing the equity had been bought to achieve had been achieved.

That is the cleanest available demonstration of what this kind of alignment is and is not. It is not a partnership and it is not loyalty. It is a joint purchase of an outcome, and it dissolves when the outcome is delivered. The ambition that produces cooperation is the same ambition that ends it.

Today the register is ordinary. In the table of major shareholders in its own annual filing, the holdings large enough to require disclosure belong to institutional asset managers: the largest a passive index house with 26,325,103 shares, or 6.83 per cent, and the second an actively managed fund group with 19,612,223 shares, or 5.09 per cent, both as reported on 18 February 2026. There is no state holding, no controlling shareholder and no customer on the share register. The most strategically consequential manufacturer in Europe is owned, in effect, by everybody's pension and everybody's mutual fund.

What it owns, one by one

Every significant acquisition this company has made closed a gap in its own machine: a part, a measurement or a calculation that it could not afford to have anyone else control. Not all of them were suppliers. Every one of them was a dependency the company decided not to have.

The computational lithography business, acquired in March 2007 for $270 million, which the company recorded as about €203 million. It makes no hardware at all. Its software predicts what a given mask will actually print and then deliberately distorts the mask so that the distortions of the optics cancel out, which is the seventh module described below. It is the cheapest of the five purchases and the one that turned a machine into a system, and it is the reason the manufacturer sells computational products alongside the hardware today.

The light source, acquired in 2013 for €1.95 billion. This is the American company that built the system of lasers and tin droplets described below. Before the acquisition it was an independent supplier to the whole industry; afterwards the light source and the machine were developed by the same organisation, which matters because the two cannot be optimised separately.

The electron-beam metrology business, acquired in 2016 for 99.7 billion Taiwan dollars, a sum its own annual report records as €2.9 billion. Its product does not print anything; it inspects, finding defects far smaller than optical inspection can resolve. What it buys the parent is information: a printing machine that can measure the consequences of its own printing closes a loop that was previously open.

A quarter of the optics company, in 2016, for €1.0 billion plus committed development funding. This is the single most important relationship in the business, because the mirrors are the component that cannot be sourced anywhere else at any price, and the stake was taken precisely so that the relationship would not depend on a contract.

An optical module manufacturer, acquired in 2020, which makes the precision glass assemblies that surround the optics.

Add them together and the pattern is explicit: roughly €6 billion of disclosed consideration, with one of the five prices, that of the optical module manufacturer, never published. Not spent on growth, not on competitors and not on adjacent markets, but on removing the possibility that someone else controls a part of its own machine.

What each of them is trying to do

The light source business is trying to raise the power of the plasma, because throughput is directly limited by how many photons per second it can produce, and every additional watt is worth hundreds of millions in customer capacity.

The optics company is trying to hold a figure of around a tenth of a nanometre on larger mirrors as the numerical aperture increases, which is the central physical problem of the next generation.

The metrology business is trying to find smaller defects faster, because inspection time is dead time on a production line.

The computational business is trying to keep its predictions accurate as features shrink below the dimensions its models were built for, because every per cent of modelling error arrives at the customer as a per cent of lost yield.

And their own suppliers are pursuing the same thing one level down. The glass-ceramic blanks for the mirrors come from a small number of specialist glassmakers who have spent decades on materials with almost no thermal expansion. The drive laser comes from a machine-tool company that had to invent a carbon-dioxide laser of unprecedented stability. Neither of those businesses would exist in that form without this one customer, and neither has an alternative buyer.

Does it have competitors

Two companies made lithography systems for the industry's previous generations and still make them for older processes. Neither offers an extreme ultraviolet system. One of them is pursuing a different approach entirely, imprinting the pattern mechanically rather than projecting it with light, which avoids the whole problem of making optics for a wavelength that nothing transmits, and which has not yet been adopted for leading-edge logic.

The serious long-term effort is national rather than corporate. China has begun mass-producing immersion deep-ultraviolet systems domestically, with reported plans for five machines in 2026 and twenty in 2027, which addresses the generation before the one in question. Its extreme ultraviolet programme has produced a reported prototype, and people involved describe 2030 as a realistic target for making working chips with it.

So the competitive answer is: nobody sells an alternative today, one alternative technology exists and is not yet used at the leading edge, and one state is building the previous generation at scale while working on this one. The monopoly is real and it is not permanent, and the timescale on which it is challenged is measured in years rather than quarters.

The architecture of the machine, completely

The product is a single machine weighing on the order of 180 tonnes, containing something like one hundred thousand components, delivered in dozens of freight containers and assembled on site over months. A standard unit costs about €180 million; the newest generation between €380 and €400 million. What follows is what is inside it, in the order the light travels.

Module one: the source

The source is a separate vessel, usually beneath the floor of the cleanroom, and it is a plant rather than a lamp.

The droplet generator ejects molten tin through a heated nozzle as a stream that breaks into droplets of about thirty micrometres at a rate near fifty thousand per second. The stream must be stable in rate, size and trajectory for weeks of continuous running, because every droplet is a target that has to be hit twice.

The drive laser is a carbon-dioxide system of tens of kilowatts: a seed oscillator feeding a chain of amplifiers, with the beam transported tens of metres through a fixed path and steered onto the falling droplet with microradian accuracy. It fires a pre-pulse that flattens the sphere into a disc, which presents a larger and flatter target, and then a main pulse that vaporises the disc into plasma. Timing between the two is measured in nanoseconds and is adjusted continuously from camera feedback on the previous droplets.

The plasma reaches roughly 220,000 degrees and radiates across a broad spectrum, of which the usable fraction is the narrow band at 13.5 nanometres that the mirrors are tuned to reflect. Conversion efficiency is low, which is why a source producing a few hundred watts of in-band light consumes on the order of a megawatt.

The collector is an ellipsoidal mirror roughly the width of a dustbin lid, surrounding the plasma, coated with the same multilayer as the rest of the optics, which gathers light emitted in all directions and focuses it to a single point. That point is called the intermediate focus, and it is the boundary between the source module and the scanner: everything upstream is the business of producing photons, everything downstream is the business of using them.

Debris mitigation runs continuously, because the tin that is not converted sprays outward and lands on the collector. Hydrogen is flowed through the vessel so that tin is carried away as a volatile compound, and the collector is still a consumable item.

A spectral purity filter removes the residual infrared from the drive laser, which would otherwise reach the wafer and expose the resist where it should not.

Module two: the illuminator

Between the intermediate focus and the mask sits the illuminator, whose job is not to carry the image but to shape the light before it arrives.

It consists of arrays of small mirrors, in two sets. The field facets divide the beam and reshape it into the narrow arc-shaped slit that will scan across the mask. The pupil facets determine the angular distribution of the light, which is to say the directions from which the mask is illuminated.

That second control is more important than it sounds. The resolution of the final image depends not only on the wavelength and the optics but on the angles at which light strikes the pattern, and different pattern types print best under different illumination shapes. Choosing that shape, together with the pattern on the mask, is a computational optimisation performed before any wafer is exposed.

Module three: the mask stage

The mask is reflective, because nothing transmits at this wavelength, and it sits face down above the optics.

It is held electrostatically. Vacuum chucks, which hold a wafer by suction everywhere else in the industry, do not work in a vacuum, so the mask is clamped by electric field. It is scanned across the illumination slit at four times the speed of the wafer below, in the opposite direction, with the two stages synchronised to a fraction of a nanometre.

Above it sits the pellicle, a free-standing membrane on the order of fifty nanometres thick, whose purpose is to keep particles away from the mask surface. It must transmit most of the light in both directions, survive the thermal load of doing so, and not sag.

Module four: the projection optics

The heart of the machine, and the part that cannot be bought from anybody else.

Six mirrors in the current generation carry the image from mask to wafer, demagnifying it four times. Each is aspheric, each is figured to a tenth of a nanometre and coated with around fifty molybdenum-silicon bilayers, and each is mounted so that its position can be adjusted in several axes by actuators during operation.

That last point is what separates this from a camera. The optics are corrected in real time. The machine measures its own aberrations and moves the mirrors to compensate, continuously, because the optics absorb roughly a third of the light at every reflection and therefore deform as they work. The entire assembly is held on a frame of near-zero-expansion material, thermally conditioned, and isolated from everything that vibrates.

In the newest generation the numerical aperture rises from 0.33 to 0.55. To achieve that without the mirrors shadowing each other, the optics become anamorphic: four times demagnification in one axis and eight in the other. The consequence is that the exposure field is halved, so a large chip is printed as two fields that must be stitched together with an overlay error small enough that the join is invisible to the circuit.

Module five: the wafer stage

The wafer moves beneath the optics on a stage that is one of the most extreme pieces of motion engineering in existence.

It is magnetically levitated and driven by linear motors, because air bearings cannot be used in a vacuum. It accelerates at several times gravity, reverses, and settles to a position accurate to a fraction of a nanometre within milliseconds, then holds that accuracy while scanning. Position is measured continuously against the metrology frame by encoders and interferometers, and the stage is actively cooled, because the power dissipated by moving that much mass that fast would otherwise expand the metal it is made of.

Before each exposure the wafer is aligned by measuring marks printed in previous layers, and its height is mapped across the whole surface, because the depth of focus is a few tens of nanometres and the surface is never perfectly flat. Corrections are applied field by field.

Module six: everything that keeps it running

Vacuum. The entire optical path is pumped down, because air absorbs the light. Wafers and masks therefore enter through load locks, and every wafer must be pumped and vented on the way in and out, which is part of why throughput is what it is.

Hydrogen. Flowed through the system both to carry away tin and to prevent carbon deposits building up on the mirrors from residual hydrocarbons.

Thermal control. Every element that matters is temperature-conditioned to millikelvin stability, because the materials expand.

Vibration isolation. The machine sits on an isolated foundation, and within it the metrology frame floats independently of the structure that moves.

Module seven: the software that is half the product

The pattern on the mask is not the pattern that is wanted on the wafer.

At these dimensions, light diffracts enough that a drawn rectangle prints as a rounded blob and adjacent features interfere with each other. The correction is computational: the mask pattern is deliberately distorted, with extra shapes added that do not print themselves but alter how the neighbouring features do, so that what emerges on the silicon is the intended geometry. That process, together with the choice of illumination shape described above, is a large optimisation run on substantial computing resources for every layer of every chip.

This is why the manufacturer sells computational products alongside the machines, and why its acquisition of an electron-beam metrology business matters: the machine prints, the metrology measures what was actually printed, and the software closes the loop between them.

What the whole sequence achieves

Each wafer is exposed field by field, a few hundred fields per wafer, at a rate of roughly one hundred and fifty to two hundred wafers an hour. Every field requires the stages to accelerate, synchronise, scan and settle. The machine repeats this continuously, for years, with the optics correcting themselves as they heat, in a vacuum, while a laser vaporises fifty thousand droplets of tin per second in a vessel below the floor.

And the output of all of it is one layer. A finished chip requires dozens of lithography steps, most of them on simpler machines, interleaved with deposition, etching, implantation and polishing, across roughly a thousand process steps and three to four months.

How they cohabit

The relationships around this company have the same shape as those documented elsewhere in this chain, and the same stability.

It cannot exploit its customers, because there are only a handful of them and the failure of any one would remove a large fraction of its market. Its customers cannot exploit it, because there is nowhere else to buy. Its suppliers cannot exploit it, because it has bought equity in the two most critical of them. And its government cannot ignore it, because the machine has become an instrument of foreign policy, which is why its sales map is redrawn by export control decisions taken in capitals.

Each party holds something the others cannot replace, and no party can use that position without destroying the thing it depends on. The 2012 funding arrangement was that structure made explicit and written down: three competitors paying jointly for a capability none of them could develop alone, in a form that prevented any of them from controlling it.

What happens if it stops

What follows is a constructed scenario rather than a forecast, and no cause is specified, because the mechanics are identical whether the interruption is physical, political or financial.

Immediately: nothing. Every machine already installed keeps running. There are several hundred extreme ultraviolet systems in the world and they do not stop because their maker does.

Within months: the fabs under construction stop. Every leading-edge plant being built anywhere, including those being built expressly to reduce the world's dependence on one island, is being built around machines that have been ordered and not yet delivered. An order book of €38.8 billion is, from the customer's side, a construction schedule. Without the deliveries, those buildings become very expensive empty rooms.

Within a year: capacity stops growing. Not falls: stops. The installed base continues to produce, and the industry's ability to produce more of anything at the leading edge is frozen at whatever it was on the day of the interruption.

Within two to three years: it begins to fall. These machines require continuous servicing, spare parts, consumable optics and software from the manufacturer. A fleet without its maker degrades, slowly at first and then not slowly.

And there is no substitute at any price. No other company sells the machine. The alternative national programme describes 2030 as a realistic target for working chips from a prototype. A second supplier, starting today with unlimited funding, would be reproducing not a product but the supplier network and the accumulated knowledge of several decades.

What it would do to artificial intelligence

The effect is the same in kind as the one described in these pages for the foundry, and worse in degree, because this is one layer further up.

The supply of new accelerators would be fixed at the level the existing installed base can produce, and then would decline as machines age out. No new leading-edge capacity would be added anywhere in the world, by anybody, including the companies designing their own silicon to escape their suppliers.

Every projection in the field assumes computation available for training grows by a large multiple each year. That assumption rests, at the bottom, on the delivery of a few dozen machines a year from one factory in the Netherlands. Remove it and the industry is left to improve by efficiency rather than by scale, which it is capable of doing, and which produces a different trajectory from the one currently being financed in hundreds of billions of dollars.

How you would build another one

The sequence, with what is known about each step.

Decades, not years. The technology took roughly two decades of development after the physics was understood, with substantial public funding in several countries and a research consortium that included the companies that would later buy the machines.

A network before a product. Thousands of suppliers had to develop capabilities that had no other customer: mirrors of that figure, lasers of that stability, vacuum systems of that cleanliness, resists that react to those photons. None of those businesses existed independently and none would have been built speculatively.

Customers willing to fund their own supplier. The decisive step was the one described above. The development was completed because three customers paid for it, and they paid because each had concluded that no alternative existed and that the structure protected them from one another.

A state that permits it. The machine is export-controlled. Whether a new entrant may buy components, hire people or sell the result is decided by governments.

And an organisation willing to lose money for twenty years. The company shipped its first production extreme ultraviolet systems long after the programme began, and the technology was repeatedly described as impossible during that period by people with good reasons.

The honest conclusion is that this is not a business that can be entered by a company. It has been attempted, successfully, exactly once, by a consortium of companies and states over twenty years, and the only current attempt at a second one is also a consortium of a company and a state, working on a timescale measured in decades rather than product cycles. Which is the answer to why nobody else has done it. Not because the physics is secret, but because the thing to be assembled is less a product than a treaty.

The parts of the parts, and how any of it is obtained

Following the machine downwards ends somewhere unexpected, and the descent is worth making because it shows where the real limits are.

The mirror is a stack of fifty molybdenum and silicon layers on a substrate. The substrate is a glass-ceramic with almost no thermal expansion, which works by balancing a crystalline phase that contracts against a glass phase that expands; it is cast and then annealed for months, because cooling it quickly introduces stresses that no later polishing can remove. The polishing is done by scanning an argon ion beam across the surface to remove material atom by atom, measuring interferometrically between passes, over months per mirror. The coating is sputtered with thickness control in picometres, because the reflection works by interference and a few per cent of error destroys it.

The laser is a chain of amplifiers whose beam is transported tens of metres and pointed with microradian accuracy at a falling droplet thirty micrometres across, twice, fifty thousand times a second. Its optics come from the same handful of specialist glassmakers.

The tin must be of a purity that leaves no residue that cannot be flushed, and the hydrogen used to flush it must be of a purity that introduces nothing itself.

The vacuum is maintained by turbomolecular pumps against outgassing from every surface inside, which is why the materials used in the chamber are selected as much for what they do not emit as for what they do.

And the mask carrying the pattern needs a defect-free multilayer blank, supplied by two Japanese companies, written by an electron-beam system supplied by essentially one Austrian company, and protected by a membrane fifty nanometres thick that must transmit the light and survive the power.

How you obtain each of these

Almost none of it is bought in the ordinary sense, and the mechanisms are the same at every level.

Equity, where the supply cannot be duplicated. The optics are secured by owning a quarter of the company that makes them. The light source is secured by owning all of it. This is the first instrument reached for.

Development funding, where the component does not yet exist. A supplier will not build a capability speculatively for one customer, so the customer pays for the development and then buys the output.

Prepayment and multi-year commitment, where capacity is the constraint. Customers commit years ahead and that commitment is what justifies the expansion.

Qualification, where quality cannot be inspected after the fact. A material or component enters the machine only after a process lasting months or years, which means switching supplier is not a procurement decision but an engineering programme, and therefore rarely happens.

And licence, where a government has an interest. The finished machine requires an export licence. So, increasingly, do several of its components. This is the only link in the chain where the answer to how you obtain it is that you may not be permitted to.

Why this machine holds up the world economy

The negative case has been made above: what fails if it stops. The positive one is more important and is rarely stated, because it requires connecting four things that are normally discussed separately.

One: everything modern is built on a falling price, and this machine is why the price falls

The defining economic fact of the last half century is not that computers exist. It is that computation has become continuously cheaper, year after year, for sixty years, which is a behaviour no other industrial product has ever displayed. Steel did not do this. Energy did not do this. Food did not do this.

That decline has a physical cause. The cost of a chip is set largely by the area of silicon it occupies, and the number of functions you can place in a given area is set by how small a feature you can print. Printing smaller features is lithography, and at the leading edge there is one machine that does it.

So the sentence that is usually said loosely, that the world economy runs on cheap computing, resolves at the bottom to a specific statement: the price of computation falls because the resolution of a few hundred machines improves, and when it stops improving the price stops falling.

Every business model built on the assumption of cheap computing, which is now most of them, is downstream of that.

Two: the measurable share, and the unmeasurable one

The measurable part is already large. Modelling by Bloomberg Economics puts the sectors that use chips as a direct input at about 5.6 per cent of total global value added, which is close to six trillion dollars.

The unmeasurable part is everything else, because there is no longer a sector that does not use computation somewhere in its production. A farm uses it in the tractor and in the futures contract that prices the crop. A hospital uses it in the scanner and in the scheduling. A bank is computation with a licence. The direct share understates the dependency in the same way that measuring the economic contribution of electricity by the revenue of power companies would understate electricity.

Three: artificial intelligence is the first industry that is explicitly limited by this one

Most industries consume computation. This one is constrained by it, and says so in public.

The capability of a model is, to a first approximation, a function of how much computation was spent training it, and the cost of using it is a function of how much computation each answer requires. Both improve when accelerators improve, and accelerators improve because they contain more transistors, switching faster, at lower cost per transistor. That improvement comes from the manufacturing node, and the node comes from this machine.

The thousandfold collapse in the price of machine reasoning over three years, documented elsewhere in this newspaper, is partly algorithmic and partly the compounding of that hardware improvement. Remove the hardware half and the curve flattens, as described above.

Which is why every plan in this industry, whether it is measured in gigawatts, in billions of dollars of committed capital or in orbital satellites, is in the end a bet on the output of one factory in the south of the Netherlands continuing to rise.

Four: the frontier of knowledge is now compute-bound

This is the part that goes beyond economics.

A growing number of scientific problems are no longer limited by instruments or by ideas, but by how much computation can be applied to them. Protein structure and drug design. Climate and weather models whose resolution is set by available cycles. Fusion and plasma simulation. Materials discovery, where the search space is too large to explore physically. Genomics at population scale.

In each of those fields the rate of progress has become a function of the rate at which computation gets cheaper. That is a new condition in the history of science, and it means that the improvement curve of a lithography machine in the Netherlands sets, indirectly, the pace of discovery in laboratories that have never heard of it.

And the honest limit of the claim

Computation is not the only input, and the machine is not the only cause. Algorithms have improved enormously and account for a large share of the gains. Energy is now a binding constraint in its own right. Data, capital and people matter.

But of all those inputs, this is the only one with a single supplier, a single factory, an annual output counted in tens of units and an order book that is reviewed by foreign ministries. The others can be obtained in many places. This one cannot be obtained anywhere else at all.

That is the whole of the argument. Not that the machine is impressive, though it is. That the rate at which the world gets richer, and the rate at which several sciences advance, are both functions of a curve that one company in one town is responsible for continuing.

How this leaves the Earth

The consequence of concentrating a capability this tightly is not only economic, and the shape of it is visible already.

It makes one small country strategically significant out of all proportion to its size. A nation of eighteen million people holds a veto on the industrial future of every other nation, exercised through an export licensing office. Its foreign policy is now consulted by governments that had no previous interest in it, and its ministers are lobbied on decisions that would ordinarily be technical.

It converts a supply chain into an instrument of policy. The company's China revenue fell from 36 per cent of system sales in one quarter to 19 per cent in the following quarter, not because demand changed but because rules changed. Every customer of this machine now has to model political risk the way it models yield.

It concentrates an enormous amount of human capability in a small number of towns. The knowledge that makes this possible lives in perhaps a few tens of thousands of people in a handful of places in Europe, the United States and Japan. It is not written down in a form that would survive their dispersal, which is why the industry's expansion plans everywhere begin with moving experienced people rather than with buying equipment.

And it makes the physical infrastructure of computation dependent on a rate rather than a stock. The world does not have a supply of lithography machines. It has a flow, forty-eight extreme ultraviolet systems in 2025, and every projection of future computing capacity, every data centre plan and every national strategy is, underneath, an assumption about that flow continuing.

The future this skeleton builds

If the structure described here succeeds, and on present evidence it is succeeding, the shape of what follows is legible.

Computation becomes a scheduled resource rather than a purchasable one. It already is, upstream: machines are allocated, capacity is reserved years ahead, and the queue is the market. As demand grows faster than the flow of machines, that logic propagates downstream until it reaches the end customer, and renting computation starts to look less like buying electricity and more like booking a slot.

Industrial policy becomes the principal form of economic policy. Every government that understands this chain is now subsidising a piece of it. The sums already committed in the United States, Japan, Europe, Korea and China exceed, in aggregate, what any of them spends on most other industrial objectives, and the spending is justified by security rather than by return.

Alignment replaces competition at the top of the chain and intensifies it below. The companies that cannot replace each other cooperate, fund each other and take equity in each other, exactly as documented here and elsewhere in these pages. The companies that can be replaced compete on price as fiercely as ever. The line between those two groups is the most valuable thing a business can know about itself.

And the physical world reorganises around a small number of buildings. Everything in the artificial-intelligence economy, the gigawatt campuses, the orbital proposals, the valuations, the national programmes, resolves at the bottom to a flow of machines from one town, built from mirrors polished for months and light made by vaporising tin fifty thousand times a second.

That is the whole of it. The most abstract industry humanity has built rests on the most physical object it has ever manufactured, and there is exactly one place that makes it.

What the filings establish

Three things, all disclosed, and all of them numbers rather than judgements.

That three hundred and twenty-seven machines, forty-eight of them at the leading edge, produced €32.7 billion of sales at a gross margin of 52.8 per cent, below the 67.7 per cent earned by a customer that cannot buy those machines anywhere else. That is the measurable price of building the hardest object in the chain instead of using it.

That €1.38 billion of its research funding and €3.85 billion of its equity came from three of its own customers, in non-voting form, with the cash handed straight back to existing shareholders, and that the last of the three was out by December 2023. The whole life of an alignment, from instrument to exit, took eleven years and is on the public record.

And that 36 per cent of its system sales in one quarter became 19 per cent in the next with no change in demand, which makes it the first component supplier in history whose order book is a foreign policy document.

Everything the world currently expects from artificial intelligence, measured in gigawatts and in orders of magnitude, is downstream of a few dozen machines a year leaving one town in the south of the Netherlands.

Sources

Filed and published figures. Net sales, net income, order book, bookings and guidance are from the company's own results releases filed with the United States Securities and Exchange Commission: the fourth-quarter and full-year 2025 release of 28 January 2026, which gives €32.7 billion of net sales, €9.6 billion of net income, a 52.8 per cent gross margin, the €38.8 billion backlog, the €13.2 billion of fourth-quarter bookings with €7.4 billion of extreme ultraviolet among them, and guidance of €34 billion to €39 billion for 2026; and the first-quarter 2026 release of 15 April 2026, which raised that guidance to €36 billion to €40 billion and reports the China share of sales at 19 per cent against 36 per cent in the preceding quarter. The unit counts, three hundred and twenty-seven lithography systems sold and forty-eight extreme ultraviolet systems recognised in 2025 against four hundred and eighteen and forty-four in 2024, are from the 2025 annual report, which also states the split of the forty-eight into four of the newest type and forty-four of the preceding one. Customer concentration is from the same disclosures.

The 2012 arrangement. The programme as proposed, including the €1.38 billion of research funding, the non-voting character of the shares and the return of the proceeds to existing shareholders through a synthetic buy-back, is set out in the shareholders' circular of 24 July 2012 published for the extraordinary general meeting of 7 September 2012, which describes a maximum participation of 25 per cent. The amount actually subscribed, €3,853.9 million, of which €3,016.1 million by the American and Korean participants to 30 September 2012 and €837.8 million by the foundry on 31 October 2012, is in the filing of 21 November 2012. The aggregate 23 per cent and the individual stakes are from the company's announcements as each participant joined.

Machine specifications. Unit prices, numerical aperture, source operating principle and optical configuration are as published by the manufacturer and in the optical engineering literature. Mirror substrate specifications on the order of 0.1 nanometres root mean square are as published in that literature.

Competitors. The status of alternative lithography approaches, domestic Chinese deep-ultraviolet production plans of five systems in 2026 and twenty in 2027, and the description of 2030 as a realistic target for the Chinese extreme ultraviolet programme, are from trade and press reporting and are identified as such in the text.

Ownership and acquisitions. The institutional holdings quoted, 26,325,103 shares or 6.83 per cent and 19,612,223 shares or 5.09 per cent as reported on 18 February 2026, are from the table of major shareholders in the annual report on Form 20-F for 2025, which lists holders disclosing at least 3.0 per cent to the Dutch market authority or 5.0 per cent to the Securities and Exchange Commission. The consideration for the electron-beam metrology business, a share swap agreement of 16 June 2016 for TWD 99.7 billion, recorded as €2.9 billion, completed on 22 November 2016, is from the Form 20-F for 2016; figures of around €2.75 billion that appeared in the press on the day of the announcement are the same price converted at that day's rate. The $270 million, about €203 million, paid for the computational lithography business is from the company's completion announcement of 8 March 2007. The consideration for the light source business in 2013 and for the 24.9 per cent interest in the optics partnership in 2016, and the disposal of their stakes by the three customer investors in 2015, 2016-2018 and 2023, are from company announcements, filings and shareholder disclosures.

What is not reported here. No figure is given for the yield, reliability or uptime of any machine at any customer, because those figures are not published by the manufacturer or by its customers. No price is given for the optical module manufacturer acquired in 2020, because none was disclosed; the figure of roughly €6 billion for the five acquisitions is therefore the sum of the four that were. And no revenue figure is attributed to any individual customer, because the company discloses concentration without naming the buyers.