The House That Makes Them All

There is a company that appears in every story this newspaper has published about the construction of the artificial-intelligence economy, always in the same position and never as the subject. It is the place the chips are made.

Not designed: made. The distinction is the whole of its business and most of its power. The accelerator that trains frontier models, the processors its competitors sell against it, the custom silicon that the largest buyers of those accelerators are designing specifically in order to stop buying them, and the telephone in the reader's pocket, are to an overwhelming degree manufactured in the same buildings, by the same company, under contracts negotiated years in advance.

That company does not have to prefer an outcome. Its filings explain why with unusual clarity.

The internal skeleton

In the quarter to June 2026 it reported revenue of $40.2 billion, up twelve per cent on the previous quarter and 33.7 per cent on the year, with a gross margin of 67.7 per cent and an operating margin of 60.3 per cent. It guided the following quarter to between $44.6 billion and $45.8 billion. The quarter before that had produced about $35.9 billion.

Those are the margins of a company with no substitute, and the interesting part is what it is deliberately doing to them.

Capital spending for 2026 is set at between $60 billion and $64 billion, allocated roughly seventy to eighty per cent to advanced process technology, about ten per cent to specialty processes and between ten and twenty per cent to advanced packaging, which is the step that has become the binding constraint on the entire industry. That is an enormous number in absolute terms and it is not the striking one.

The striking one is that the company has told its owners, in advance, that two of its strategic decisions will reduce its margin. Ramping its newest two-nanometre process is expected to dilute gross margin by three to four percentage points. Starting up its plants outside Taiwan, in Arizona, in Kumamoto and in Dresden, is expected to dilute it by a further two to three. Between five and seven points of margin, announced ahead of time, as the price of two things: being first on the next process, and not being in only one place.

Very few companies can afford to publish that sentence. It is the clearest statement available of what a dominant position is actually worth: the ability to spend your own margin on removing your own risks, and to tell the market you are doing it.

The external skeleton: customers who are enemies

Nvidia's external structure, as set out in these pages, is a portfolio of shareholdings. This company's is something stranger and more powerful: a customer list.

Its annual filing discloses that the ten largest customers accounted for 78 per cent of net revenue, that the largest single customer was 19 per cent, and that the second largest was 17 per cent. A year earlier, that second customer had been 12 per cent.

Read those three numbers together and a transition appears in the accounts of a manufacturer before it appears anywhere else. The second-largest buyer of the world's most advanced manufacturing capacity grew from roughly an eighth to roughly a sixth of a company that itself grew by a third. That is what the artificial-intelligence build-out looks like from underneath, expressed not as a narrative but as a line in a concentration disclosure.

The composition of the revenue says the same thing from another angle. High-performance computing, the category that contains data centre silicon, was 43 per cent of revenue in 2023, 51 per cent in 2024 and 58 per cent in 2025, at NT$2,192,931 million. Smartphones, which built this company and dominated it for a decade, are now 29 per cent. In two years the centre of gravity of the most important manufacturer on Earth moved from the pocket to the data centre, and it did so without the company changing anything about what it sells.

Geographically, 75 per cent of net revenue comes from North America.

Who owns it, and what it owns

Two questions that are usually asked of a company and rarely of this one, and the answers are stranger than the business itself.

Who invests in it

Its largest single shareholder is a state fund. The National Development Fund of Taiwan held 1,653,709,980 shares as of 31 July 2026, which is 6.38 per cent of the company. That is the largest holding anybody has, and it is not a controlling one.

The historical figure next to it is the one worth pausing on. At the founding in 1987, that same fund provided 48.3 per cent of the start-up capital. The state put up almost half the money to create a company that did not exist in an industry that did not exist in that country, and has since been diluted to roughly a sixteenth of it.

Everything else is dispersed. American depositary receipts held through a single depositary bank account for 20.49 per cent of the share count. Among the institutional holders are the sovereign fund of another country, a European central bank's investment arm, and the two largest index managers in the world. There is no controlling shareholder of the company that manufactures the world's most advanced semiconductors, and the single largest voice in its ownership holds six per cent.

Compare that with the structure described elsewhere in this newspaper, where two individuals hold majority voting control of two of the largest companies on Earth through shares carrying ten votes each. The most strategically consequential manufacturer in existence is governed by nobody in particular.

What it invests in

Here the structure inverts in a way that is genuinely unusual, and it answers a question this newspaper examined elsewhere in these pages in the abstract.

Its plants outside Taiwan are not subsidiaries. They are joint ventures, and the minority partners are its own customers.

The Japanese plant, in Kumamoto, is owned 86.5 per cent by the company, with Sony Semiconductor Solutions at 6.0 per cent, Denso at 5.5 and Toyota at 2.0. The European plant, in Dresden, is owned 70 per cent by the company, with Bosch, Infineon and NXP each holding 10.

Those six minority partners are not financial investors seeking a return on a fabrication plant. They are buyers of chips. An image-sensor maker, two car-parts manufacturers, a carmaker and two European semiconductor firms have each put equity into the factory that will supply them, in the country where they operate.

What they are purchasing with that equity is not profit. It is certainty of supply, and a seat at the table when capacity is allocated. Having lived through a period in which assembly lines stopped for want of components that cost a few dollars, they have concluded that the cheapest insurance available is to own a piece of the plant.

This is the clearest documented example in modern industry of the alignment described in these pages: companies with no common product, no common market and no particular affection for one another, pointing at the same objective because each of them needs the same thing to exist. Nobody had to persuade anybody. The shortage did the persuading.

Where that leads

If this is the model, and the evidence suggests it is becoming one, the consequence is a change in how industrial capacity gets built.

For forty years, manufacturing concentrated wherever it was cheapest and most efficient, and that logic produced the geography described later in this article: a handful of addresses with no redundancy. The emerging model is the opposite. Capacity is built in several places at once, more expensively, financed jointly by the companies that will consume it and by the governments that want it inside their borders, and co-owned by both.

That arrangement is less efficient by design. It costs the operator between two and three points of gross margin, which it has disclosed in advance, and it costs the governments subsidies, and it costs the customers capital they would rather have spent on their own products. All three are paying for the same thing, which is the right not to depend on a single location.

What it produces, if it works, is a world with more fabrication plants than pure economics would ever have built, each one partly owned by its users. That is not how the semiconductor industry has functioned at any point in its history. It is, however, exactly how ports, pipelines, railways and electricity grids were financed in every previous industrial era, by the people who could not operate without them.

Why it does not have to care who wins

Here is the structural fact that makes this company different from every other participant in the story.

The firms on that customer list compete with each other ferociously. Some of them exist specifically to displace others on the same list. The largest buyers of accelerators are designing their own accelerators in order to stop buying them, and those designs are manufactured in the same fabs as the accelerators they are meant to replace. A victory by any of them is, from the manufacturer's position, a transfer between two of its own customers.

This is a different kind of power from the one described in these pages previously. A company with ninety per cent market share in a product can be displaced by a better product. A company that manufactures everyone's product is displaced only by the arrival of another manufacturer, and the arrival of another manufacturer takes a decade, costs hundreds of billions and requires a workforce that does not currently exist.

It also means the company allocates rather than sells. When capacity is scarce, which it has been continuously, the question is not what the price is. It is who gets the wafers and when, and that decision shapes which products reach the market and which do not. Allocation is the quiet instrument of this industry, and it sits here.

Does it have competitors, and why do they sell less

It has several. Measured by revenue, which is the only measure that comes from published accounts rather than from estimates, the gap is not a lead. It is a different category.

Of roughly $53.5 billion of global foundry revenue in the second quarter of 2026, this company took about 72.5 per cent. Second place was a Korean conglomerate at 5.9 per cent, down from 6.5 in the previous quarter. Third was a Chinese manufacturer at 5.4 per cent, rising. Then two older firms at 3.9 and 3.3 per cent. The American company currently attempting to re-enter the business does not appear in the top ten, and a Japanese state-backed start-up is still developing its first process.

Sixty-six percentage points separate first from second.

The obvious question is why, and it is worth being explicit about what can be established and what cannot, because this is an industry in which the decisive number is never published by anybody.

It is not distribution. There is no distribution in this business. Perhaps fifty organisations in the world commission leading-edge silicon, and every one of them knows precisely what every foundry can do. Nobody loses an order because a customer failed to hear about them.

And nobody outside the companies knows the yields. Yield, the proportion of chips on a wafer that work, is the number that decides the economics of a fabrication plant, and no foundry on Earth discloses it. Every figure in circulation, including the ones repeated confidently in the trade press, is an estimate produced by supply-chain research houses from fragments, and such estimates are frequently wrong and occasionally placed by someone with an interest in them. This newspaper does not report them as facts, and the reader should treat any article that states a yield percentage as a certainty with the same caution.

What can be established, from published accounts and corporate structure, is the following, and it is sufficient.

The economics of a foundry are decided by yield, and yield is decided by accumulated volume. A customer pays for wafers; what it receives is working chips. If one plant delivers more working chips from the same wafer, its customer's cost per chip is lower by that proportion, and no commercial argument survives that difference. Yield improves by running volume and correcting the defects that only appear at volume, which means the leader's advantage compounds with every wafer it processes and cannot be bought by a competitor at any price. The second plant to attempt a process is not behind by the time it takes to build a plant. It is behind by the number of wafers the first one has already ruined and learned from.

Switching has a cost measured in years. Designing a chip for a particular process takes one to three years and a great deal of money, and moving that design to another foundry is not a transfer but a redesign. Customers therefore commit generations in advance. Losing one generation means losing the customer for two.

And the structural reason, which is a matter of corporate fact rather than estimate. The company at the top of this list was founded on a promise not to compete with its own customers, and it does not design or sell chips of its own. Its principal competitor manufactures for customers while also selling memory, telephones and its own silicon. A fabless designer handing over its most advanced design is, in that case, handing it to a company that sells against it. The pure-play foundry model exists precisely because of that problem, and it is the reason the model won.

None of this means the competition is finished. The second-placed company signed a foundry contract reported at $16.5 billion with a single customer, the largest of its kind, and counts major designers among its clients. What the published numbers show is a gap of sixty-six points and a route back that runs entirely through a figure that nobody will ever publish.

What each link in the chain is trying to do

The foundry is the centre of the story and it is not the whole of it. Behind it stands a chain in which almost every link is a monopoly or close to one, and none of them can be replaced by spending money quickly. Each is pursuing something specific, and the pursuits only make sense together.

The machine that prints

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

What it sells costs about €180 million for a standard machine and between €380 and €400 million for the newest generation, of which it expects to ship between five and ten in 2026 within a total of more than sixty units. Its order book exceeds €38 billion, which is more than a year of revenue sold before it is built. In the first quarter of 2026 it reported €8.8 billion of sales at a 53 per cent gross margin.

What it is trying to do is stay the only company that can print a feature smaller than the one everybody else can print. That is the entire strategy, it has worked for two decades, and the reason it works is that the machine is not really a machine. It is an assembly of the best optics, the best vacuum systems and the best precision engineering that several countries can produce, delivered in crates and assembled on site over months.

The rest of the equipment

Printing the pattern is one step of several hundred. Depositing the layers, etching them away, polishing the surface flat, implanting the ions and inspecting the result are each done by equipment from a small number of American and Japanese firms, each dominant in its own step and dependent on the others for theirs.

What they are all pursuing is the same thing from different angles: being the indispensable step. A fabrication plant is a sequence, and a sequence has no weakest link that can be tolerated, because every wafer passes through all of them.

The surface it is printed on

Two Japanese companies supply roughly fifty-three per cent of the world's silicon wafers. In 2025 they committed about a billion dollars to add two hundred thousand wafers a month of the ultra-flat grade that the two and three nanometre processes require, because at those dimensions the flatness of the starting material is a limit on the yield of everything built on it.

What they are pursuing is a specification rather than a volume. Anyone can make a silicon disc. Making one flat enough for a pattern measured in the width of a few dozen atoms is a different industry.

The chemistry nobody discusses

This is the most exposed point in the entire chain and almost nothing is written about it.

Photoresist is the light-sensitive coating in which the pattern is formed. Without it there is no chip of any kind, at any node, anywhere. Four Japanese companies accounted for seventy-six per cent of the global market, and the raw material for those chemicals comes largely from petrochemical feedstock that arrives by sea.

In 2026, disruption to a maritime chokepoint curtailed supplies of that feedstock sharply enough that Japanese suppliers formally warned their customers, who are the memory and logic manufacturers of the world, about procurement difficulties. The most advanced manufacturing on Earth depends on a liquid made by four firms in one country from an input that passes through a strait.

What those companies are pursuing is reliability rather than growth, which is the usual posture of a supplier whose product is irreplaceable and whose failure is catastrophic for everyone downstream.

The memory

An accelerator is useless without memory attached to it at enormous bandwidth, and that memory is made by three companies. One holds roughly fifty-six per cent of the market for the high-bandwidth type, with the other two sharing most of the rest. All three have their capacity for 2026 sold out.

Note the direction of power in that sentence. Capacity sold out means the manufacturer is allocating, not selling, exactly as the foundry is. And a memory maker's product must pass qualification by the accelerator designer before it can be shipped at all, which means the three of them compete to satisfy a specification written by their customer.

What they are pursuing is a position in the next generation, because in memory the qualification decides the next two years and the decision is made once.

The design layer

A new manufacturing process does not arrive at a customer as a factory. It arrives as software: a set of design rules, verified cell libraries and simulation models, without which no engineer can lay out a circuit that the plant can actually build.

That layer is supplied by two American software companies and a small number of design service firms, and it is the reason a foundry's ecosystem matters as much as its equipment. A process that nobody can design for is a process nobody will use.

What that layer is pursuing is to be the environment in which the chip is conceived, because whoever supplies the environment is present in every project from the first day and is extremely difficult to remove from it.

The customers

And at the end, the buyers, who are the subject of the first half of this piece: ten of them accounting for 78 per cent of revenue, the largest at 19 per cent, the second at 17 per cent, and several of them engaged in trying to destroy one another.

What each of them is pursuing is the same scarce thing, which is capacity in the newest process at the earliest date. That is the only contest in this industry that cannot be won with money alone, because the capacity is allocated and the allocation is decided upstream.

How they cohabit

Set the chain out and a pattern appears that is unusual in modern industry.

Almost every layer is a monopoly or a near-monopoly, and not one of them can exercise it against the others. The lithography company could in principle raise prices without limit; it would gain nothing, because its only customers are three or four manufacturers and the collapse of any of them removes a third of its market. The foundry could in principle squeeze the wafer suppliers; it would be squeezing the flatness its own yield depends on. The memory makers compete viciously for a qualification written by a customer who needs all three of them to survive, because a single supplier of memory would be an unacceptable risk to the customer.

Each link holds a knife to the others' throat and none of them can use it. That is not cooperation and it is not friendship. It is the mutual hostage structure described elsewhere in these pages, assembled accidentally over forty years by companies that were simply specialising, and it is more stable than any contract could make it.

The second property is that they all move on the same calendar. A new process node requires new lithography, new chemistry, new wafer specifications, new design rules, new packaging and new memory, and all of them have to be ready in approximately the same quarter. Nobody can go faster than the slowest, and nobody can afford to be the slowest twice.

And the third is that the whole chain is funded in advance by the customers. Capacity is reserved years ahead, equipment is ordered against those reservations, materials capacity is expanded against those orders. The money travels backwards up the chain long before any chip travels forward down it.

The geography, and what is being done about it

Roughly nine in ten of the world's most advanced chips are manufactured on one island, and by the estimate of the research firm Counterpoint, close to ninety-nine per cent of the chips used to train frontier models originate there. That concentration is the single largest physical risk in the world economy, and it is being addressed, slowly and expensively, by the company itself.

The commitment in Arizona now stands at about $265 billion, with reports of as many as twelve fabrication plants and four packaging facilities planned. Plants are starting in Japan and in Germany. And the margin cost of all of it, two to three points, is being disclosed in advance rather than discovered later.

The timetable is the part that matters and it is measured in years. Advanced packaging capacity outside Taiwan arrives towards the end of the decade on announced schedules. Until then, every scenario in which the world's computing requirements keep growing runs through the same few square kilometres.

How that leaves the Earth

The consequence of all this is geographic, and when it is written down it is difficult to believe.

The most advanced logic is manufactured in essentially one place. The machine that makes that manufacturing possible is made in essentially one town, in one country, by one company. Roughly half of the world's silicon wafers come from two firms in one country. Three quarters of the photoresist comes from four firms in the same country, from a feedstock that arrives by sea through a small number of straits. The memory comes from three companies in two countries. The design environment comes from two companies in one country. And the advanced packaging that joins the logic to the memory, which is the step that currently limits how many accelerators can exist, is concentrated in the same island as the logic until new capacity opens elsewhere towards the end of the decade.

The computation of the entire planet has roughly six addresses.

This is not a conspiracy and nobody designed it. It is the result of forty years of specialisation, in which each company did the sensible thing of concentrating on the step it was best at, and the aggregate of all those sensible decisions is a supply chain with no redundancy at any level.

The duplication now under way, measured in hundreds of billions of dollars and in new plants in Arizona, Japan and Germany, is an attempt to convert six addresses into eight or ten. It is being paid for partly by governments and partly out of the margin of the companies themselves, which have told their shareholders in advance what it will cost them. On the published timetables, it arrives towards the end of this decade.

Until then, every forecast about artificial intelligence, every plan for a data centre, every projection of economic growth that depends on computation, and every device in every pocket, rests on a chain in which almost every link is a single point of failure, and in which the links cannot afford to harm each other because each one would be destroying its own only market.

That is the most important industrial fact in the world, it is disclosed in full in documents anyone can download, and it is almost never stated in one place.

What happens if it stops

What follows is a constructed scenario, not a forecast and not a prediction about any country. It is an attempt to describe, using published figures, what the world looks like if this company ceases to supply it. The cause is left unspecified on purpose, because the mechanics are the same whether the interruption comes from conflict, from a blockade, from an earthquake or from a loss of water or power to the plants.

The first week: nothing happens. The industry runs on inventory, and inventory is measured in weeks for finished components and a few months for some categories. Shops keep selling, factories keep assembling, data centres keep installing what is already on pallets. The first visible effect is financial rather than physical: the shares of every company on the customer list reprice immediately, because the market understands supply chains faster than consumers do.

The first quarter: the queue empties. Products at the front of the queue are the ones with the newest chips, which means the most advanced telephones, the newest accelerators and the newest computers stop first, not last. Allocation ceases to be commercial and becomes political, because when there is nothing to sell there is only the question of who gets what exists.

Within a year: it reaches everything with a circuit in it. The precedent is documented and it was far milder. A partial, temporary shortage of ordinary chips a few years ago stopped assembly lines across the automotive industry worldwide, for components that cost a few dollars each, and the lost production was counted in millions of vehicles. That was a shortage of mature semiconductors, not of advanced ones, and it corrected itself in months.

The modelled version of the larger event has been published. Bloomberg Economics estimated the cost of a conflict over the island at around $10 trillion, roughly ten per cent of global output, which is a larger shock than the pandemic, the war in Ukraine and the financial crisis of 2008. In its modelling, a conflict scenario reduces global output by 10.2 per cent, with Taiwan's own economy down 40 per cent, mainland China's by 16.7 and the United States' by 6.7. A blockade without fighting still removes 5 per cent of world output. The same analysis notes that about 5.6 per cent of total global value added, nearly $6 trillion, comes from sectors that use chips as a direct input.

And the part that cannot be fixed with money: it cannot be replaced. The second-largest foundry has 5.9 per cent of the market and the third 5.4. Running flat out and diverting every wafer, they do not approach 72.5 per cent. A new leading-edge plant takes three to five years and tens of billions of dollars, and that is to build the building and install the equipment. The yield, which is what actually determines whether the chips work, is accumulated over years of production and lives in people and in process data rather than in blueprints.

There is also a less obvious bottleneck. Even chips already manufactured and sitting in warehouses cannot become accelerators without advanced packaging, which joins the logic to its memory, and that capacity is concentrated in the same place. A supply of dies without a supply of packaging is a supply of nothing.

What it would do to artificial intelligence

The effect on artificial intelligence specifically is different from the effect on everything else, and more interesting, because it is delayed rather than immediate.

Close to ninety-nine per cent of the chips used to train frontier models originate in one place. If that supply stops, no new training capacity is built anywhere in the world, from that day, for as long as it takes to stand up an alternative. But the capacity that already exists keeps working, and there is a great deal of it. The models in service continue to serve. The research continues on the hardware already installed.

So nothing visible happens for months. What happens instead is that the curve flattens.

Every projection in this field assumes that the amount of computation available for training grows by a large multiple each year, and every plan for the next generation of models is denominated in that growth. Remove the supply of new accelerators and the industry is left with a fixed quantity of computation that then begins to decline, because hardware at that density and that utilisation fails at a measurable rate and the failures cannot be replaced.

The consequences, in order of arrival. Training runs get scheduled rather than commissioned, because the capacity is finite and shared. The price of inference stops falling, which reverses the single most important trend in the economics of the technology, the thousandfold collapse described elsewhere in this newspaper. Efficiency stops being a virtue and becomes the only remaining axis of progress, so research shifts from larger models to smaller ones, from training to extracting more from what is already trained. The companies that had been designing their own silicon discover that designing it was never the hard part.

And the plans that depend on abundance stop being plans. The orbital data centres, the gigawatt campuses, the hundred-billion-dollar commitments reported in these pages are all denominated in accelerators that would not exist.

The honest conclusion is twofold and both halves matter. The field would not die, because the installed base is enormous and because constraint has historically produced better engineering than abundance. But the specific future that the whole industry is currently financing, the one measured in gigawatts and in orders of magnitude, requires an uninterrupted supply from a small number of buildings, and there is no version of it that does not.

The blueprint: the machines, one by one

What a fabrication plant actually is, is a sequence of about a thousand process steps applied to a disc of silicon over three to four months, using roughly fifteen categories of machine. This is what each of them does and what physics limits it.

The crystal puller

The input is a disc of silicon that is a single continuous crystal, with no grain boundaries anywhere in it. It is made by the Czochralski method: a seed crystal is dipped into molten silicon at about 1,414 degrees and withdrawn slowly while rotating, so that the melt solidifies onto the seed in the same lattice orientation. The result is a cylindrical ingot, 300 millimetres across, weighing a couple of hundred kilos, which is then sliced, lapped and polished.

The specification that matters is flatness. Across a 300 millimetre surface, the variation permitted for the newest nodes is measured in nanometres, because the lithography that follows has a depth of focus of a few tens of nanometres and cannot print on a surface that wanders outside it.

The lithography scanner

This is the machine the industry is organised around, and it is the most complex object manufactured in series anywhere.

It weighs on the order of 180 tonnes, arrives in dozens of freight containers and takes months to install and calibrate. A standard extreme-ultraviolet unit costs about €180 million and the newest generation between €380 and €400 million.

Its light source is not a lamp. Molten tin is ejected as droplets at a rate of around fifty thousand per second. Each droplet is struck twice by a carbon-dioxide laser: a first pulse flattens it into a disc, a second vaporises it into a plasma at roughly 220,000 degrees, hotter than the surface of the sun. That plasma radiates at 13.5 nanometres, which is the wavelength the whole machine exists to produce.

Everything after that is a fight against the fact that extreme ultraviolet light is absorbed by essentially all matter, including air and glass. So there are no lenses. The entire optical path is mirrors, in vacuum. Each mirror is a stack of forty to fifty alternating layers of molybdenum and silicon, each layer a few nanometres thick, engineered so that reflections from each interface add in phase. Even so, each mirror returns about seventy per cent of the light, and with ten reflections between source and wafer only a few per cent of what was generated arrives. The rest becomes heat in the optics, which then have to be held to their figure while absorbing it.

Those mirrors are the flattest objects made by humans. Their manufacturer describes them as being, if scaled up to the area of a country, smooth to within a fraction of a millimetre.

The mask carrying the pattern is also reflective rather than transparent, for the same reason. The wafer stage and the mask stage then move in opposite directions in a synchronised scan at a four-to-one ratio, with the stages accelerating at several times gravity while holding position to a fraction of a nanometre. A modern unit processes on the order of one hundred and fifty to two hundred wafers an hour.

The newest generation raises the numerical aperture from 0.33 to 0.55, which improves resolution, and does it with anamorphic optics that magnify differently in the two axes. The consequence is that the printable field is halved, so large chips have to be printed in two exposures and stitched together.

The track

Bolted to the scanner is the machine that applies the photoresist, bakes it, and develops it after exposure. Resist is spun on as a liquid film of tens of nanometres and its chemistry is the quiet limit of the whole process.

Chemically amplified resists work by having each absorbed photon generate an acid molecule, which during a bake catalyses many further reactions, which is how a weak signal becomes a developable pattern. At 13.5 nanometres the photons are energetic and therefore few: the number arriving in a single feature is small enough that the statistical variation in their arrival, which is shot noise, shows up as roughness on the edge of the printed line. Below a certain size you are not fighting optics any more, you are fighting counting statistics.

The deposition tools

Layers are built up rather than carved. Chemical vapour deposition flows precursor gases that react on the hot wafer surface. Atomic layer deposition does it one atomic layer at a time: a precursor is admitted until the surface is saturated and the reaction stops itself, the chamber is purged, a second reactant completes the layer. That self-limiting behaviour is how a gate dielectric about one nanometre thick is laid down uniformly across 300 millimetres.

Physical vapour deposition, or sputtering, knocks atoms off a metal target with plasma and lets them condense on the wafer, which is how the metal layers are put down. Epitaxy grows crystalline silicon or silicon-germanium continuing the lattice of the wafer below, used to strain the channel and raise carrier mobility.

The etchers

Material is removed by plasma. In reactive ion etching the wafer sits on an electrode that develops a negative bias, so ions from the plasma are accelerated vertically into the surface, which makes the etch directional: it cuts down and not sideways, which is the only way to produce a vertical wall. Atomic layer etching does the inverse of atomic layer deposition, modifying one layer and then removing exactly that layer, for steps where a nanometre of overshoot destroys the device.

The implanter

Doping is done ballistically. Atoms are ionised, accelerated through tens of kilovolts to megavolts, mass-selected by a magnet so that only the intended species arrives, and fired into the lattice, where they come to rest at a depth determined by their energy. The impact damages the crystal, so the wafer is then annealed, often by a laser sweeping across it for milliseconds, long enough to repair the lattice and activate the dopant and short enough not to let it diffuse.

The polisher

After each stack of layers the surface is uneven, and lithography cannot print on uneven ground. Chemical-mechanical planarisation presses the wafer against a rotating pad flooded with a slurry that is simultaneously abrasive and chemically reactive, removing material until the surface is flat to within a few nanometres across the whole wafer.

The metrology and inspection tools

Every critical layer is measured. Scanning electron microscopes measure printed feature sizes. Optical and electron-beam inspection hunt for defects a few tens of nanometres across on a 300 millimetre surface, which is the equivalent of finding a grain of sand on a football pitch, repeatedly, at speed. Overlay metrology verifies that each layer is aligned to the one beneath it within a couple of nanometres, because a transistor whose contact misses its gate is not a slow transistor, it is an open circuit.

The transport system

The wafers never travel by hand. They move in sealed pods on overhead rails threaded through the ceiling of the plant, dispatched by scheduling software, covering tens of kilometres inside the building over the months of the route. The cleanroom they move through is held at particle counts millions of times lower than ordinary air, with the air changed hundreds of times an hour in laminar flow from ceiling to floor, and the whole structure isolated from ground vibration because the scanner is aligning to fractions of a nanometre.

The transistor being built

The architectures have changed three times. The planar transistor, a gate lying on top of a channel, ran out when the gate could no longer turn the channel off. The fin structure stood the channel up as a vertical blade so that the gate could wrap it on three sides. The current generation stacks the channel as horizontal sheets with the gate wrapping all four sides, which is what is meant by gate-all-around or nanosheet, and it is the structure in the two-nanometre processes now ramping.

Underneath those are two other changes. The gate insulator stopped being silicon dioxide, because below about 1.2 nanometres it leaks by quantum tunnelling, and became a hafnium-based material with a higher dielectric constant. And power delivery is moving to the back of the wafer, so that the front side carries only signal wiring and the power rails come up from behind, which reduces the voltage lost in the wires.

Above the transistors sit fifteen to twenty layers of copper interconnect, each patterned, filled and polished in turn. At current dimensions the delay of the wiring dominates the delay of the transistors, which is why so much of the engineering effort goes into layers that do no computing at all.

The packaging line

A finished wafer is still not a product. The dies are cut out, and the ones that will become accelerators are placed on a silicon interposer carrying vertical connections through its own thickness, alongside stacks of memory, and the assembly is mounted on a substrate. That is the step that is currently the industry's binding constraint.

The newest joining method dispenses with solder entirely: two surfaces of copper and dielectric are polished flat, brought into contact and annealed so that the copper pads fuse directly, which permits connection pitches below ten micrometres and the vertical stacking of dies.

The equation that governs the economics

All of the above exists to serve one number. Yield falls approximately as the exponential of defect density multiplied by die area, which has a consequence that decides the industry: for the same manufacturing quality, a large chip yields far worse than a small one, and the penalty is not proportional but exponential.

That single relationship explains why an accelerator of eight hundred square millimetres is so much more expensive to make than a telephone processor of a hundred, why the industry moved to assembling big products out of smaller dies, and why a manufacturer with slightly lower defect density does not have a slightly better business. It has a different business.

Down to the bottom: the parts of the parts of the parts

The machines above are themselves assembled from components that are, individually, as difficult as the machines. Following the chain downwards is the only way to see where the real chokepoints are, and the bottom of it is not where anyone expects.

The scanner, taken apart

The drive laser. The source of extreme ultraviolet light begins with a carbon-dioxide laser of tens of kilowatts, built by a German machine-tool company, consisting of a seed oscillator and a chain of amplifiers whose beam is transported tens of metres and aimed with an accuracy of microradians at a moving target the width of a human hair. It fires twice per droplet, fifty thousand times a second, and must hit a falling sphere of molten metal both times.

The tin. The droplet generator ejects molten tin through a nozzle at a controlled rate and size, around thirty micrometres, with a stability that has to hold for weeks of continuous operation. The plasma it produces sprays tin debris in every direction, which lands on the collector mirror beneath it, which is why hydrogen is flowed through the vessel continuously to carry the tin away as a volatile compound.

The collector. An ellipsoidal mirror roughly the width of a dustbin lid sits around the plasma, gathering the light that is emitted in all directions and focusing it into the machine. It is coated in the same multilayer as the rest of the optics and it lives inside the debris field, which means it degrades and is replaced.

The projection optics. The assembly that carries the image from mask to wafer is made by a German optics company in which the scanner manufacturer bought a quarter stake in 2016 precisely in order to secure it. It contains six to ten mirrors.

The mirror, taken apart

The substrate. Each mirror begins as a block of glass-ceramic engineered to have almost no thermal expansion, so that absorbing the heat of the discarded light does not change its shape. Two materials in the world are used: a lithium-aluminosilicate glass-ceramic from a German manufacturer, and an ultra-low-expansion titania-silica glass from an American one. Both work by balancing a phase that expands against a phase that contracts, so that the net change with temperature is close to zero. The raw blocks are cast and then annealed for months, because cooling them quickly would introduce stress that no amount of later polishing can remove.

The figuring. The block is ground and polished conventionally, and then corrected by ion beam figuring: a focused beam of argon ions is scanned across the surface under computer control, removing material atom by atom, dwelling longer where the surface is high. Between passes the shape is measured interferometrically and a new dwell map is computed. The specification for a mirror working at this wavelength is a figure accuracy and a roughness on the order of 0.1 nanometres root mean square, which is approximately the diameter of a single atom, held across a surface the size of a dinner plate. Each mirror takes months.

The coating. The finished surface is then coated with around fifty alternating layers of molybdenum and silicon, each a few nanometres thick, deposited by sputtering with thickness control measured in picometres, because the reflection works by constructive interference and an error of a few per cent in layer thickness destroys it.

The mask, taken apart

The pattern itself is carried on a mask, and the mask is as hard as the optics.

The blank is a low-expansion glass substrate carrying the same multilayer reflector, and it must be free of defects, because any flaw in the mirror beneath the pattern prints on every wafer it is used for. Two Japanese companies supply essentially all of them.

The writing is done by electron-beam systems that expose the absorber pattern directly, with the newest generation using many beams in parallel because a single beam would take too long. One company in Austria dominates that tool.

The pellicle is a membrane stretched above the mask to keep particles off it, which must be strong enough to survive handling, thin enough to transmit the light, and stable under the power passing through it. It is on the order of fifty nanometres thick.

The wafer, taken apart

The silicon. It begins as quartzite reduced with carbon in an electric arc furnace to metallurgical silicon, which is then converted to trichlorosilane, distilled, and decomposed onto heated rods to produce polysilicon of a purity measured in parts per billion. That polysilicon is melted and pulled into a single crystal.

And the crucible it is melted in. The molten silicon sits in a crucible of fused quartz, and because the melt slowly dissolves the crucible, the purity of that quartz sets the purity of the crystal. The inner layer requires high-purity quartz of a grade that essentially one place on Earth supplies.

That place is a mining district around a town of about two thousand two hundred people in the mountains of North Carolina, which produces something between seventy and ninety per cent of the world's high-purity quartz and around eighty per cent of the semiconductor grade. Two companies work it, one Belgian-owned and one Norwegian. The deposit was formed some three hundred and eighty million years ago when two tectonic plates collided and produced pegmatite of extraordinary chemical purity fifteen miles below the surface, and no geological equivalent has been found anywhere else.

The crucibles are consumables. They are replaced every few hundred hours.

So the recursion terminates here: the most advanced manufacturing in human history rests, at its base, on sand from one county, mined by two firms, consumed continuously.

The chemistry, taken apart

Photoresist is a polymer, a compound that releases acid when struck by a photon, and a solvent. The solvent chain runs back to naphtha, which is a petroleum fraction, which arrives by sea. Four Japanese companies account for about three quarters of the world market.

The process gases are their own chain. Excimer lasers for the older lithography consume neon, most of which has historically been produced as a by-product of steelmaking in a small number of plants. Etching and deposition consume tungsten hexafluoride, hafnium compounds for the gate dielectric, and a range of fluorinated gases. Cooling and leak detection consume helium, which comes from a handful of natural gas fields.

How you would actually obtain any of it

The question that follows is procurement, and the honest answer is that almost none of it is bought in the ordinary sense.

The scanner is allocated, not sold. One company makes it, its order book runs years ahead, and its output is distributed among customers it has supplied for a decade. Beyond that, the machine is subject to export licensing by the government of the country where it is made, so whether a given buyer may receive one at all is a political decision rather than a commercial one.

The optics are secured by ownership. The scanner manufacturer did not rely on a contract for its mirrors; it bought roughly a quarter of the company that makes them. That is the pattern that recurs at every level of this chain, and it is the same instrument described elsewhere in this newspaper: where a supply cannot be replaced, the buyer takes equity.

Capacity is secured by prepayment. Customers of the foundry commit years in advance, and the foundry commits to its own suppliers the same way, which is how a materials company justifies building a plant for a product that does not yet have demand. Money moves up the chain long before product moves down it.

Materials are secured by long-term contract and by stockpiling, which is the only defence available against a supplier that cannot be duplicated. After the hurricane that closed the quartz district in 2024, every buyer in the world discovered simultaneously how many months of crucible supply it was holding.

And the people are secured by moving them. The overseas plants described above were started by relocating hundreds of engineers from the original site, because the knowledge is not written down anywhere that would allow it to be transferred by any other means.

The summary is uncomfortable and precise. At every level of this chain, from the quartz to the finished accelerator, the mechanism by which supply is obtained is not price. It is allocation, prepayment, equity and licence. The market, in the sense of a buyer choosing among sellers, barely exists anywhere in it.

How you would found one, from beginning to end

The question that follows from everything above is whether this can be done again, and by whom. The answer is specific, and it is worth setting out as a sequence rather than as an opinion, because each step has a number attached to it.

Step one: the capital, before anything exists

A single leading-edge fabrication plant costs in the region of twenty to thirty billion dollars to build and equip. For scale, the company described here is spending between $60 billion and $64 billion in 2026 alone, and has committed around $265 billion to its American expansion for a planned dozen plants and four packaging facilities.

That money is spent before a single chip is sold. There is no stage at which a smaller version works, because the economics depend entirely on volume: the equipment costs the same whether you run it at ten per cent or ninety.

Step two: get in the queue for the machines

You cannot buy your way past this one. The machine that prints the smallest features is made by a single company with one hundred per cent of that market, which expects to ship a little over sixty units worldwide in 2026, of which between five and ten are the newest generation at €380 to €400 million each. Its order book exceeds €38 billion.

Those units are allocated to customers who have been buying for a decade and who are expanding. A new entrant joins the back of a queue measured in years. And whether you are permitted to receive one at all is decided by export control regimes rather than by the seller.

Step three: land, power and water, guaranteed for decades

A leading-edge plant consumes electricity and ultrapure water on a scale that commits a region. Sites are chosen for guaranteed supply of both over the life of the asset, which is why these decisions involve national governments rather than planning departments, and why the plants announced in the last three years are accompanied by state subsidies in every jurisdiction where they are being built.

Step four: the people, which is the real wall

A plant of this kind requires thousands of process engineers whose expertise exists only inside other plants of this kind. It cannot be trained from textbooks, because what is being learned is which of tens of thousands of parameters matters, and that knowledge is held by people who are already employed.

The company described here has had to move hundreds of its own engineers abroad to start its overseas plants. That is the clearest possible statement of where the asset lives.

Step five: lose money for years on purpose

Yield, as set out above, is accumulated rather than purchased. A new process begins below the threshold at which production is economic, and improves by running wafers and studying the ones that fail. Every one of those wafers is paid for by the operator.

This is the step that eliminates almost everybody, because it requires an owner willing to fund years of losses in exchange for a position that may never arrive.

Step six: the ecosystem, which is a cold-start problem

No customer can design a chip for your process unless you supply the design rules, verified libraries and simulation models that let them do it, and unless the two or three design software companies support your process in their tools. Building that costs money and requires volume. Volume requires customers. Customers require the design environment to exist first.

Every new foundry faces that circle, and the only way through it is to pay for the ecosystem before there is anything to sell.

Step seven: win the customer two generations early

Designing a chip for a process takes one to three years. A customer committing to you is committing before your plant has proved anything, and if you miss your schedule their product does not exist. The first customer therefore has to be either desperate, subsidised, or your own corporate parent.

How the one that worked actually did it

The company in this article was founded in 1987, and the lesson of its founding is not technological.

The state provided 48.3 per cent of the start-up capital. The process technology was obtained through a partnership with an established European manufacturer rather than invented from nothing. And the innovation that made it succeed was a business model rather than a process: it would manufacture for others and never compete with them, which no existing chip company could credibly promise because they all sold chips of their own.

That combination, state capital, transferred technology and a structural promise nobody else could make, is the actual recipe, and it is why the attempts since have taken the same shape. The Japanese project now building towards a two-nanometre process is state-backed with technology from an American partner. The Chinese manufacturer gaining share is state-backed and works without access to the critical machine. The American company attempting to re-enter is an existing manufacturer trying to convince the market it will not compete with the customers it is asking for.

And the door that is actually open

If the leading edge is closed, and it is, the useful part of the answer is where the industry is still foundable, because capacity is not only built at two nanometres.

Mature and specialty processes, which make the chips in cars, appliances, industrial equipment, power systems and sensors, run on equipment that is decades old, is widely available second hand, and costs hundreds of millions rather than tens of billions. The shortage that stopped the world's assembly lines a few years ago was a shortage of these, not of advanced logic.

Packaging and testing is the second door, and at this moment it is the bottleneck of the entire industry rather than a sideline. It is less capital-intensive than wafer fabrication by an order of magnitude and the demand is visible in everybody's published plans.

And the third is specialisation: power semiconductors, radio frequency, photonics, sensors, materials. Each is a real business with real customers and none requires a twenty-billion-dollar building.

That is the honest shape of it. The centre of the board cannot be entered. The edges of it are being entered constantly, by companies nobody writes about, and several of them are already indispensable to the firms described in this article.

What the filings establish

Three things, none of them opinion.

That the most consequential manufacturing company in the world has a revenue concentration of 78 per cent in ten customers, several of whom are attempting to destroy each other, and is therefore indifferent to the outcome of the competition that defines this decade.

That the transition to artificial-intelligence infrastructure is visible in its accounts earlier and more cleanly than in anyone else's, because every participant has to pass through it: 43, then 51, then 58 per cent.

And that it is spending between five and seven points of its own gross margin, deliberately and with advance notice to its owners, to be in more than one place and to be first on the next process. Those are the two risks it cannot insure against, so it is buying them down with the only currency it has in surplus.

Everything else in the artificial-intelligence economy is a question of who wins. This is the one position in it where that question does not arise.

Sources

Filed documents. The company's annual report on Form 20-F for the 2025 financial year, filed with the United States Securities and Exchange Commission in 2026, for customer concentration, revenue by platform and geographic distribution of revenue (filing). Quarterly results and the accompanying management presentations for the first and second quarters of 2026, published by the company, for revenue, margins, guidance, capital expenditure allocation and the stated margin dilution from the two-nanometre ramp and from overseas plants (investor relations).

Ownership and joint ventures. Shareholding figures are from the company's published shareholder information as of 31 July 2026 and from its listing documents; the founding capital contribution of 48.3 per cent by the state fund is a matter of record. Equity stakes in the Japanese and European joint ventures are as announced by the participating companies.

Scenario modelling. The estimated cost of a conflict over Taiwan, including the figures of $10 trillion, 10.2 per cent of global output under a conflict scenario and 5 per cent under a blockade, and the observation that about 5.6 per cent of global value added comes from sectors using chips as direct inputs, are from Bloomberg Economics modelling published in 2024. The scenario in this article is a construction built on those magnitudes and names no country as a cause.

Materials and physics. The Czochralski process, chemically amplified photoresist and extreme-ultraviolet shot noise, reactive ion etching, atomic layer deposition and etching, ion implantation and annealing, chemical-mechanical planarisation, gate-all-around transistor architecture, high-permittivity gate dielectrics and backside power delivery are described as documented in the standard semiconductor process literature. Mirror substrate materials of near-zero thermal expansion, ion beam figuring, and the figure and roughness specification on the order of 0.1 nanometres root mean square for extreme-ultraviolet optics are as published in the optical engineering literature. The relationship between yield, defect density and die area follows the Poisson model standard in the field.

High-purity quartz. The concentration of semiconductor-grade high-purity quartz supply in the Spruce Pine district of North Carolina, its two operators, and the geological origin of the deposit are as documented in industry and press reporting following the disruption of 2024.

What is not reported here. No yield figure appears in this article. Yield is not disclosed by any foundry, and every figure in circulation is an estimate produced by supply-chain research firms. The same applies to the division-level results of competitors that do not report their foundry operations separately.

Press reports, identified as such in the text. The figure for the share of the world's most advanced chips manufactured in one place, and the estimate of close to ninety-nine per cent for chips used to train frontier models, are attributed to Counterpoint Research as reported in 2026. Figures for the total committed investment in Arizona and the number of planned plants there are drawn from company announcements and trade coverage of them.