The AI Pipeline, Part 2: One Machine and One Island
This is part 2 of the AI Pipeline series. Part 1 covered the ground floor: the sand and minerals, and the three-company EDA layer that draws every chip on x86 and ARM architectures. This part covers what happens next, and it is the layer where the concentration gets almost comical: one Dutch machine prints the chips, one Taiwanese company fabricates most of them, three companies sell the memory, and the supporting cast includes a food company you have in your kitchen.
A century of shrinking switches
Worth a moment of history to appreciate why the machine matters. In the 1940s, computers were built from vacuum tubes, which made machines room-sized, hot, unreliable, and slow. The transistor, invented in 1947, replaced them with a solid-state switch, and by the late 1950s the problem had inverted: transistors were so small and reliable that wiring thousands of them together became the bottleneck, the problem the literature called the tyranny of numbers. The integrated circuit, built independently by Jack Kilby and Robert Noyce, put the wiring inside the silicon itself. Kilby's first version was germanium; Noyce's, the first monolithic silicon IC, is the direct ancestor of everything in your pocket.
The Intel 4004 in 1971, the first commercial microprocessor, had 2,300 transistors. From there, Moore's Law took over: roughly a doubling of transistor count every two years, sustained for five decades, from 2,300 transistors to over a hundred billion on a single AI GPU today. That compounding is the entire material basis of the AI boom.
And it only worked because the printing resolution kept shrinking, which meant the light used to print kept moving to shorter and shorter wavelengths: G-line, I-line, deep ultraviolet, and finally extreme ultraviolet. Every wavelength transition forced the industry to reinvent its light sources, its optics, and its photochemistry. EUV at 13.5nm was the hardest transition yet, and it is the reason one company in the Netherlands now sits at the center of the world.
The machine that rules the world
Here is a fact I think about constantly. The most important company in the AI supply chain is not NVIDIA, not OpenAI, not Google. It is ASML, a Dutch company from a town of about 48,000 people (Veldhoven), whose machines print the patterns that become chips. ASML is the only company on earth that makes EUV lithography systems, the ones that can print at 13.5 nanometre wavelengths, structures thousands of times finer than a human hair.
An EUV machine weighs about 150,000 kilograms, contains around 100,000 parts, and ships in roughly 40 freight containers plus three cargo planes and 20 trucks. Assembly on site takes about 250 engineers six months. The newest High-NA version costs around $400 million, more than a Boeing 787-9 Dreamliner. Inside it, light is generated by firing a laser at tin droplets moving at kilometers per second, tens of thousands of times per second. The physics of getting usable light at 13.5nm is the reason every older wavelength approach failed and decades of prototypes had to die before the first production machine worked.
The mirrors that shape that light are the smoothest objects ever manufactured: polished to below one atom of roughness, equivalent to a mirror the size of Germany with a single one-millimeter bump on it. And those mirrors come from exactly one company, Carl Zeiss SMT, which ASML's own annual report describes as the sole supplier of critical lenses, mirrors, illuminators and collectors, whose prolonged interruption would effectively halt production.
The economics of this monopoly are obscene and entirely rational. Decades of failed prototypes, unproven physics, and supplier co-investment went into EUV before the first production machine existed. There is no fast follower because there is nothing to copy cheaply. Which is why the export-control fights, which part 3 gets into in detail, all route through one building in Veldhoven.
The fab: one company prints the AI world
Design a chip on your laptop, send the file to a fab, and the fab prints it. Except there is effectively one fab that can print the advanced stuff. TSMC holds around 70-73% of the entire foundry market (Samsung has about 7%, and everyone else fights over scraps), and over 90% of the most advanced AI chips. In Q1 2026 the split was TSMC 73%, Samsung 7%, SMIC 5%, UMC 4%, GlobalFoundries 3%.
Why so lopsided? Yields and timing. Samsung's 4nm yields ran as low as 35% against TSMC's roughly 70%, which is why Qualcomm took its flagship Snapdragon back to TSMC. GlobalFoundries quit the 7nm race entirely in 2018 because the money did not work, and now makes mature and specialized chips. Intel is trying to become a foundry competitor with its Intel Foundry business, but it is climbing a mountain the leaders have already descended. Every new node costs more, the machines cost more (that $400M High-NA tool again), and only one customer base is big enough to amortize it: the AI industry, which keeps paying whatever TSMC asks.
The moat extends past the fab. Advanced packaging, the step that bolts GPU dies and memory stacks into one module (TSMC's CoWoS process being the standard for AI chips), is around 90% TSMC too. The bottleneck for AI GPUs in the last two years was not lithography, it was packaging capacity. It is also a market heading for about $80 billion by 2033, with the OSAT players (ASE, Amkor, JCET) holding about 59% of the broader advanced packaging market while TSMC keeps the AI-critical front-end to itself. Demand shows up everywhere: Foxconn's Q2 2026 profit rose 35%, driven by AI server and packaging work.
Where TSMC actually is on the curve: its 2025 annual report shows 3nm alone at 24% of wafer revenue, with the 2nm ramp underway and TSMC Arizona rising in the US. The company's scale means a single node can carry a quarter of a fab giant's revenue.
There is a business-model reason TSMC stays dominant that gets overlooked: it is a pure-play foundry, meaning it does not compete with its own customers. Samsung and Intel both sell their own chips, so every fabless designer that hands TSMC a design is handing it to a company that will never turn around and compete with them in the market. Neutrality is the actual product. That, plus the yield compounding, is why the "silicon shield" language around Taiwan exists, and why the whole world is now funding TSMC fabs in Arizona and Japan as a hedge.
Memory: the quiet oligopoly
AI compute is as much a memory story as a compute story. Token generation is bandwidth-bound, which is why HBM (high bandwidth memory, stacks of DRAM dies sitting next to the GPU die) governs how fast an AI chip actually runs. The next generation designs fuse the GPU and HBM even tighter.
HBM supply is a three-company oligopoly: SK hynix, Samsung, and Micron. At the end of 2025 the split was roughly 57% SK hynix, 22% Samsung, 21% Micron, and SK hynix has been the leader with around 62% share, riding an HBM-led memory supercycle. South Korean firms together hold roughly 75-85% of the next-generation HBM market, worth on the order of $55 billion. Samsung shipped the first commercial HBM4 and unveiled HBM4E with NVIDIA at GTC 2026. Micron, meanwhile, broke ground on an HBM advanced packaging facility in Singapore to add capacity. The broader DRAM market is the same three names, an oligopoly so settled that it behaves like a utility with pricing power.
The hidden layer: monopolies inside the monopoly
And the periphery of this periphery is even more concentrated, which is the part almost nobody talks about. Japan supplies about 88% of the world's coater/developer machines, 53% of silicon wafers, and 50% of photoresists. One Japanese company, Ajinomoto (yes, the MSG company), makes the build-up film used in high-end package substrates, with roughly 95% dependence on a single supplier, and the EU's own supply-chain analyses flag that dependence as a vulnerability. Another Japanese company, Nittobo, controls about 90% of the T-glass cloth that goes into substrates.
Sit with that for a second. The AI buildout of 2026 can be slowed by a food company's film division and a specialty glass fiber cloth. The AI pipeline is not a chain of a few monopolies. It is a chain of monopolies inside monopolies, and several of the narrowest ones have no second source at all.
The chips, once made, get bought by an even smaller set of companies than the suppliers, and that is where the geopolitics goes from background to open warfare. That is part 3.