Every AI Chip Is Made In One Place

PUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDY

By

Nvidia does not build the chips that run the AI boom. Neither does Apple, or AMD, or Google, or Amazon. They design them. They write the blueprints for the most complex objects humanity manufactures, and then they send those blueprints to a single company on the other side of the planet and wait for the boxes to come back. That company is Taiwan Semiconductor Manufacturing Company, and it fabricates roughly ninety percent of the world's most advanced chips. Every frontier AI accelerator, every flagship phone processor, every leading-edge laptop chip traces back through one set of factories run by one firm. The most important supply chain of the century does not branch. It funnels, and at every stage it narrows toward a single point.

The Foundry

In 1987 an engineer named Morris Chang founded TSMC on an idea the rest of the industry thought was backwards. Until then a chip company built what it designed and designed what it built. Chang proposed to build only, and to design nothing. TSMC would be a pure foundry, a contract manufacturer that printed anyone's design and competed with none of its customers. The model freed a generation of companies to design chips without ever owning a factory, and it quietly concentrated the factory half of the industry into the one firm that did it best.

Manufacturing at the leading edge compounds. Each generation of chip is harder to make than the last, the factories cost more, the yields start lower, and the only way up the learning curve is to already be on it. TSMC has been on it longer than anyone. By late 2025 it held about seventy percent of the entire foundry market and far more at the top, close to ninety percent of the world's three-nanometer production and a comparable grip on the two-nanometer node now ramping. Its rivals are not close. Samsung trails at a fraction of its advanced-node share, and Intel, once the most advanced manufacturer on Earth, has reportedly ordered its own cutting-edge chips from TSMC. Nvidia's Blackwell accelerators, Apple's M-series processors, and AMD's newest CPUs are all fabricated by TSMC. Thousands of companies design chips. One company builds the ones that matter.

PUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDY

The Machine

TSMC cannot print a leading-edge chip by itself. To draw circuits a few atoms wide it needs extreme ultraviolet lithography, a way of patterning silicon with light at a wavelength of 13.5 nanometers. Exactly one company on Earth makes the machines that generate that light. It is a Dutch firm called ASML, and without it the modern chip industry stops at a wall it cannot cross.

Inside the machine, a stream of molten tin droplets falls through a vacuum, and a high-power carbon-dioxide laser strikes each droplet twice, fifty thousand times a second, vaporizing the tin into a plasma roughly forty times hotter than the surface of the sun. That plasma emits the 13.5-nanometer light. Because extreme ultraviolet is absorbed by air and even by glass, the light cannot be bent through lenses. It is steered by mirrors from the German optics maker Zeiss that rank among the smoothest objects ever made, polished so finely that enlarged to the size of a country their tallest bump would measure a fraction of a millimeter. The entire light path runs in vacuum. ASML spent more than six billion euros across seventeen years building the system, shipped its first production unit in 2013, and now sells each machine for well over a hundred million dollars, with the newest generation priced near three hundred and eighty million. No machine on Earth costs more to build, and no company but ASML knows how.

TSMC makes the chips, and ASML makes the one tool TSMC cannot work without. In October 2022 the United States turned that fact into a weapon, barring the export of advanced chipmaking equipment to China and forbidding ASML from selling its extreme ultraviolet machines there at all. China has since poured tens of billions into catching up and remains, by the estimate of ASML's own chief executive, ten to fifteen years behind. The order book of a single Dutch company is now an instrument of foreign policy.

PUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDY

The Island

The deepest concentration is geographic. More than ninety percent of TSMC's most advanced capacity sits in one place, the island of Taiwan, which a far larger neighbor claims as its own territory and has refused to rule out taking by force. The dependence is so complete that strategists call it the silicon shield, the theory that the island is too valuable to the world economy for anyone to permit a war over it. The most important technology of the era rests on that theory holding.

Governments have noticed. The United States passed a fifty-two-billion-dollar CHIPS Act in 2022 to pull fabrication back onto American soil, and TSMC is now building advanced fabs in Arizona. It moves the needle less than the price suggests. The newest nodes and the research that produces them stay in Taiwan, and one Taiwanese industry institute projects that even in 2029, after the foreign fabs come online, the island will still hold sixty-one percent of the world's sub-six-nanometer capacity. The center of gravity does not move, because the talent, the suppliers, and the accumulated knowledge do not move with a building.

The chips that run modern artificial intelligence are designed in many places, but they are built by one company, on machines from one other company, mostly on one island within reach of a rival power. It is the most concentrated dependency in the modern economy, and it is the same fault line the brittleness of centralized systems traces everywhere else, drawn this time straight through the foundation of the entire industry.

Some of this cannot be undone. You cannot decentralize a twenty-billion-dollar fab or a three-hundred-and-eighty-million-dollar lithography machine across a network of ordinary buildings. Fabrication is the one layer of the computing stack that is irreducibly centralized, a chokepoint the world inherits rather than chooses. But a chip is only born in Taiwan. Where it runs for the next decade, who owns the computation it performs, and whether the power that feeds it collects in a few warehouses or spreads across many, are not settled inside a fab. Those layers are still open, and what gets built on them is the question this publication exists to answer.

PUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDYPUDDY

Last updated:

PUDDY