Hokkaido and Magdeburg

From the Megamachine series · Signpost to the Megamachine

Japan bought itself capability. Germany wanted to buy itself a tenant.

I. A Wafer in the Snow

On 18 July 2025, in Chitose on Japan’s northern island of Hokkaido, a man held a round disc of silicon up to the cameras. On it sat working transistors of the two-nanometre class, the finest semiconductor technology humans can currently make. Until then only three companies in the world had managed it: TSMC in Taiwan, Samsung in South Korea and Intel in the USA. Each of them had needed decades. The company now showing the wafer was called Rapidus. It was barely three years old.

Construction of the factory had only begun in September 2023. In December 2024 the lithography machine arrived from the Netherlands, a device the size of a bus. In April 2025 the first exposure ran. In July the transistors were alive.

A transistor works or it does not. In “Technical Intelligence” we described what this incorruptibility means: no press release, no expert opinion and no committee can replace the fact that the machine runs. On that July day it ran.

At about the same time, on 24 July 2025, Intel’s chief executive told his staff that the company would not build its planned chip factory in Magdeburg. The German government had promised around ten billion euros in subsidies for it. In Hokkaido, in those weeks, a capability came into being. In Magdeburg a field remained.

II. How Japan Lost the Lead

The story has a prehistory, and it is instructive. In the 1980s Japan dominated the semiconductor industry. About half of all the world’s chips came from Japanese factories, and their memory chips were so good and so cheap that American manufacturers got into trouble. In 1986 Washington responded with a semiconductor agreement that dictated volumes and prices to Japan.

But the trade war alone does not explain the fall. In 1987 Morris Chang founded a company in Taiwan with a strange idea: it would not design chips itself, only manufacture them, for anyone who needed them. That idea became TSMC. The big Japanese conglomerates stuck to their old model. Each built its own chips for its own televisions, cameras and computers. They had been too big, too interlocked and too successful to rebuild themselves. Japan’s share of the world market fell from about half to less than a tenth.

The state tried to rescue them. It merged the memory divisions of several conglomerates into a new company, Elpida, and propped it up with public money. In 2012 Elpida went bankrupt. The lesson was painful but clear: whoever merges the old structures gets a bigger old structure.

III. Start Anew, Don’t Restructure

Rapidus is the answer to that lesson. In 2020 the American company IBM offered to industrialise its research on two-nanometre transistors in Japan. IBM had the knowledge but no factory of its own any more. In August 2022 Rapidus was founded; the name comes from the Latin word for fast. Toyota, Sony, NTT, NEC, SoftBank, a large bank and other companies put in money. But none of them got the company. Rapidus is not a division, not a merger and not a rescue. It is a new foundation with its own mission.

That is the decisive difference from Elpida. No one tried to persuade the old conglomerates to change. Something new was started alongside them, and they were made to pay for it. The idea is familiar to us. In our working paper “Can the Young Untangle the Party?” we asked whether renewal can come from the entangled organisation itself or only from a place next to it that has nothing yet to defend. Rapidus is such a place.

The sceptics were loud. No company had ever started directly at the top. A country whose own production had stalled somewhere around forty nanometres wanted to skip twenty years of development. Some called it a fantasy at the taxpayers’ expense.

IV. Knowledge Travels with People

How do you skip twenty years? Rapidus gave an answer as old as craftsmanship. The company sent around 150 engineers to Albany in New York State, to IBM’s research centre, many of them experienced people from Japan’s old industry. They stayed there, some for years, working side by side with those who had developed the process. Then they came back to Hokkaido in waves.

What they learned there is in no manual. Between a process that works in the laboratory and one that works in a factory lie thousands of small decisions that are written down nowhere. You learn them only by standing alongside and taking part.

This is not a Japanese invention. In 1789 the young mill worker Samuel Slater left England, which banned the export of textile machinery and even the emigration of skilled workers. He took nothing with him, no drawing, no model. He had Richard Arkwright’s machines in his head because he had worked on them for years. In America he rebuilt them from memory. In England they called him “Slater the Traitor”, in America “the father of American industry”. The Republic of Venice, for its part, forbade its glassmakers on Murano to leave the lagoon, because it knew that its secret lay not in books but in hands.

Rapidus also placed a technical bet. In conventional factories silicon wafers are processed in batches. In Chitose each wafer is handled individually, and every step is recorded by sensors. That is more expensive per wafer, but it yields many times the data, and data is the fastest way to learn. The company is betting that at the start, the speed of learning counts for more than size.

In “Bigger Oars” we described how apparatuses economise on those who do the work. In this case Japan did the opposite. It did not send the steersmen travelling, but the rowers.

V. Magdeburg: Buying an Address

At the same time Germany chose a different path. In 2022 Intel announced it would build a chip factory in Magdeburg for more than thirty billion euros. The federal government promised around ten billion euros, the largest single subsidy in the history of the Federal Republic. In 2024 Intel put the project on ice; in July 2025 it gave it up. The company was itself in crisis and cutting jobs worldwide.

One has to be clear about what Germany wanted to buy. The factory would have been an Intel factory, with Intel’s process, run by Intel, according to decisions taken in California. Germany would have got jobs, taxes and a building. The knowledge would have stayed with the company. When the company leaves, the knowledge leaves with it, and that is exactly what happened, before the first stone was laid.

Meanwhile a factory is being built in Dresden in which TSMC holds seventy per cent and Bosch, Infineon and NXP ten per cent each. It is being subsidised with around five billion euros. From 2027 it is to make chips for cars and machines, at feature sizes TSMC has mastered for many years. That is sensible, because German industry needs exactly these chips. But here too the Taiwanese company runs the factory. Germany is the location, not the manufacturer.

The pattern goes beyond chips. In Heide in Schleswig-Holstein, the Swedish battery maker Northvolt wanted to build a factory with support from the federal and state governments. In March 2025 Northvolt went bankrupt. The politics of relocation always follows the same logic: you pay a foreign company to come. You buy an address. In “Plunder, Lease and Inheritance” we described how one can lease pieces of the state. Here it is the other way round: the state leases itself a tenant. If the tenant fails to arrive, the field remains.

VI. What Must Be Said in Fairness

Rapidus may fail, and fail thoroughly. TSMC’s lead is enormous. The Taiwanese company has long been making two-nanometre chips in volume, for Apple and the big artificial-intelligence firms, and its capacity is sold out for years. Rapidus runs a pilot line. Volume production is to start in 2027. Running a factory around the clock with stable yields, says one analyst, is among the hardest things in the industry: “Even Samsung has struggled.” Another asks who is actually going to fill the factory. Rapidus has no firm major customer so far. In October 2026 the company presented a network of seventeen partners to help customers design chips. That is progress, but not yet business. A government official says soberly that whether Rapidus succeeds will only be known in ten years.

The price is high. By March 2027 the Japanese state will have put the equivalent of around fourteen billion euros into Rapidus, and more will be needed for full volume production. Japan even changed its laws so that the state itself can hold shares. If the project fails, the country will have paid for one of the most expensive lessons in its industrial history. And TSMC is now building in Japan itself, in Kumamoto on the southern island of Kyushu, courting exactly the Japanese customers Rapidus needs.

Dresden is no mistake either. A factory for the chips found in every car is better than none. And Japan itself brought TSMC to Kumamoto with billions, so it does both: it buys relocations and capability.

So the difference does not lie in Japan doing everything right. It lies in what remains when things go wrong. If Rapidus fails, Japan is left with hundreds of engineers who know a two-nanometre process from the inside, a factory full of data and an industry that has learned. If a relocation fails, a field remains.

VII. What Remains

A country can put its money into the future in two ways. It can pay someone who already knows how, so that he comes. Or it can send its own people to where it is done, so that they learn it and bring it back. The first is faster and safer, as long as the other party stays. The second is more expensive and riskier, but what has been learned no one can take away again.

Japan experienced this itself in the 1980s, and Germany is experiencing it now. In “Manager Twilight” we wrote that Germany has stopped building carriages and only steers. That applies to its industrial policy too. It administers relocations, negotiates subsidy rates and invites foreign corporations. That knowledge is not bought but learned, and that people have to be sent to learn it, hardly features in it.

The wafer from Chitose does not yet prove that Japan will make it. But it proves something else: the lead everyone thought unassailable could be closed, and not with money alone, but with people who were allowed to learn. A country that buys itself tenants remains a landlord. A country that buys learning can one day build for itself.