Taiwan builds the AI backbone, someone else banks the money

Taiwan holds nearly everything the AI boom stands on. It barely gets a cut of the revenue. One side sells subscriptions and tokens; the other pours metal, glues substrates and spins fans. The money settles where the user is. Taiwan gets what's left.

For TSIA's 30th anniversary, DigiTimes laid out an inconvenient map. The island already builds the physical layer of the AI boom: advanced logic, packaging, substrates, testing, power delivery, cooling, server assembly. All of it sits inside one unusually dense industrial base. Nowhere else comes close.

What actually holds the island up

The list looks boring only at first glance:

• advanced logic — the heart of any accelerator;
• packaging — where separate dies become one chip;
• substrates — the thin part nobody can source fast enough;
• testing — without it, no batch ships;
• power and cooling — where data centers melt;
• server assembly — the final production line of the AI era.

Every line is a separate business. Together they form an industrial base you cannot copy in a year or relocate by decree. The US and Japan keep trying. It is slow going.

Where the trap sits

Assembling a server is not the same as selling compute. Margin lives in the chip, the platform, the subscription. Taiwan sells parts, not outcomes. The more powerful AI gets, the pricier the parts — and the deeper the dependence on someone else's demand. A customer can walk. A part sits in the warehouse.

Then there is risk number two. Clients do not want the whole chain on one island. Relocation is no longer talk; it is construction. The faster even part of it moves, the less leverage Taiwan keeps at the table.

Why this matters now

The question has changed. It used to be: who builds the fastest chip. Now it is: who holds the bottleneck. Today the bottleneck is packaging, substrates, cooling. Those are made in Taiwan. That is leverage. But leverage only works when you use it.

Right now Taiwan plays contractor. Reliable, irreplaceable — still a contractor. Fine while the client pays. Bad the day the client finds a second address.

There is one way out: sell the system, not the component. Own racks, own clouds, own models running on your own silicon. Then margin stays on the island instead of flying into somebody else's earnings call. Plenty of talk, fewer products. DigiTimes puts the question right in the headline: Taiwan can build the AI backbone, but it cannot yet cash in on it.