TSMC gets the land for what comes after A14

TSMC has approval to expand the Longtan section of the Hsinchu Science Park for process technology at A14 and beyond. Three wafer plants. About 104 acres. A development cost reported at NT$95 billion, roughly $3 billion, before a single scanner. The first plant is pencilled for commercial production in 2030. Later phases of the same expansion are described as sub-A14 — whatever TSMC will call the node after 1.4 nanometres when it has to print it on a slide.

Intel is talking 14A for 2028. TSMC is talking A14 volume in 2028 and the node after that in 2030, on land it has only just been allowed to use. Those dates will move. The fact that does not: the next node is being sited in Taiwan, in a science park that already cannot hire as fast as it pours, while Arizona and Kumamoto are still ramping what TSMC already knows how to make.

Longtan is not a clean story. TSMC put off moving there in 2023 after residents pushed back. The park bureaucracy has been grinding the third phase through the National Science and Technology Council and the Executive Yuan for most of this year. Approval for the land is not approval for the tools, and it is not a workforce. Hsinchu’s technician market is the same market every other TSMC site in Taiwan is already bidding in. A new campus does not create a new labour pool. It tightens the old one.

3

Fabs in Longtan phase three

NT$95bn

Reported development cost, before tools

2030

First plant, commercial production

What the land approval covers

A science-park approval is a political document. It says the state will tolerate the traffic, the water, and the power draw. It does not say the ultrapure-water plant is sized, or that Taipower has a substation, or that the vocational schools will graduate the night shift. Hsinchu’s water politics are not new. Every additional fab in the northwest corridor is another claim on a dry island that already runs foundries like a national project. Arizona’s problem is the same shape in a different climate: the desert, the interconnect, the people. Taiwan’s version is denser and already built. That is why TSMC still puts the next node there.

Arizona Fab 21 is in high volume on nodes TSMC has already qualified. Kumamoto is the insurance policy Japan paid for. Neither is where A14-beyond will be debugged. Process integration still sits with the teams that have done the last four nodes in Hsinchu and Tainan. A 2030 production date on a node that does not yet have a public name assumes High-NA, or heroic multi-patterning, or both — and a mask path that, as of this month, has a 2031 pilot. The land and the lithography are not on the same clock.

Relative to N2, A14 is being briefed as more than 20 percent higher logic density, 10 to 15 percent more performance, 25 to 30 percent better energy use. Those are foundry numbers. They become die-cost numbers only when the process is in yield, the mask is stable, and the packaging can take the device. CoWoS is already sold out. A faster transistor on a substrate you cannot buy is a paper.

Suppliers into the park — vacuum, ultrapure water, construction, sub-fab, the people who keep a tool in spec — will feel 2030 this winter as RFQs, not as a ribbon. The building is the easy part. The hire is not. TSMC can get land. It cannot print technicians. The same McKinsey-for-SEMI shortfall that has Arizona, Texas and New York bidding for process techs applies, denser, in Hsinchu. A third phase does not create a fourth shift. It bids against the first two.

For a company that buys wafers, 2030 is a capacity conversation with a foundry already booked solid on N3. It is not a purchase order. Geography as insurance — Arizona, Kumamoto — is real and it is not the leading edge. The leading edge stays where the process integration team already sits. That, this week, is still Taiwan, on land the park has only just released.