Do not pave the way for competitors

Now that businessman Elon Musk has approached Taiwan Semiconductor Manufacturing Co (TSMC) to discuss collaboration on his Terafab project, the issue Taiwan should be focused on is not whether TSMC would establish more factories in the US, but whether cracks are beginning to appear in the division of labor underpinning the chip foundry.

TSMC is evaluating a potential investment in Texas — but the structure of the partnership, ownership stakes, manufacturing processes and timeline have yet to be determined, Reuters said.

It is too early to conclude that TSMC’s potential Texas fab would be Terafab, but when a hyperscale artificial intelligence (AI) customer wants not only to buy chips, but also to own a semiconductor fab, the larger question is: Is TSMC dealing with a major customer or future competitor?

The Terafab project, a massive chip factory designed to build more than 1 terawatt of AI computing power annually, is different from a typical start-up foundry. Tesla’s job openings describe it as a highly vertically integrated semiconductor manufacturing hub, bringing logic chips, memory, advanced packaging, testing and photomask production under one system, and requires 2-nanometer-class Gate-All-Around process capabilities. That would bring Tesla’s autonomous driving and robotics, SpaceX’s space computing, and the AI computing needs of Musk’s AI start-up, xAI, into a single supply chain. When a company needs enough computing power to support an entire fab, semiconductor manufacturing could shift from an outsourced project to strategic infrastructure.

However, a fab cannot simply be replicated with money. Extreme ultraviolet lithography machines can be purchased and fabs can be built, but the hardest things to acquire are process integration, yield ramp-up, equipment management, materials qualification and the discipline of mass production accumulated over decades. This is precisely why Musk needs TSMC. If the arrangement follows a model similar to Japan Advanced Semiconductor Manufacturing Inc or European Semiconductor Manufacturing Co, with the customer taking an equity stake while TSMC retains majority ownership and operational control, it would still be an extension of an existing business model.

However, if Terafab holds the controlling stake while TSMC provides process technology and operational expertise, the implications would be entirely different. TSMC’s most valuable asset is its ability to integrate processes, equipment, materials, yields and customer confidentiality into a manufacturing system capable of sustained operation. If that capability is exported to a customer-controlled factory, the moat of the chip foundry would be redefined.

Taiwan should not focus solely on how much advanced process technology is being moved overseas. It should also ask who controls the fab, which generation of process technology is being transferred and whether the other party is gaining access to production capacity or manufacturing capabilities. Manufacturing for a customer overseas and assisting that customer in acquiring the ability to manage an advanced fab are vastly different things. What the nation needs to preserve is its Taiwan centered ecosystem of research and development, process integration, key talent and supply chain collaboration.

Musk on Saturday last week wrote on X: “Just discussions, but something may come of it.” If the outcome is simply another major customer, TSMC would secure a long-term order. However, if it instead helps a customer develop advanced semiconductor manufacturing capabilities, what would be rewritten is not merely the story of a fab in Texas, but the division of labor that TSMC founder Morris Chang (張忠謀) created.

TSMC can manufacture more chips for the AI era, but it should not, in building fabs for its customers, pave the way for competitors.

Liao Ming-hui is an assistant research fellow at the Chung-Hua Institution for Economic Research.