In the particularly complex semiconductor manufacturing chain, foundries play a central role. TSMC, the Taiwanese sector giant, has quickly established itself as one of the main beneficiaries of the surge in artificial intelligence.Thanks to its ability to produce the world's most advanced chips at the best price, its dominant position is unquestionable... although not uncontested. Some clients are looking for alternatives, while the race to ramp up output increases the risk of an imbalance between supply and demand. To make matters worse, Washington's push for Made in the United States is disrupting this well-oiled machine.

A long-time client on the way out?

TSMC relies mainly on two major clients: Nvidia, which delivers a sizeable margin, and Apple, whose regular orders provide financial stability and help justify the opening of new plants. Long an exclusive partner of the foundry, the Cupertino company nonetheless appears to be exploring other options.

According to a SemiAnalysis report, Apple is considering diversifying its suppliers. While nothing has been decided yet, such a move could weaken TSMC's revenue, regardless of how demand for artificial intelligence evolves. Apple has already signed a contract with Samsung for the supply of CMOS image sensors, manufactured in the US, from 2027. In parallel, the company is also taking an interest in Intel's 14A process, which could power its future A16 chips. According to SemiAnalysis, the probability that Apple will favor Intel remains low (30%), but it is not negligible.

A production chain under strain

TSMC faces a constant dilemma: preserving profitability while meeting exponential demand. The rise of artificial intelligence does not remove this challenge, quite the contrary. To meet growing needs, the foundry must continuously open new production lines. Yet these are only economically viable if they run at at least 80% of capacity.

So when Nvidia wants to increase output, TSMC must ensure that this demand will last. That requires tracing the chain upstream, all the way to hyperscale infrastructure (Amazon Web Services, Alphabet, CoreWeave, Oracle, Meta, etc.), to confirm a durable volume of orders. This process shows just how complex it is to maintain a high level of production, but also highlights TSMC's vulnerability: a sudden drop in demand from cloud giants would trigger a domino effect, hitting Nvidia and, in turn, the foundry's profitability.

A costly bet in the United States

Another sticking point for TSMC: its footprint on US soil. In a tense geopolitical context, the foundry has been trying to meet the US administration's expectations by building several plants in the United States. But this diplomacy-driven strategy is proving economically risky.

A SemiAnalysis report, cited by BusinessKorea, reveals a yawning gap between production costs in Taiwan and those in Arizona. Manufacturing a 5 nm wafer costs $6,681 in Taiwan, versus $16,123 in the United States, i.e. 2.4x more. The reasons: higher wages, materials that are twice as expensive, significant procurement costs and depreciation that is four times higher. As a result, gross margin falls to 8% on US soil, compared with 62% in Taiwan. Constraints that TSMC's management had already flagged before this expansion began, but geopolitical risks, along with generous subsidies, ultimately carried the final investment decision.

Under these conditions, TSMC's expansion in the United States implies a significant shortfall and much higher CAPEX. While the company remains a pillar of the AI revolution, it will need to show agility to navigate this increasingly restrictive environment. Even so, the Taiwanese foundry retains a strategic position thanks to already-strong relationships with its customers, as well as cutting-edge processes that give it a sizeable moat over smaller or less structured competitors.