Taiwan Semiconductor Manufacturing said on October 15 that it will build another 3-nanometer factory in Tainan, with production scheduled to begin in the first half of 2028. The site will stand alongside the company’s 3-nanometer capacity in Arizona and in Kumamoto, Japan, and will add to output for the most advanced processors available.
The announcement coincided with a picture of who is waiting for that output. Nvidia, AMD and Broadcom are the largest buyers of leading-edge wafers for graphics and central processors, and Apple, Google, Amazon, Microsoft and Meta have designed their own accelerators at the same foundry. All of them are queued for the same lines.
Sajiv Dalal, the company’s chief executive for North America, said he receives at least ten calls a day asking about capacity. Analysts described the situation as a contest over who can command the most supply and the best terms, which is a polite way of describing buyers with more demand than the industry can currently satisfy.
The financial picture explains why the company can commit capital years ahead. TSMC reported third-quarter net income of NT$763.1 billion, about $24 billion, up 68.7 percent from a year earlier and its sixth consecutive record quarter, beating the NT$746.3 billion analysts expected. Revenue rose 51 percent to NT$1.49 trillion, and operating margin reached 58.7 percent, above the top of its own forecast range.
Guidance for the fourth quarter was set at $52.3 billion to $54.3 billion, a midpoint that implies a 14 percent increase from the prior quarter and 58 percent from a year earlier. Three-nanometer production accounted for 33 percent of wafer revenue in the third quarter, which is the number that makes the Tainan decision legible: the node that buyers are fighting over is also the node paying for the expansion.
The company is also evaluating investment in Texas and has said it is willing to work with the fab that Elon Musk has proposed, while keeping operating control of its own plants. Control is the condition that matters. A foundry earns its margin by deciding how a line runs, and a partnership that cedes those decisions would trade the company’s main advantage for someone else’s capital.
Adding capacity in three countries costs more per wafer than building in Taiwan, and the company has said customers will pay a share of the difference. Silicon made in Arizona or Kumamoto carries costs that Tainan does not, and the price list for the Taiwan base book does not capture what a customer pays for wafers produced abroad.
The shape of demand has changed since the last shortage. Accelerator vendors now consume a larger share of leading-edge output than phone and personal computer makers did in earlier cycles, and they order fewer wafers at far higher value. A customer building accelerators will pay more for allocation than a customer assembling handsets, which changes how the company negotiates.
For equipment makers and materials suppliers, the expansion is a commitment that can be read as a forecast. A fab that breaks ground now places orders for lithography, deposition and inspection tools long before it produces a wafer, and those orders are how the supply chain decides which technologies to develop next. The company’s own capital spending has been the largest single signal in the industry for two decades, and each addition to the list of sites is a statement about what its customers say they will need four and five years from now.
Neither the schedule nor the demand is fixed. Fab construction has slipped before, and a line that starts production in 2028 depends on customers whose product plans will change in the meantime. If demand for accelerators cools, an industry that built capacity for a shortage ends up with idle equipment, and the present scarcity becomes the next glut. The company has said it builds where it holds commitments, which moves part of that risk onto the buyers who are queuing today.
Memory is the constraint that sits outside the foundry’s control. Prices for DRAM and flash have risen faster than wafer costs, and a system builder now faces increases across several components at once. Chips that cannot be paired with enough memory do not ship, which limits how much extra wafer capacity translates into finished systems.
Analysts said the company’s ability to raise prices and fill new lines reflects the absence of an alternative at equivalent scale. Samsung Foundry and Intel both need to fill their own advanced capacity, and neither has matched yields on the most advanced nodes. That leaves buyers with a choice between waiting and paying.
For the buyers, the practical result is that accelerators and custom chips will remain allocated rather than freely available, and that the queue forms years in advance. A company that wants volume in 2028 is negotiating now, which is why the calls reach Mr. Dalal at a rate of ten a day.
The Tainan fab will not answer those calls quickly. Construction takes years, and equipment for a new line is committed long before the building is finished. The company is buying time with capital, and its customers are spending time in a queue.
