TSMC's Capacity Becomes Fierce AI Battleground as Record Quarter Extends

TSMC makes the vast majority of the world's most advanced semiconductors, including the graphics processors and central processors from Nvidia, AMD and Broadcom that are needed to run AI workloads. A growing number of technology companies also rely on TSMC to make custom in-house AI accelerators. Last year, TSMC said it had capacity to make more than 17 million 12-inch equivalent wafers, but customers want more.

"The competition is extremely fierce," Jordan Klein, an analyst at Mizuho Securities, told CNBC. "It's a game of who can get the most leverage and supply." TSMC North American CEO Sajiv Dalal told CNBC he gets at least 10 calls a day about wafer capacity.

Klein said that if compute demand continues to grow at the current rate over the next 12 to 18 months, capacity constraints at TSMC could threaten some of the industry's large infrastructure investments and the "aggressive revenue growth targets" of AI companies. U.S. hyperscalers are projecting well over $700 billion in capital expenditures this year, and analysts expect that figure to jump in 2027. A global shortage of memory chips also threatens AI ambitions. The logic chips TSMC manufactures are more complex and expensive to produce.

Because of that complexity, GPU and CPU leaders such as Nvidia and AMD work out lengthy contracts and bespoke deals with TSMC behind closed doors years in advance. Nvidia CEO Jensen Huang and AMD CEO Lisa Su have both visited Taiwan multiple times this year.

Sid Sheth, CEO of AI chip startup D-Matrix, has built a relationship with TSMC for more than 20 years, and his company relies on TSMC to manufacture its processors. Sheth said that for TSMC customers, many roads lead to Kevin Zhang, a senior vice president of the Taiwanese company and a central figure in determining which chipmakers get access to its factory lines. Zhang is "currently probably the most popular person on the planet," Sheth said in an interview.

Sheth said Jerry Qubain, D-Matrix's vice president of manufacturing, has been visiting TSMC since 1992 and goes to Taiwan three to four times a year. "This was even before Nvidia was founded," Sheth said. "That's a good example of a person we brought on board because of the relationships he has at TSMC." Qubain joined D-Matrix in 2024. TSMC declined to participate in this story.

TSMC had 534 customers in 2025. While it does not name or rank them, its top 10 customers accounted for 78% of net revenue, with its top two bringing in 19% and 17%, respectively. That leaves startups competing for capacity against customers with far bigger orders. Sheth said TSMC is "selective," but has an incentive to back promising newcomers. "They were very supportive of Nvidia and Broadcom when they were very young companies, and now these are their biggest customers today," Sheth said. "They want to see a repeat of that."

TSMC is trying to bring much more capacity online, with new fabs in Taiwan, Japan and Arizona. It is making advanced chips in the United States for the first time. In July, TSMC announced it was adding $100 billion and "several more" fabs to the six planned in Arizona, bringing its total commitment to $265 billion. TSMC is also reportedly considering an entirely new fab location in Texas.

All of this will take years to come online. In the meantime, analysts say TSMC is adding tools such as ASML's extreme ultraviolet lithography machines to existing factories to ramp up capacity. ASML beat profit expectations in its earnings report Wednesday and said customers are interested in placing orders through 2030.

TSMC is not the only option. Capacity constraints create an opening for the other two leading advanced chip manufacturers, Samsung and Intel, to win customers. "People want a second source to be able to twist TSMC's arm and keep them honest on pricing," Dan Hutcheson, an analyst at TechInsights, told CNBC. That second source might be Intel, the only U.S.-based advanced logic manufacturer. Although its foundry business has struggled with delays and setbacks for years, Hutcheson says all the big chipmakers are now testing there.

Trade association SEMI expects 102 new chip fabs will come online globally by 2029, including 15 in the United States and 78 in Asia. China is the primary growth engine, expected to nearly triple capacity by 2030.

Editor's Summary TSMC's record quarter and 58% fourth-quarter revenue growth forecast underscore how AI demand is straining its advanced chip capacity. Customers including Nvidia, AMD, Broadcom and major cloud and AI firms are competing for limited wafer supply, while TSMC expands in Taiwan, Japan and Arizona and rivals Samsung and Intel seek an opening. The constraints could affect AI infrastructure spending and revenue targets if compute demand keeps growing at its current pace.