TSMC has become the single point of failure for the AI industry — and for the first time, it is admitting it cannot build enough.
On the company's April 16 earnings call, CEO C.C. Wei put it plainly: demand is stronger than their own expansion plans. TSMC is now constructing three new 3-nanometer fabs simultaneously, in Taiwan, Arizona, and Japan, a move the company has never made in its history for a node already at full maturity. The first will come online in early 2027; the last not until 2028. Building and ramping a semiconductor fabrication plant takes three to five years. There are no shortcuts.
This is the number that matters from TSMC's eighth consecutive record quarter, and it isn't the 58 percent profit jump to $18.2 billion. It is the gap between what AI developers need and what physics allows TSMC to build.
The manufacturing process that TSMC calls 3nm, essentially a measurement of transistor density that lets more powerful AI chips fit in the same physical space, now accounts for 25 percent of TSMC's revenue, up from 6 percent in the third quarter of 2023. HPC/AI chips, the processors that train and run AI models, now account for 61 percent of revenue. The number used to be smaller.
TSMC is deploying $52–56 billion in capital spending this year alone, up 37 percent from $40.9 billion in 2025. The company expects to remain supply-constrained through the end of the decade.
"The demand is stronger than our own expansion plans," Wei said. It takes two to three years to build a new fab, he added, and another one to two years to ramp it up.
What makes this structurally different from an ordinary capacity crunch is the unprecedented multi-region expansion of a mature node. Historically, TSMC does not add significant capacity to a process technology once its manufacturing recipe is stable and yields are high. According to EE Times, the 3nm node is past that point, yet TSMC is now building it in three locations simultaneously: Tainan, Arizona, and Kumamoto, Japan. The Taiwan facility will start production in the first half of 2027, the US site in late 2027, and Japan in 2028.
For cloud providers, AI model labs, and defense contractors whose roadmaps depend on leading-edge silicon, the practical implication is blunt: plan around 2028 as the earliest point when TSMC's supply constraint might ease. Not 2026. Not 2027. 2028 at the earliest, assuming nothing slips.
TSMC's market capitalization sits at roughly $1.7 trillion, nearly double Samsung's, and the company's shares have gained 35 percent year-to-date. SemiAnalysis noted that TSMC's share of foundry revenue has grown from 54 percent in 2020 to 72 percent in 2026, and the gap continues to widen. Reuters reported that first-quarter profit leapt 58 percent to a record T$572.5 billion ($18.2 billion), with revenue rising to NT$1.134 trillion ($35 billion). CNBC reported capital spending to rise as much as 37 percent to between $52 billion and $56 billion, with high-performance computing accounting for 61 percent of revenue and advanced chips at 7-nanometer or smaller comprising about 74 percent of total wafer revenue.
