Morris Chang was born in China in 1931 and moved to the United States at 18, studying at Harvard before transferring to MIT, where he earned bachelor’s and master’s degrees in mechanical engineering. He later completed a PhD in electrical engineering at Stanford in 1964. [1][2]

After three years at Sylvania Semiconductor, Chang spent about 25 years at Texas Instruments, rising from supervising a production line to running TI’s integrated circuits division and later its entire semiconductor business as a group vice president. He then served briefly as president and chief operating officer of General Instrument. [1][2]

In 1985 Chang moved to Taiwan to become president of the Industrial Technology Research Institute (ITRI), a government-backed research organisation, a role he held, later as chairman, until 1994. K.T. Li, a Taiwanese minister without portfolio, was instrumental in recruiting him, and Taiwan’s premier asked him to turn research into economic value for the island’s industry. [1]

Chang founded Taiwan Semiconductor Manufacturing Company in 1987. His reasoning, as he later described it, started from Taiwan’s limits: the island had little strength in chip design, marketing or intellectual property, but it could manufacture well. A company that only manufactured, and did so for others, would play to that strength and avoid the weaknesses. [1][3]

Raising money was not easy. Chang recalled that Philips was the one significant industrial investor willing to commit when others hesitated, and the project had strong political sponsorship in Taiwan’s government. [1]

TSMC’s early customers were large integrated device manufacturers, which sent it work they did not want to do in their own fabs, and small Taiwanese firms. Growth accelerated from around 1991 and 1992 as the fabless industry expanded. In a 2007 interview Chang said TSMC had earned about 110 percent of the pure-play foundry industry’s total profits over its history, meaning its profits exceeded the combined results of all its competitors. [1]

When TSMC started, the chip industry was built around integrated device manufacturers that designed and made their own chips, and some of TSMC’s first orders were overflow work these companies chose not to run in their own fabs. TSMC describes itself as having pioneered the pure-play foundry model, focusing exclusively on manufacturing its customers’ products rather than selling chips of its own design. [3][1]

That focus is the heart of the model. Chang argued it would let independent designers start chip companies without the huge capital cost of a fabrication plant. The customer designs the chip and sells the product; the foundry supplies the manufacturing, spreading the cost of each new factory across many customers. By 2025 TSMC served 534 customers and made 12,682 different products using 305 process technologies. [1][3]

Over time even some integrated manufacturers exited manufacturing. AMD, which once ran its own fabs, co-founded GlobalFoundries with Mubadala in 2009 as a separate manufacturing company. Today AMD’s Instinct MI300X accelerator is built on TSMC 5 nm and 6 nm processes. [4][5]

The companies that dominate AI hardware design are TSMC customers. NVIDIA’s H100 GPU uses a TSMC 4N process customised for NVIDIA, and its Blackwell generation uses a custom TSMC 4NP process, joining two dies with a 10 TB/s link. AMD’s MI300X, with 153 billion transistors and 192 GB of HBM3, combines TSMC 5 nm and 6 nm chiplets. [6][7][5]

Apple is another cornerstone customer. Apple says it is the first and largest customer of TSMC’s Arizona fab, which it expected to produce tens of millions of Apple chips in 2025, with packaging and testing to be done nearby by Amkor. [8]

AI has reshaped TSMC’s business. Its revenue grew 35.9 percent in 2025 to $122.4 billion, and high-performance computing, which includes AI accelerators, rose to 58 percent of net revenue from 51 percent a year earlier. In the second quarter of 2026 it reported revenue of $40.2 billion and a gross margin of 67.7 percent, with 77 percent of wafer revenue coming from 7 nm and more advanced processes. [9][10]

Packaging has become as important as the wafer. TSMC’s CoWoS technology mounts GPU dies and HBM stacks on a large interposer, and TSMC says demand rose significantly after generative AI arrived in late 2022. Its CoWoS-L variant, at 3.5 times the size of a lithography exposure field, entered volume production in 2024 to handle the largest AI packages. [11]

TSMC is also at the front of process technology. Its 2 nm N2 process, the company’s first to use nanosheet transistors, began volume production in the fourth quarter of 2025, and in April 2026 TSMC set out a roadmap through A14 in 2028 and A13 and A12 in 2029, with A12 aimed at data-centre products. [12][13]

Most of TSMC’s manufacturing remains in Taiwan, where it runs six 12-inch GIGAFAB facilities alongside older 8-inch and 6-inch fabs. It also operates fabs in Nanjing, Arizona, Japan (through the majority-owned JASM) and Washington state, and began building a specialty fab in Dresden in 2024. Its managed fabs had capacity of more than 17 million 12-inch-equivalent wafers in 2025. [3]

Its US presence has expanded quickly. In March 2025 TSMC raised its planned US investment to $165 billion. In July 2026 it announced another $100 billion for Arizona, bringing the planned total there to $265 billion for ten fabs, two advanced packaging plants and an R&D centre. The first Arizona fab has made N4 chips since late 2024, a second fab using 3 nm is expected to start volume production in 2027, and about 30 percent of TSMC’s 2 nm and more advanced capacity is expected to be in Arizona eventually. [14][15]

The foundry model is the reason a handful of design companies can compete to build AI accelerators at all, but it also concentrates the most advanced manufacturing and packaging in one company. That makes TSMC’s capacity plans, pricing and geography a leading indicator for GPU supply: when CoWoS or leading-node capacity is tight, new accelerators stay scarce and rental rates stay high. On Kovara you can see how that plays out in prices, comparing GPU cloud rates across providers and regions, or ask Kova how a new generation’s supply is affecting cost per hour.