How TSMC built an unassailable monopoly in semiconductor manufacturing

Executive overview

Most technology companies talk about chips. TSMC makes them — all of them. The company manufactures nearly every leading-edge chip in every iPhone, Nvidia GPU, AMD processor, and custom silicon from Apple, Amazon, and Microsoft.

Morris Chang founded TSMC at 56 after being passed over for CEO at Texas Instruments. His insight: create a pure-play foundry that manufactures chips for fabless designers, a market that barely existed in 1987. That bet unlocked an industry-defining flywheel: more customers fund more capex, which funds better process technology, which attracts more customers.

No amount of money can replicate 40 years of process power — TSMC is the hardest thing to build in the world, and nobody else is close.

Morris Chang's rise through semiconductors

• Born in China in 1931, fled three wars before age 18; reached Harvard at 18, transferred to MIT to study mechanical engineering
• Failed MIT PhD qualifying exams twice; joined Sylvania's semiconductor division for $1 more per month than Ford offered
• Self-taught electrical engineering by studying Shockley's textbook and interrogating a senior engineer at the hotel bar each night
• Joined Texas Instruments in 1958 — the same year the integrated circuit was invented — and quadrupled yields on a failing IBM production line within four months
• Invented learning-curve pricing at TI with BCG: automatically lower prices each quarter to drive volume, accelerate yield improvement, and gain market share; made TI the world's largest IC company
• Promoted to VP of semiconductors, then sidelined to consumer products he couldn't turn around; resigned in 1983 after 30 years

The founding of TSMC

• After TI, spent a year at General Instrument — a financial-engineering firm, not a technology one — before resigning again
• Recruited by Taiwan's government minister KT Lee to run ITRI, the country's Bell Labs equivalent; treated it as semi-retirement after all other options closed
• Given three days to write a business plan; concluded Taiwan's only viable strength was semiconductor manufacturing, not design, IP, or marketing
• Created the pure-play foundry model: manufacture chips for fabless customers, taking no design role — the opposite of every incumbent's approach
• Jerry Sanders of AMD had recently declared "real men have fabs"; no one thought fabless chip companies would exist at scale
• Raised $220 million at a zero-dollar pre-money valuation; Morris received no founder equity and bought his stake later on the open market
• Taiwanese government owned 50%; Philips (the only willing Western partner) owned 28%; remainder was strong-armed from Taiwanese business leaders
• TSMC and ARM were both founded in 1987

Building the foundry flywheel

• Early revenue came from IDM "dregs": excess capacity other chip companies didn't want, which evaporated as soon as they needed it back
• Morris had seen IC designers at TI unable to leave because they couldn't raise enough capital to build their own fabs; TSMC removed that barrier
• Fabless companies followed: Qualcomm, Broadcom, Marvell, and Nvidia (which raised only $20 million and never opened a fab) all launched on TSMC
• Flywheel: fabless growth → TSMC revenue → 50% gross margins → capex → better process technology → more fabless customers
• At the leading edge, competitors shrank from 22 companies to 14 to 6 to 2 (TSMC and Samsung); Intel has fallen behind
• Morris retired in 2005 at 74, returned in 2009 at 78 to capture the mobile and cloud opportunity; personally closed the Apple deal, committing $9 billion and 6,000 employees to build a dedicated fab in 11 months

The semiconductor value chain

• Five layers: IP (ARM, architecture licensing) → EDA software (Synopsys, Cadence) → fabless chip design (Apple, Nvidia, Qualcomm) → fab equipment manufacturing (ASML) → foundry (TSMC)
• ASML, the sole maker of extreme ultraviolet lithography machines, was co-founded with Philips in 1984 — the same Philips that was TSMC's first outside investor
• EUV machines cost $200 million each, require four 747s to ship, and need a dedicated ASML crew to assemble and operate; ASML makes roughly 50 per year
• EUV creates chips by firing a laser at molten tin 50,000 times per second, generating plasma that emits extreme ultraviolet light; the required precision exceeds Apollo mission calculations
• No known mirrors reflect EUV light; ASML had to invent a new mirror type and contract a single German company for the laser
• TSMC has ordered out ASML's production years in advance; the Netherlands has blocked EUV machine exports to China

Why TSMC cannot be replicated

• Process power: 40 years of accumulated manufacturing know-how, relationships with ASML, and engineering depth; cannot be bought or fast-tracked
• Scale economies: $17 billion operating profit on $48 billion revenue (2020); $100 billion capex planned over three years; 50%+ gross margins on what others treated as a commodity
• Switching costs: porting a chip design to another foundry takes years of deep process re-integration
• Counter-positioning: IDMs could not adopt the pure-play foundry model without destroying their integrated margins
• China's SMIC and any government spending program cannot close the gap — the knowledge required to run EUV at yield takes decades to accumulate
• TSMC recently began raising prices for the first time since Morris introduced learning-curve pricing, a direct signal of monopoly pricing power

Geopolitical risk and the Taiwan question

• Virtually all leading-edge manufacturing sits on an island claimed by China; this is the primary long-term risk
• A Chinese annexation of Taiwan would cut off leading-edge chip supply to every major technology company and defence program in the world
• TSMC is building a plant in Arizona and a fab in Japan, primarily because governments are subsidizing cost; Morris has said the leading edge belongs in Taiwan's ecosystem
• It is unclear whether TSMC's process power could survive evacuation; the Toyota Production System failed to transfer cleanly under the far lower pressure of a joint venture with GM
• Intel saw EUV first, invested $4 billion in it by 2012, then determined it wouldn't work — and to this day has not shipped a chip manufactured using EUV, a textbook example of the innovator's dilemma