How Morris Chang built TSMC into the world's essential chip foundry

Executive overview

Semiconductors are a scale game: the manufacturer with the most volume drives down cost fastest, locking out rivals permanently. TSMC's edge was doing one thing no integrated chipmaker would — serving customers without competing with them.

Morris Chang founded TSMC in 1987 on two bets: that a pure-play foundry could outmanufacture anyone, and that fabless chip design was coming. Both proved correct. The learning curve — price ahead of cost, grab volume, compound — is the strategic engine behind everything TSMC has done since.

The company that never competes with its customers has become the only manufacturer operating at the leading edge of Moore's Law.

The NVIDIA relationship and the 40nm crisis

• Jensen Huang cold-mailed Chang in 1997 when NVIDIA had 50 employees and faced bankruptcy; Chang visited the next day after his San Jose office ignored the letter.
• NVIDIA's first TSMC chip saved the company and launched what became one of TSMC's five largest customer relationships.
• In 2009, Chang returned as CEO after his successor laid off ~700 workers via performance reviews — a move Chang considered dishonest and operationally damaging in a Moore's Law industry.
• A yield crisis on the 40nm node — TSMC's most important node for that era — had been left unresolved; Chang spent the first four to five weeks back dissecting every aspect.
• He flew to Jensen's home, had pizza and salad with no business discussion until 8pm, then offered a compensation package of over $100M with a 48-hour acceptance deadline.
• Jensen accepted within two days; the relationship continued with many billions of dollars of subsequent business.

The 28nm bet and Apple's arrival

• Chang set R&D spend at a fixed 8% of revenue regardless of recession, ending the annual budget fight and giving the R&D team confidence to think big.
• The R&D team called 28nm the "sweet spot" — analogous to hitting a tennis racket's sweet spot — and Chang tripled capital expenditure to nearly $6B, over board objection.
• The 28nm node coincided with the smartphone era; Apple arrived not through a sales call but through Foxconn chairman Terry Gao bringing Apple COO Jeff Williams to Chang's home for dinner.
• Williams offered TSMC a 40% gross margin guarantee; Chang's existing margin was 45%, on its way to a target of 50%, so he said nothing and moved on.
• Apple wanted 20nm, not 28nm — an unplanned detour that forced TSMC to develop an intermediate node.
• Chang agreed to supply only half of Apple's stated demand, funding the gap through corporate bonds rather than cutting dividends or issuing new equity.

The Intel interlude and Samsung shock

• In February 2011 Williams paused discussions: Intel's CEO had approached Tim Cook to pitch Intel as Apple's chip supplier.
• Chang was not alarmed — he assessed TSMC as equal or better to Intel on technology, better on manufacturing, and far ahead on customer trust.
• He flew to Cupertino anyway; Tim Cook saw him instead of Williams and said Intel "just does not know how to be a foundry."
• TSMC's pure-play model was the decisive factor: Intel's existing chip design business created an inherent conflict of interest that no Apple executive could ignore.
• After completing 20nm for Apple, TSMC fell behind Samsung on 16nm — Samsung had skipped 20nm and got there first, winning Apple's initial 16nm orders.
• Chang emailed Williams immediately; Williams flew to Taipei and confirmed Apple would switch fully to TSMC once TSMC's 16nm was ready, which happened within six months.

The IBM-Qualcomm story and the learning curve

• Through the 1990s, Qualcomm's main foundry was IBM; when Qualcomm quietly shifted volume to TSMC, Chang immediately understood IBM Semiconductor was in structural trouble.
• IBM then asked TSMC to co-develop the 130nm node — a request Chang refused without deliberation, knowing it would transfer capability to IBM and create dependency.
• IBM gave the co-development to UMC, which accepted and later deeply regretted it.
• The learning curve framework: as production volume of any product rises, unit cost falls predictably; the winner is whoever reaches maximum volume first.
• Chang absorbed the theory directly from BCG's Bill Bain over two years at Texas Instruments, where he refined it for semiconductor economics.
• Practical implication: price ahead of where cost will land at full scale, crowd out competitors, aggregate demand, then earn the economics of dominance.
• Every major TSMC capital decision — the 28nm tripling of CapEx, taking half of Apple's order, the 8% R&D floor — flows from internalising this framework.

The fabless origin and TSMC's structural advantage

• Chang saw fabless coming before TSMC existed: at General Instrument, he met Gordon Campbell who dropped his funding ask from $50M to $5M once he decided not to build a fab.
• For TSMC's first several years, most business was overflow manufacturing from integrated device manufacturers — their least critical work at the lowest margins.
• This was deliberate: build volume, build yield expertise, build customer trust, wait for the fabless wave.
• The pure-play rule — TSMC does not design chips, does not make end products, does not compete at any part of the value chain — is what makes customer trust possible at scale.
• Intel's customers in Taiwan uniformly disliked Intel; none of them trusted it. TSMC's customers actively collaborated with it.
• The Hsinchu Science Park ecosystem — TSMC, ASML, Synopsys, Cadence, ARM, MediaTek, Qualcomm offices, and two universities — is physically instantiated and took decades to form; it cannot be airlifted.

Why one foundry dominates

• Fab construction now costs ~$20B per facility; the CapEx threshold rises with each node, structurally eliminating players who cannot sustain it.
• TSMC's net income and CapEx have grown roughly in parallel — they spend almost everything they earn to stay ahead, then earn more because they are ahead.
• R&D is a separate budget on top of CapEx: an 8%-of-revenue floor that compounds as revenue grows.
• Older fabs, fully depreciated, continue generating high-margin revenue for automotive, CMOS sensors, and legacy replacement parts — a cash engine Intel does not have because Intel repurposes fabs for the leading edge.
• Proprietary packaging technologies (CoWoS for AI chips) make dual-sourcing harder, deepening lock-in at each new generation.
• The semiconductor market grew from $26B in 1987 to $527B — TSMC rode and shaped the entire curve.