Jensen Huang Turned Down the TSMC CEO Job Twice: I Already Have a Job
In 2013 Morris Chang seriously considered handing the TSMC he founded 26 years earlier to Jensen Huang. Huang answered with five words — ‘I already have a job’ — and said the same thing both times. TSMC's market cap then was 10 times Nvidia's.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
No money to tape out, so build the chip in a computer first
Nvidia's first two chips both failed, and by the end of 1996 the company had been cut down to just over thirty people. The normal process in the chip industry is tape out, test, fix the design, go again — two or three rounds is common — but Huang only had enough money on his books to tape out once. He took a third of his cash, about one million dollars, and bought a machine that could simulate an entire chip inside a computer, to find as many errors as possible before going into a real fab, and then decided the first version would go straight to mass production. A manager asked him how he knew he would get it right in one pass. He answered: I know it has to be right, because if it isn't, the company is gone. In August 1997 the third chip shipped, and in its first four months it moved more than 1 million units.
Betting the whole lifeline on one fab with no contract in hand
From 1998 Huang put almost every Nvidia chip on TSMC alone. At the time most fabless chip companies worked with two or three fabs at once, playing them off on price while keeping a fallback, and Nvidia took that path thoroughly enough. More important, Huang made this bet without a long-term contract in hand — around 1997 the two sides discussed a formal manufacturing arrangement, but day-to-day orders were still placed one at a time, and TSMC never promised to reserve any particular capacity long term. Only two things held up the bet: Morris Chang's verbal commitment, and Huang's line that sounded almost like bragging.
When TSMC makes a mistake, it is Nvidia that goes bankrupt
In 1998 a chemical step at the end of a TSMC production line went wrong, causing large-scale breaks across a whole batch of wafers, and more than half the chips in that Nvidia lot were scrapped. Because it was hard to tell by ordinary sampling which ones were usable, Nvidia threw up a temporary test line and hired a large crew to sort them one chip at a time, and TSMC sent two production planners from Hsinchu to California for a full month. That quarter Nvidia's revenue fell from more than 28 million dollars to less than half, the net loss came close to 10 million dollars, and in the end it had to sell part of its equity to several downstream companies for cash. This is the fundamental risk of the design-only model: the fab belongs to someone else, the mistake was someone else's, but the money lost, the undelivered goods and the walk to the edge of bankruptcy can only be yours.
A customer diverting a little business has to be written into risk factors
In 2002 TSMC's most advanced process generation could not get its yields up, Nvidia's new flagship shipped half a year late, and once it launched review sites nicknamed it the hair dryer for its power draw and heat; rival ATI used the opening to overtake it with new products. In March 2003 Nvidia announced an alliance with IBM, handing its next-generation flagship to IBM's new fab in New York state, in a contract estimated at more than 100 million dollars. The same day the two companies jointly issued a press release reaffirming the partnership, but three months later TSMC filed its annual report with the SEC and for the first time put the matter into risk factors, writing that it could not guarantee it would not lose Nvidia's business in the future. The previous year Nvidia had still contributed about a fifth of TSMC's revenue; that year its share fell from two tenths to one and a half tenths. In the end IBM could not get its yields up either, and in the summer of 2004 Nvidia's consumer chips came back to TSMC, with the flagship following in 2005.
After the pizza dinner, a 48-hour ultimatum
After the 40nm yield failure Nvidia demanded compensation, and then-CEO Rick Tsai insisted TSMC was not at fault and would not pay a cent; Huang said that if they could not agree, they would go to arbitration. After Morris Chang returned in 2009 he spent half his time on this matter, talking for dozens of hours with everyone involved, then shut himself in with his staff to study it for a full three weeks. In early July he emailed Huang, saying only that he would come to Silicon Valley the following week and would be at his house at 6:30 p.m. on July 15 for salad and pizza, not a word about business. At 8 o'clock the two went into the study and closed the door, and Chang gave his settlement proposal: more than 100 million dollars in compensation, TSMC's only and final offer, not open to haggling, valid for just 48 hours, otherwise arbitration was the only route. Within two days Huang accepted. The figure was roughly equal to 3% to 4% of TSMC's net profit that year, and about 3% of Nvidia's full-year revenue.
Morris Chang wanted Huang to succeed him, and was refused twice
In 2013, choosing a successor for the second time, Morris Chang seriously considered Huang, who was outside the company. In his memoir he listed three reasons: character, career, and expertise in semiconductors, on top of 16 or 17 years of personal friendship between the two. Chang asked Huang to his face whether he was interested in being TSMC's CEO, spent about 10 minutes laying out the expectations and the pay, and made a point of stressing that the compensation would certainly be more than Huang was earning now. Huang did not interrupt for the full 10 minutes, did not ask a single question, and at the end said only five words: I already have a job. A few weeks later Chang made another phone call, with no change in wording and no sweeter terms, and Huang's answer did not change by a single word. At the time TSMC's market cap was 90 billion dollars, a full 10 times Nvidia's.
The AI bottleneck moved from inside the chip to the packaging
Training a model is essentially doing arithmetic on enormous quantities of numbers over and over: the processor does the computing, the numbers sit in memory, and in the past the two were packaged separately and soldered onto the same circuit board, with data moving over the copper traces in the board. The traces are long and limited in number, and against AI-scale computation they ran out. TSMC's answer was to use wafer-fab processes to build a large slab of silicon as a base, place the processor and high-speed memory side by side on it, and put them into a single package — a technology called COWOS. In 2011 the two companies together produced the first test chip using this packaging, and in 2016 the first mass-production chip shipped, with few people paying attention at the time. In early 2023 TSMC could supply only about 8,000 COWOS wafers a month, while Microsoft, Google, Meta and a batch of model companies were all waiting on Nvidia cards, and the orders all jammed at the same packaging step. Over three years TSMC expanded capacity roughly tenfold, with about 60% of it going to Nvidia.
A design flaw is Nvidia's fault, not TSMC's
In the summer of 2024 a new generation of Nvidia AI chips was found to have a design flaw before mass production, yields would not come up, and shipments slipped by a quarter, and rumors quickly spread in the market that the two companies were blaming each other. On October 23, asked directly by reporters in Copenhagen, Huang said flatly that these were all fake news, and stated clearly that the design problem was 100% Nvidia's fault, and that TSMC's job was to help them solve the yield problem and remake the chip at astonishing speed. What has kept this partnership going so long was never the absence of mistakes, but that after a problem arises, both sides rarely push the blame onto the other.
In their own words · checked verbatim
I know it has to be right, because if it isn't, the company is gone.
我知道它一定是对的，因为它如果不对，那公司就没了。
Jensen Huang14:11
Please don't do this to us, because in the future we will be your biggest customer.
请你不要这样对我们，因为将来我们会是你们最大的客户。
Jensen Huang20:16
TSMC has learned to dance with 400 partners, while Intel has spent all these years used to reading the dance.
台积电已经学会了跟400个舞伴一起跳舞，而英特尔这么多年一直习惯读舞。
Morris Chang1:20:22
Figures
| Nvidia market cap | 5 trillion dollars | 1:00 |
| Morris Chang's age when he founded TSMC | 55 | 8:08 |
| Nvidia's third chip shipments in its first four months | More than 1 million units | 14:11 |
| Nvidia's share of TSMC revenue (2001) | About 17% | 26:22 |
| 40nm settlement compensation | More than 100 million dollars, about 3% to 4% of TSMC's net profit that year | 48:39 |
| TSMC vs Nvidia market cap in 2013 | TSMC 90 billion dollars, 10 times Nvidia | 1:27:26 |
| TSMC COWOS monthly capacity in early 2023 | About 8,000 wafers | 1:19:21 |
| Nvidia's prepayment to TSMC at the end of 2021 | Nearly 7 billion dollars | 1:13:18 |
| Largest customer's share of TSMC revenue in its 2025 annual report | 19% | 1:25:25 |
Glossary
- COWOS (Chip-on-Wafer-on-Substrate)
- TSMC's advanced packaging technology, placing the processor and high-speed memory on the same silicon interposer to solve the AI compute data-transfer bottleneck.
- Tape-out
- Running a chip design through a complete pass on a real production line to produce the first batch of wafers, in order to verify that the design can be built.
- Yield
- The share of chips cut from a wafer that are actually good, which directly determines the cost of each working chip.
- Fabless chip company
- A company that only designs chips and does not build its own fabs, outsourcing all manufacturing to a foundry.
- Silicon photonics
- Technology that introduces optical signals to move data between chips, replacing copper traces that are slow and hot.
How to listen
Founders and investors watching the semiconductor supply chain, AI compute bottlenecks and long-term founder relationships; anyone who wants to know why TSMC is hard to replace and why Nvidia cannot do without it.
The dinner-party opening and biographical setup from 0:00-3:00 can be fast-forwarded; the core mechanics start in the middle.