Jensen Huang in Taiwan With Lex Fridman: Pick Partners, Not Suppliers

NVIDIA CEO Jensen Huang shows Lex Fridman around Taiwan, explaining how he chose TSMC, why he opposes broad export controls, and why he wants his suppliers to get rich. Visiting Taiwan w/ Jensen Huang is a 90-minute travel video rather than a sit-down podcast: Lex follows Jensen through the Computex show floor, an NVIDIA staff party, two night markets and a Taiwanese breakfast, and the substance arrives in conversations between dishes. They cover the new 100 percent liquid-cooled rack, Taiwan’s semiconductor history, export controls and the “flywheel,” and the cold call and letter that led to NVIDIA’s 30-year partnership with Morris Chang’s TSMC.

TLDW

Jensen Huang argues that Taiwan sits at the center of the world’s computing supply chain because of engineering culture and a generosity he saw in its night markets as a child. On export controls, he says denying technology works for bullets and refined rare earths but that the better long-term strategy is to maximize the reach of America’s technology and economic flywheel, with only very selective restrictions. Over breakfast he tells how an unanswered phone line, a long letter and an evening call from Morris Chang started NVIDIA’s relationship with TSMC. His lesson from 30 years of it is that you choose who to work with rather than what to buy, keep agreements loose, and always leave something on the table for partners.

Thoughts

Huang’s best business lesson is that for anything as hard to build as a chip, you are choosing a partner, not making a purchase. He puts it plainly over breakfast: “You really are choosing not what to buy, but who to work with.” Huang’s reasoning is specific to chipmaking. Building a chip with a foundry is a joint project full of unknowns, so no contract can spell out every decision in advance, and he compares an exhaustive one to an ironclad prenup: the more specific it gets, the more brittle the relationship becomes. NVIDIA and TSMC instead run on what he calls “rough justice,” where sometimes one side wins and sometimes the other, and it evens out over 30 years. That only works with a partner whose values you trust, which is why he says he picked TSMC over UMC across the street. Morris Chang kept telling him what TSMC would not do and spent the meeting on governance, integrity and long-term thinking. Most procurement processes are designed to remove trust from the equation. Huang’s claim is that in the hardest kind of manufacturing, trust is the cheaper and more durable contract.

The companion idea is to always leave something on the table. Huang warns that squeezing suppliers by forcing everyone to bid until “nobody makes any money but you” is remembered on the island forever. He then offers a striking data point: last year his supply chain’s stocks rose about 300 percent while NVIDIA’s stayed flat, and he frames that as proof the platform is big enough for everyone to rise with it. There is self-interest here, since a healthy, well-funded supply chain is exactly what a company planning to sell “a few trillion dollars of computing” needs, and every one of those partners has to invest hundreds of billions to keep up. But that is the point. Generosity toward suppliers is not charity when your growth depends on their capacity. It is how you make sure they build it.

On export controls, Huang gives a more careful answer than the usual “let us sell everywhere” line. He concedes that denial works for physical, finite things like bullets, bombs and refined rare earths, where the other side cannot simply make its own. For technology platforms, he argues, the right thing to optimize is a flywheel: installed base drives usage, usage drives wealth, wealth funds R&D, and R&D keeps the lead. America’s advantage, in his view, is that an American stack built on English and American standards deploys almost everywhere, while China’s does not, and “technology without customers is useless.” The weakness is that NVIDIA is the most interested party in this debate, and Lex’s own reporting from earlier in the trip cuts the other way: Chinese AI labs such as MiniMax say they are compute starved, and Huang agrees that scarcity is pushing them toward remarkably efficient architectures. Restrictions may slow rivals while also sharpening them. Huang’s answer is a balance, “extremely selective” controls plus maximum platform reach, and he never says where exactly the line sits.

A detail from the foundry story deserves more attention than it gets. In the mid-1990s, Huang says, foundry capacity was so scarce that using someone else’s fab often meant licensing your technology to them, and SGS-Thomson would not take NVIDIA’s business at all. He notes in passing that this “kind of reminds me of” some early deals with AI companies, where access to capacity came with something licensed away. It is a quiet admission that the AI industry is living through the same bargain NVIDIA once faced, with NVIDIA now holding the scarce resource. It also explains why TSMC’s “no strings attached” offer meant so much to a startup that, in his words, might not have had to die after all.

The sharpest technical point comes on the Computex floor. Huang says hyperscale x86 CPUs were designed to be rented, so the economics rewarded as many cores as possible and the metric was dollars per core. AI does not rent cores; it generates and sells tokens, so a CPU’s job becomes moving through the processing loop as fast as possible. That is the case for NVIDIA’s Vera CPU, which Lex cites at 1.8 times over x86, and it is a useful frame beyond chips. When the unit of sale changes from capacity to output, the hardware that wins changes with it.

Key Takeaways

Taiwan produces more than 90 percent of the world’s most advanced semiconductors, and Lex frames it as the place every computer, phone and AI data center depends on.
NVIDIA’s newest rack at Computex is 100 percent liquid cooled, up from 80 percent on Grace Blackwell, and assembly time fell from two hours to five minutes.
Huang says x86 server CPUs were built for renting cores, while AI sells tokens, so CPUs should be designed to run the generation loop as fast as possible.
He is nervous before every keynote, uses no teleprompter, and says the whole island must invest hundreds of billions to build a few trillion dollars of computing.
He sees the AI boom as America’s best chance to rebuild its grid and manufacturing, and calls Taiwan, which is sending engineers to US fabs, a strategic partner in that effort.
He traces Taiwan’s chip industry to government minister K. T. Li, who drew Philips and Morris Chang to the island, and to IBM’s open PC chipset specifications, which Taiwanese engineers designed to and TSMC and UMC manufactured.
NVIDIA has about 1,000 people in Taiwan, mostly hardware, supply chain and operations engineers, and plans a new building for 4,000.
Huang credits Taiwan’s place at the center of the supply chain to generosity and hard work, the same traits he saw in night market vendors as a child.
On export controls he argues for a technology and economic flywheel over broad denial, with “extremely selective” limits on specific technologies.
He calls English and American standards a global flywheel advantage that Chinese technology stacks do not have.
In the 1990s scarce foundry capacity often required licensing your IP to the fab, which he compares to some early AI capacity deals.
TSMC never returned his calls, so he wrote Morris Chang a long letter, and Chang eventually phoned him one evening and invited him to Taiwan.
He chose TSMC over UMC because Chang focused on values, governance and long-term partnership and promised to “jump through hoops,” which Huang says TSMC has done for 30 years.
His partnership rules are to keep agreements loose, accept “rough justice,” make sure partners know you want them to win, and always leave something on the table.

Chapters

0:47 Why Taiwan Matters

Lex introduces his first trip to Taiwan, to attend Computex and spend several days with Jensen Huang. He describes the island as two-thirds mountains yet responsible for more than 90 percent of the most advanced chips, comparing it to ancient wonders whose construction is still a mystery. He credits culture and an education system that prizes engineering, and says he wants the US and China to compete while remaining friends.

4:27 Computex: Liquid Cooling, Vera and Keynote Nerves

On the show floor, mobbed by fans, Huang shows off a rack with no cables or hoses that is fully liquid cooled and takes five minutes to assemble instead of two hours. He explains why the Vera CPU is built for generating tokens rather than renting cores, visits SK Hynix and admires old GeForce and Titan cards on display. Afterward he admits he is nervous before every keynote, does not rehearse because everything comes together at the last moment, and tried to distill why the whole Taiwanese ecosystem must invest to build out AI.

14:57 Reindustrializing America and Taiwan’s Tech History

Huang says the AI boom is the best chance to bring manufacturing back to the US, and that Taiwan is training people and sending engineers to help build American fabs. Asked why so many hardware companies come from Taiwan, he credits K. T. Li’s push for science and manufacturing, Philips’s early role in what became TSMC and ASML, and IBM’s open chipset specifications, which let Taiwan supply the motherboards that a young Michael Dell assembled into PCs. Lex shares a line from a Taiwanese woman who learned to spell CoWoS before Gucci.

19:59 Raohe Night Market

The pair work through Raohe night market with a crowd in tow: buns fresh from the oven, sweet potato balls, a spicy noodle Lex calls possibly the best food he has eaten, shaved ice and the famous oyster pancake. A vendor refuses Huang’s money. Between bites Huang says NVIDIA has about 1,000 people in Taiwan, mostly hardware, supply chain and fab engineers, and plans a building for 4,000.

29:44 A Childhood Scar and Compute-Starved Labs

Huang recalls going to the night market with his parents because they could not afford restaurants, and getting cut across the face by a cook’s knife at almost five. He says he still loves it for the vendors who share what little profit they have. Lex describes visiting MiniMax in Shanghai and Unitree in Hangzhou, where every AI company told him it was compute starved, and Huang agrees that the constraint is forcing very efficient architectures.

33:46 The NVIDIA Party

At an NVIDIA employee gathering with Kavalan whisky and a banquet table, Huang hands out food to the press. He describes turning a corner that day to find several thousand people lined up and not knowing what to do. He argues that Taiwan’s capacity for generosity and helping others, combined with very hard work, is why the island is the world’s supply chain. On stage he jokes that he started dating his wife in college because she had a 25-cent weekly allowance.

43:11 Humility, Export Controls and the Flywheel

Over braised pig’s feet, Huang says fame does not weigh on him because he has always wanted others to succeed, and that goodwill is not zero-sum. Asked to steelman export controls, he says denial works for finite physical goods but that long-term security comes from prosperity, and the right target is a flywheel of technology and economic leadership. He argues America’s global reach through English and American standards is a major advantage, and that every country has a sovereign right to control its technology but the US should maximize its platform’s reach.

55:19 Ningxia Night Market in the Rain

Despite rain, the group heads to Ningxia night market, where TSMC founder Morris Chang is introduced. Huang explains why eggs appear in everything, the QQ bouncy texture prized in noodles and sesame balls, and why it is impossible to finish a meal when hosts keep bringing food: in some cultures their generosity matters more than your appreciation for the food. They try egg crepes, peanut ice cream rolls with cilantro, green mangoes, grass jelly, durian and plums.

70:16 Breakfast and the Search for a Foundry

Over soy milk, fried dough, rice rolls with pork floss and soup dumplings, Huang tells the founding story. Around 1996 the PC boom created a fabless chip industry and a scramble for foundry space, mostly in Japan and Europe because Intel, AMD and TI built their own products. SGS-Thomson turned NVIDIA away, and fabs often demanded IP licenses for access. TSMC, then a roughly $300 million company, never answered his calls, so he wrote Morris Chang a long letter until Chang called one evening while the office played PlayStation.

83:13 Choosing TSMC and Rough Justice

Huang visited UMC and TSMC, across the street from each other, and chose TSMC because Chang talked only about values and promised to jump through hoops for NVIDIA. He argues that with a foundry you choose who to work with, not what to buy, that tight contracts make relationships brittle, and that “rough justice” evens out over decades. He closes by saying partners must know you want them to win, and that he was glad to see his supply chain’s stocks triple while NVIDIA’s stayed flat.

Notable Quotes

“x86 CPUs, the hyperscale CPUs were designed to be rented.”
Jensen Huang, on why NVIDIA’s Vera CPU is built for tokens

“You know, goodwill is not a zero-sum game. It’s not limited by physics.”
Jensen Huang, on wanting others to succeed

“So the question is, are you optimizing for someone else not having access to this technology, or are you optimizing for a flywheel?”
Jensen Huang, on export controls

“Technology without customers is useless.”
Jensen Huang, on America’s global flywheel advantage

“You really are choosing not what to buy, but who to work with.”
Jensen Huang, on picking TSMC as NVIDIA’s foundry

“The problem is, is the more specific you want it to become, the more brittle that relationship’s gonna be.”
Jensen Huang, on why he keeps partner agreements loose

“Nothing gives me more pleasure than last year when my entire supply chain’s stock price went up 300%, mine stayed flat.”
Jensen Huang, on leaving something on the table for partners

Watch the full video here.