Night Markets and Chip Foundries: A Week in Taiwan with Jensen Huang

Lex Fridman opens by saying he is visiting Taiwan for the first time, partly to hang out with NVIDIA CEO Jensen Huang and partly to attend Computex, a huge conference in Taipei. He describes the island as two-thirds mountains yet responsible for more than 90% of the world's most advanced semiconductors. Every computer, smartphone, GPU and AI data center, he says, depends on this one place.

He compares the puzzle of Taiwan's dominance to the mystery of how ancient sites such as Göbekli Tepe or the Egyptian pyramids were built. Part of the answer, he suggests, is culture and an education system that prizes engineering, and it has large technological and geopolitical implications. He hopes the US and China can compete but remain friends, and sees understanding each other's culture and working together as part of that.

Fridman also praises Huang as someone who has kept his humility and listens to everyone around him as a fellow human being. He adds that Huang is a foodie, so there is no better guide to Taiwanese food.

At a booth, Huang shows Fridman a system he says is 100% liquid cooled, with no cables or hoses. The previous Grace Blackwell generation was only 80% liquid cooled, and assembly has dropped from two hours to five minutes. In his view, things that are easy to build are also more reliable, and the product is in full production.

Discussing NVIDIA's Vera CPU, which Fridman says was shown at 1.8x over x86, Huang explains why the design logic differs. Earlier hyperscale CPUs were designed to be rented, so the goal was as many cores as possible and the economics were dollars per core. AI does not rent cores; it generates tokens and sells them, so the fundamental purpose is to go through that loop as fast as you can.

Along the Computex floor they meet partners such as SK Hynix, where Huang praises its high-bandwidth memory, and a booth with a pile of old GPUs. Huang jokes that voice control will dominate in future, while Fridman says he finds it hard to think while talking and prefers typing.

Asked if he was nervous before the keynote, Huang says he is always nervous before public speaking. Everything comes together at the last moment, there is no teleprompter, and he doesn't know all his words in advance. His aim was to distill a big message down for the whole island: if NVIDIA is to build and sell a few trillion dollars of computing, every company in the ecosystem must invest hundreds of billions.

He says he still loves gamers, since that is where NVIDIA began, and advises that no matter how high you rise you should not let your feet leave the ground. Fridman adds that some of his happiest moments came playing video games on NVIDIA hardware.

After a mobbed signing session, Fridman says it is inspiring to see an engineer treated like a rock star, because engineers build the world and rarely get the love they deserve. He thinks it will inspire kids to become engineers. Both men remark on the humility they met from engineers in Taiwan and, Fridman adds, in China.

Huang says the US needs to bring manufacturing back and calls the AI boom the best chance to invest in the grid, sustainable energy and manufacturing again. With demand so large, he argues, you can afford to manufacture anywhere, so you might as well manufacture in the US as well. He adds that AI and robotics in manufacturing make this a good time to reindustrialize.

He stresses that Taiwan is a strategic partner in this effort. Setting up world-class semiconductor fabs in the US is hard because there isn't enough local talent, so Taiwanese companies send engineers, invest money and train people. He hopes people see how real that partnership is from a technology perspective.

Asked why so many hardware companies come from Taiwan, Huang credits K. T. Li, who had a 1950s vision of bringing science and technology manufacturing to the island and helped persuade Morris Chang to return. He notes that Philips, which spun off the fab that became TSMC and the equipment business that became ASML, was attracted to Taiwan. He calls it the beginning, an implanted seed.

The timing coincided with the PC revolution. IBM opened the specifications for the chipsets on its compatible motherboards, from memory and cache controllers to the graphics controller, and Taiwanese engineers designed chipsets manufactured at TSMC and UMC. That made PC parts available to many companies, including a young Michael Dell, who bought motherboards from Taiwan, licensed an operating system from Microsoft and became a PC maker.

Fridman shares a line he heard from a woman who grew up in Taiwan, that as a girl she learned to spell CoWoS before she could spell Gucci. He takes it as a sign of how much the culture values engineering, and thanks Huang for bringing cool back to technology.

The two visit the Raohe night market, sampling oyster pancake, noodles, sweet potato balls, shaved ice and more. Vendors mob Huang, and one refuses to take his money. Fridman calls the noodle possibly the best food he has ever had and the oyster pancake the reason he came.

Huang recalls that as a child he went with his parents to night markets, which were the family's idea of eating out because they didn't have money for restaurants. He tells how, at about five years old, he wandered up to a cook who was cleaning a knife and was cut across the face, and his father ran him to the hospital. He says he always loved the night market afterward, because of how good the food was and how hard everyone works while still sharing something.

Huang also recounts his recent trip to China, visiting AI companies and robotics firms. He says they are all compute starved, but that scarcity forces them to invent efficient architectures, and he notes that constraints are the mother of invention. Despite being severely constrained, he says, they are still getting funded.

At an NVIDIA employee gathering, Huang plays host, handing out food and telling guests to enjoy Taiwanese hospitality. He describes that day's booth tour, where several thousand people lined up and cheered, as mind-blowing. He says the island's capacity for generosity, kindness and helping others is why its companies are the supply chain for the world, and adds that Taiwanese are willing to work super hard.

At a lunch of braised pig's feet, Fridman asks whether the weight of millions looking up to him ever gets heavy. Huang says it doesn't, because he has always been nice to others and wanted them to succeed, even sharing his homework as a kid. He argues that celebrating others tends to help you succeed too, noting that goodwill is not a zero-sum game and is not limited by physics.

Fridman asks Huang to steelman the case for and against export controls. Huang grants that holding back a force multiplier such as rare earths, bullets or bombs can improve your own position, and that this is generally not untrue for physical things. But he says that if the goal is long-term national security and prosperity, the better strategy is to optimize for a flywheel in which technology leadership and wealth feed each other, since the safest countries are the richest.

He compares this to platform technologies, where installed base and customer use drive more consumption, which funds faster scaling. He advocates being extremely selective about access to certain technologies but otherwise maximizing platform reach. He adds that the flywheel is delicate: it can seem to take off overnight after decades, but it can also slow quickly.

Fridman asks whether America flourishes with a flywheel that includes China or not, and says that competing as friends differs from competing as enemies, a line which if crossed could destroy everyone's flywheel. Huang agrees that countries should compete but do everything to avoid conflict, and Fridman adds that the harder it is to decouple, the harder it is to have conflict.

Huang notes that AI is not a single model but a stack of technologies, from energy upward, and that leadership in different layers sits in different places. He argues that American companies benefit from English-language reach and an American stack deployable around the world, which China lacks. He also says every country has a sovereign right to decide how to use its own resources, while the US should seek the largest possible platform reach.

At the Ningxia night market, Huang takes Fridman through fried chicken, egg crepes, ice cream wrapped in rice paper with cilantro, green mangoes, taro balls and more, in the rain. They joke that Taiwan makes semiconductors and fruit. Huang says he grew up on dishes such as the ginger soup his mother made nearly every week.

On the question of how much to eat, Huang says that in some cultures not finishing food is rude, and in others finishing everything is rude, so you are either honoring their generosity or your appreciation of the food. In Taiwan, he says, the host's generosity matters more. Fridman frames it as an optimization problem where you must figure out the objective function by country.

Huang explains that texture matters as much as taste in Taiwanese food, and one prized texture is "QQ", a bouncy, chewy elasticity found in the best noodles and sesame balls. They joke that Taiwanese dishes such as pig's feet, turnip cake and pork floss have a branding problem, while another quip pokes fun at OpenAI's naming.

Over a Taiwanese breakfast, Huang tells how NVIDIA found TSMC. In the mid-1990s, just after Windows 95, everyone wanted foundry capacity, and some foundries made customers license their technology to get it. NVIDIA, a startup making graphics chips, was turned down by SGS-Thomson, then heard of TSMC and UMC, which he says had no strings attached.

He recalls looking up TSMC's US office in the Yellow Pages and reaching only a recording, leaving messages that went unreturned, and finally writing Morris Chang a long letter. One evening Chang called back, and Huang told his colleagues, who were playing PlayStation, to be quiet. Chang invited him to Taiwan, which Huang describes as his first trip back in 30 years.

Huang says he chose TSMC over UMC, which he liked equally, because Chang said "we can't do that" to many of his requests and spoke only about corporate governance, integrity, partnership and long-term thinking. Chang promised that TSMC would jump through hoops for NVIDIA, and Huang says it has for 30 years. He adds that in a chip foundry you can't know every detail up front, so you are choosing who to work with, not what to buy.

Huang compares an overly specific contract to an ironclad prenup that makes the relationship brittle, and prefers loose, "rough justice" in which each side sometimes wins. He says partners need to know you want them to win, and that you should always leave something on the table. He says nothing pleased him more than when his supply chain's stock rose 300% last year while his stayed flat, because the platform was big enough for everyone to rise.