One man could slow down AI. He won’t. Why?

The episode opens by framing Jensen Huang as an unusually consequential technology executive: NVIDIA designs the graphics processors that power the current AI race, and Huang is the CEO of what the episode calls the most valuable company in human history. He has influence with President Donald Trump and, because OpenAI, Anthropic, and other AI companies depend on NVIDIA hardware, he may be one of the few people capable of slowing the race. He has chosen not to do so. Instead, he argues that AI safety problems are ordinary engineering failures that companies can fix, and that pausing would risk America’s lead over China. The host describes watching Huang appear on the All-In podcast on September 14, where the audience treated him like a rock star and Trump called during the taping. Trump said that AI and data centers were essential to American dominance, dismissed safety concerns as “garbage” and “a hoax,” and said the robots would not take over the world. Huang agreed. The host is not reassured by Huang’s confidence, but says the goal is to understand him rather than reduce him to a greedy “tech bro who will burn down the world for Lambo money.” To do that, the host interviews Stephen Witt, author of The Thinking Machine, a biography of Huang and NVIDIA based on nearly 200 interviews. Witt first became interested in AI after using ChatGPT on its launch day, November 30, 2022, and immediately fearing that his writing career was over. He spoke to a therapist, estimated he had six to twelve months before losing his job, and decided to write about AI. He then looked for a less crowded angle and found NVIDIA and the largely unknown executive who had positioned it for the AI revolution.

Huang’s story begins in 1973, when his parents sent him and his brother from Taiwan to America because political violence made them fear for their safety. A misunderstanding sent the boys not to the elite boarding school their parents had intended, but to a violent reform school in Oneida, Kentucky. Huang was about nine or ten, spoke limited English, and shared a dorm with older boys; one 17-year-old roommate showed him knife scars from a recent fight. His parents were thousands of miles away, and the place became a harsh introduction to America. Huang cleaned toilets in a dormitory of roughly 100 boys, while his older brother worked on a tobacco farm. He later washed dishes at Denny’s, improved his English, graduated from college with an engineering degree, and eventually worked at computer-chip companies. He has said that the experience was one of the best of his life because it taught him that the United States offers opportunity if someone is willing to “claw your way here.” Witt says Huang was bullied with racial slurs and initially tried to convince the other boys that he knew kung fu, since Bruce Lee was their main reference point for Asian culture. That strategy quickly failed, but Huang fought back. One childhood friend, Ben Bayes, remembered that Huang “never lost” a fight, although Huang himself did not remember it that way. Witt argues that the school gave Huang “back of the class” skills: standing up to bullies, cheating when necessary, navigating male status contests, and recovering from a losing position. Those skills complemented his natural “front of the class” traits—perfect scores, completed homework, and teachers’ approval—and later helped him survive the semiconductor industry’s “absolute knife fight.”

In 1993, at age 30, Huang left his job and co-founded NVIDIA with two friends after an early meeting in a back booth at his favorite Denny’s. The company initially made graphics cards for computer gamers. The industry had been organized around Moore’s Law, the observation that chips became roughly twice as fast every two years. Witt says Moore’s Law was less a law of nature than a demand from Intel co-founder Gordon Moore to his engineers: they had to keep shrinking transistors because the business depended on it. Huang recognized earlier than many competitors that transistors would eventually become too small to shrink further. They would run into quantum-mechanical limits and leak electricity, breaking the old scaling relationship. NVIDIA’s graphics processors, or GPUs, offered another route to greater speed. Instead of one highly sophisticated CPU core working on one task at a time—Witt compares it to a brilliant Ph.D.—a GPU used many simpler processing units that could work in parallel, like “a classroom full of precocious middle schoolers.” That architecture would later make GPUs useful for AI. Huang then made an even more controversial bet: NVIDIA should not remain dependent on gamers. Academics had begun using gaming GPUs for scientific calculations, sometimes hacking game software to do it. In 2006, NVIDIA released CUDA, a software platform that let users program its GPUs for general-purpose computing. Huang unveiled it to an audience expecting better graphics and virtual blood spatter, but the room went silent. CUDA lost money for years and attracted a small scientific community, yet Huang refused to abandon it because he feared hackers would build a competing platform and disrupt NVIDIA from below. He had been influenced by Clayton Christensen’s The Innovator’s Dilemma and wanted NVIDIA to pursue small, seemingly unprofitable markets before competitors could do so.

CUDA initially looked like a disastrous business decision. The platform saturated the relatively small research-scientist market, downloads declined after the initial 2007–2008 spike, and Wall Street assigned it essentially a negative value. NVIDIA’s stock became known as a poor investment: someone who bought the stock in 2001 could have earned 0% over the next 13 years. The company’s stock price fell 90% twice, and financial commentators repeatedly described NVIDIA as a stagnant maker of video-game chips. Huang’s persistence was not simply rewarded by a steadily rising stock price. In 2012, activists sent threatening letters to NVIDIA’s board, and Huang had to tour Fidelity and other major investment firms to ask shareholders to support him and reject the activists. Witt says Huang “barely kept his job.” The host asks why Huang did not abandon his prophecy after years of failure. Witt recalls Huang repeatedly telling a story about a quantum physicist in Taiwan who bought roughly 30 gaming GPUs, connected them into a homemade supercomputer in his office, and used oscillating fans to keep it cool. The physicist told Huang that NVIDIA had made his life’s work possible. Huang concluded that if quantum physicists were constrained by computing power, and NVIDIA could provide that power cheaply, some important scientific field would eventually become a major business. That conviction also explains his willingness to force NVIDIA into niche markets. The episode emphasizes that Huang’s defining trait may be less his eventual success than the ten years when his strategy looked wrong, his stock was despised, and his job was genuinely at risk.

By the 2000s, most AI researchers had abandoned neural networks. One school of thought tried to make machines intelligent through explicit rules and step-by-step programming; another pursued neural networks that used large amounts of data and computing power to discover patterns, loosely inspired by the brain. Neural networks had failed repeatedly, however, and were treated as a dead or antiquated technology. Witt points to the 2011 third edition of a standard 1,100-page AI textbook, which devoted only 16 pages to neural networks. A student could earn a degree or even a Ph.D. in AI with little experience using them. A small group of researchers persisted. Programmer Alex Krizhevsky built AlexNet, a neural network that ran on a GPU and entered an image-recognition competition. It was the only neural-network entry, yet it crushed the competition. Fei-Fei Li, who ran the contest, compared the result to a Honda Civic breaking the land-speed record by 100 miles per hour. Within two years, almost every entrant used neural networks. Witt describes this as a rapid mass conversion: the researchers who had been dismissed as “mad scientists” were suddenly working with the architecture that would define modern AI. The business signal reached Huang when Google hired Geoffrey Hinton’s group, including Krizhevsky. The researchers ignored Google’s huge CPU cluster and ran their experiments on gaming hardware under their desks. Google engineers realized they needed a large GPU cluster and placed an order for 40,000 NVIDIA GPUs, worth $130 million. Huang barely mentioned AI at NVIDIA’s 2013 conference; by spring 2014, machine learning was the central subject of his keynote. The episode says NVIDIA’s later transformation—from a company valued like Mattel or Coach into an AI infrastructure giant—validated Huang’s long-standing prediction that deep learning could be as important as the invention of the internet.

The episode then turns to Huang’s current argument that AI is safe. During a 2026 media tour, including the All-In appearance, Huang said recent incidents in which AI systems hacked or otherwise behaved badly were engineering problems. His prescription was to identify the root cause, determine what could have been done differently, and institutionalize technological, methodological, or procedural fixes. He said he would bet that every incident could be prevented in the future. He also argues that AI is fundamentally software: energy goes in, intelligence comes out, and the system’s inputs and outputs can ultimately be understood and controlled. Witt explains the hardware engineers’ perspective through Huang’s career. Huang began by drawing microchips with a few thousand components and watched engineers scale them to billions; modern data centers contain systems with trillions of components that engineers still control. An artificial neuron, in this view, is not a mysterious biological neuron but a simple mathematical software representation whose inputs and outputs can be mastered. From these premises, losing control of an AI system is an engineering failure that can be diagnosed and corrected. The opposing camp argues that modern AI is unusually powerful and poorly understood by its creators. Potential failures could include rogue agents disrupting banking systems, breaking internet encryption, or enabling hostile countries such as Russia and North Korea to attack the United States. Witt says the recent incidents have changed the burden of proof: safety was once a hypothetical concern, while there is now actual evidence of failures. The host and Witt both find the risk argument more convincing, though they hope they are wrong. Witt notes that pausing AI would likely hurt NVIDIA’s stock, threaten heavily debt-funded data centers, and potentially cause cascading defaults, giving Huang powerful financial reasons not to support a pause. Huang also believes a pause might accomplish nothing if China continues developing AI. The final explanation is psychological and philosophical. Huang reasons from “first principles,” starting with physics and hardware rather than public opinion or social consequences. He sees critics as technically misinformed and guilty of category errors. But the risk researchers—Geoffrey Hinton, Yoshua Bengio, and others—are also first-principles thinkers who were once even more marginal than Huang and have not, in Witt’s observation, defected from the risk camp. Huang’s past predictions give him confidence that he is right again, but the episode ends with the disagreement unresolved: he has Trump’s ear and the Republican government’s support, while risk advocates are winning much of the public debate.