On November 8, 2006, NVIDIA unveiled CUDA (Compute Unified Device Architecture) and its first compatible GPU, the G80 — sold to consumers as the GeForce 8800. Marketed as a product launch, it was in fact an internal wager. Huang had decided that every NVIDIA GPU, including the ones shipped to teenage gamers, would carry a full general-purpose instruction set and toolchain. The cost was real: roughly twenty percent more transistor area per die, higher per-unit prices, lower gross margins — and almost no customer who actually needed the capability.
The intellectual origin was Stanford professor Ian Buck, whose 2004 paper "Brook for GPUs" laid the academic groundwork. Huang recruited him in 2005. Turning a research project into a company-wide strategy, however, required a CEO willing to commit. In *The Nvidia Way* (2024), Tae Kim records the boardroom exchange: a director asked when this market would reach a billion dollars. Huang answered, "I don't know. Maybe five years, maybe ten. But if we don't build it, AMD and Intel will."
The following decade was a slow grind. During the 2008 financial crisis, NVIDIA's stock collapsed from $37 to $6 while analysts hammered on the R&D burn — about $500 million a year, or 15 to 20 percent of revenue. In 2009, Intel launched Larrabee in an attempt to displace NVIDIA; AMD pushed OpenCL to route around CUDA. Huang did not flinch. Then, in September 2012, a graduate student named Alex Krizhevsky in Geoffrey Hinton's lab at the University of Toronto trained AlexNet on two GeForce GTX 580 cards and slashed the ImageNet error rate from 26 percent to 15 percent. CUDA had finally surfaced its killer application. Even so, as late as 2016 the data-center segment was still only about 12 percent of NVIDIA's revenue.
The lesson is subtler than it first appears. CUDA was not a bet on AI — nobody in 2006 could have forecast the deep-learning revolution. Huang was betting on the deeper proposition that accelerated computing would replace general-purpose computing, because Moore's Law was slowing and parallel architectures would win. As Kim records, his 2006 boardroom argument was simply: "Transistors will stop doubling every two years, but application demand for compute won't." That is not vision. That is physics. Between 2006 and 2022, CUDA quietly hardened into the moat of the AI era. PyTorch, TensorFlow, cuDNN, TensorRT — all of it sits on CUDA. When ChatGPT detonated in late 2022, the world discovered there was no AI chip you could buy that didn't depend on it.
Sources: Tae Kim, *The Nvidia Way: Jensen Huang and the Making of a Tech Giant* (W.W. Norton, 2024), Ch. 6–8; Stephen Witt, *The Thinking Machine: Jensen Huang, Nvidia, and the World's Most Coveted Microchip* (Viking, 2025), Part II; Krizhevsky, Sutskever, Hinton, "ImageNet Classification with Deep Convolutional Neural Networks," NIPS 2012.On April 5, 1993 — Jensen Huang's thirtieth birthday — he and two partners, Chris Malachowsky and Curtis Priem, founded NVIDIA in a Denny's diner in East Palo Alto, California. He still drops in occasionally. In 2023, the restaurant mounted a plaque on its exterior wall reading "Birthplace of NVIDIA." They started with $40,000, no product, no customer, and a single bet: that 3D graphics would go mainstream.
In 1995, NVIDIA shipped its first product, the NV1 — and it nearly killed the company. The NV1 used quadratic texture mapping, while Microsoft's newly released Direct3D had standardized on triangle polygons. NV1 was incompatible with what was about to become the industry's universal API. By 1996, nearly every partner had returned the chip and NVIDIA had nine months of cash left on the balance sheet.
At an all-hands meeting in 1996, Huang said the line that has since become a kind of corporate totem: "We are 30 days from going out of business." He then did three things. He cut headcount nearly in half, from a hundred down to about fifty. He killed the proprietary quadratic-surface roadmap, admitted the mistake, and committed wholesale to triangle polygons and DirectX. And he bet the next product on an unproven workflow: a software emulator. Lacking the cash to run real silicon test wafers, the team built a software simulation of the entire chip, debugged it there, and aimed for first-pass silicon success — something almost no fabless company attempted at the time. Kim records that when the RIVA 128 finally shipped in 1997, NVIDIA had one month of cash left in the bank.
RIVA 128 sold a million units in its first year. In 1999, the GeForce 256 launched and NVIDIA coined the term "GPU." The company went public that January; the stock rose sixfold by year-end. "We are thirty days from going out of business" has opened every executive meeting since — not as a marketing slogan, but as a genuine internal discipline. On the Acquired podcast in 2023, Huang put it this way: "Every morning I wake up and my first thought is, today might be the day this company dies. That keeps me from ever relaxing. If you're not afraid, you make arrogant decisions."
Sources: Tae Kim, *The Nvidia Way*, Ch. 3–4; Stephen Witt, *The Thinking Machine*, Part I; Jensen Huang on the Acquired Podcast, October 15, 2023 (three-hour interview with Ben Gilbert and David Rosenthal); Don Clark, "Nvidia's Founder on His Company's Wild Ride," *Wall Street Journal*, March 18, 2024.Huang's management style is regarded in Silicon Valley as the anti-Google template. He has roughly sixty direct reports — more than any other Fortune 500 CEO, where the typical span is eight to fifteen. He does not hold one-on-ones; he considers them an instrument of information silo-ing, and believes the CEO's job is to keep information flowing in front of everyone at once. Each week, every executive emails him a "Top 5 things" list — five lines, the five most important issues. As Kim puts it, "He reads thousands of these and uses the patterns inside them to find where the whole company is starting to go wrong."
His other ritual is the round-table meeting. NVIDIA's executive sessions routinely seat thirty to fifty people around a single table, and anyone present may challenge any decision, including Huang's. He genuinely believes that being publicly questioned is the engine of executive growth. He himself is regularly challenged by mid-level managers in those rooms, and he does not interrupt. At Stanford in 2024 he said: "I learn in public. I make mistakes in public. If you protect your ego from being challenged, you stop learning."
The signature black leather jacket was bought for him at Neiman Marcus in the 1990s by his wife, Lori Mills. He has worn essentially the same jacket to almost every public appearance since — there are more than a dozen near-identical versions in his closet. Lori was his lab partner at Oregon State University. When Huang arrived in the early 1980s, there were only about two hundred Asian students on the entire campus; he and Lori paired up in freshman year. At an OSU alumni event in 2024 he recalled: "I was the worst student in my electrical engineering class. The best student was sitting next to me." They married in 1985.
On work intensity — he puts in roughly fourteen-hour days and rarely takes weekends. But in a 2023 interview he was careful to qualify it: "I don't reply to email at 2 AM as a flex. I reply to email at 2 AM because I'm awake at 2 AM thinking about whether we're going to die." His fear of failure was forged early. In 1972, when he was nine, his parents sent him and his older brother from Taiwan to a boarding school in Kentucky called the Oneida Baptist Institute — essentially a reform school for troubled boys. Daily life included cleaning dormitory bathrooms and being beaten by older students. On 60 Minutes in 2024 he said simply: "Those two years taught me that a person can endure far more pain than they think. That's why I never complain anymore."
Sources: Tae Kim, *The Nvidia Way*, Ch. 2 and 12; Jensen Huang, "View From The Top," Stanford Graduate School of Business, March 6, 2024; Jensen Huang on 60 Minutes, November 12, 2023 (interview by Bill Whitaker); Patrick McGee, "Nvidia's Jensen Huang: 'When you're famous, things you say carry too much weight,'" *Financial Times*, April 5, 2024.NVIDIA crossed three trillion dollars in market capitalization in 2024, becoming the most valuable semiconductor company in human history. The controversies surrounding Jensen Huang have grown at roughly the same pace.
First, the de-facto-monopoly claim around CUDA. CUDA is closed source. PyTorch and TensorFlow can technically be run on AMD ROCm or Intel oneAPI, but in production deployments more than ninety percent of the workload still runs on CUDA. In July 2024, France's Autorité de la concurrence published a report explicitly identifying NVIDIA as a risk for "abusive conduct" in the AI-accelerator market. In August 2024 the U.S. Department of Justice opened an antitrust investigation, focused on whether NVIDIA effectively forces customers buying H100 and B200 GPUs to take the rest of the stack — DGX systems, NVLink, InfiniBand networking. The parallel with Microsoft in 1998 is hard to miss: a platform whose user base is too large to opt out of, suddenly being told that "infrastructure" implies "neutrality."
Second, the total dependence on TSMC. NVIDIA is fabless. Every advanced GPU is fabricated at TSMC's plants in Tainan and Nanjing. In 2024 the H100, H200, and B200 all run on TSMC's 4N or 5N nodes. If conditions across the Taiwan Strait deteriorate, NVIDIA has no Plan B — Samsung and Intel Foundry are at least two years behind TSMC at the 3-nanometer and 5-nanometer nodes. Stephen Witt devotes a chapter of *The Thinking Machine* to this exposure: Huang flies to Taipei at least four times a year and maintains close personal relationships with Morris Chang and C. C. Wei. It is the largest single geopolitical risk on NVIDIA's balance sheet.
Third, the double-bind in China. After the U.S. Commerce Department's October 2022 AI-chip export controls, NVIDIA launched the H800 (an "export-compliant" H100 with reduced NVLink bandwidth); when that too was banned in 2023, it shipped the H20. The strategy of trimming performance to remain salable has been criticized in the U.S. Congress as exploiting loopholes, and inside China as offering second-class silicon. By 2024, Huawei's Ascend 910B had begun taking meaningful market share inside China — a reality Huang has to live with. His Chinese customers (ByteDance, Alibaba, Tencent) keep buying his chips while simultaneously accelerating their own substitutes. In November 2024, after Washington tightened the rules again, Huang warned publicly: "If America won't let NVIDIA sell to China, China will build it themselves." The line did not land well in Washington.
Fourth, the theater of the launches. The annual GTC keynote has become something close to a religious convention for the AI industry, and Huang's stage claims regularly invite charges of over-promising. The Hopper performance numbers he announced at GTC 2022 turned out, on third-party benchmarking, to depend on carefully cherry-picked workloads. *The Information* reported in late 2023 that the "AI factory" pitch is in practice a bundle — compute, networking, software, and services priced together on the same GPU, making line-item comparison difficult. Huang rejects the monopoly framing. His standard answer: "We earn our market every day."
Stacked together, the four pressures define Huang's next decade. A company that nearly died in 1996 now controls the most consequential technology stack of the next thirty years — and the ethical weight of that position is something he has not yet had to publicly reckon with. Internally he still preaches that NVIDIA is thirty days from extinction. Externally, the market and the regulators see a company that is already too large to fail. Closing that gap between the internal story and the external one is the management problem of his coming decade.
Sources: Autorité de la concurrence, "Avis sur le fonctionnement concurrentiel du secteur de l'IA générative," July 2024; Stephen Witt, *The Thinking Machine*, Part IV; David McLaughlin and Leah Nylen, "Nvidia Hit by Antitrust Probe From US Justice Department," Bloomberg, August 1, 2024; *The Information*, "Inside Nvidia's Bundle Strategy," December 2023; Jensen Huang interview at the DealBook Summit, December 4, 2024.Note: remove any single beat and the explanation for "how NVIDIA became NVIDIA" cracks. The six-year stretch between CUDA (2006) and AlexNet (2012) is the period of strategic loneliness — and it is what determined the 2023 explosion.
The decision to ship CUDA in November 2006 did not produce a single quarter of board-acceptable return for ten years. R&D burned $500 million annually; the market refused to buy; analysts asked, every earnings call, when the project would be killed. Huang's one-sentence answer was always the same: "Physics tells us the CPU will hit a wall, and when that happens the world will need the GPU." Put yourself in his position. If you can see a trend that won't pay off for ten years — but every KPI, every shareholder, every colleague is telling you there is no market for it now — could you place the bet? And the harder question: how do you distinguish "seeing the future" from "betting wrong, stupidly"? Huang himself says he has no answer. He has only one rule: If I don't do this and ten years from now I regret it enough to want to die, then I do it. Is there anything on your list that meets that test?
Jensen Huang is the only sitting Fortune 100 CEO who casually wishes "pain and suffering" upon graduating students — and means it as a blessing. His 32-year run at Nvidia is built on a single counterintuitive instinct: a company that believes itself to be perpetually 30 days from bankruptcy makes sharper decisions than one that believes itself to be winning. He has lived this instinct three times. In 1996, after the NV1 disaster, he halved his staff and bet the company on software emulation. In 2006, when no customer needed general-purpose GPU compute, he committed roughly 15% of annual revenue to CUDA for sixteen consecutive years. In 2022, when ChatGPT exposed Nvidia as the only viable AI infrastructure vendor, he refused to raise prices proportionally to demand — preserving customer goodwill that competitors are still trying to penetrate.
The harder lesson, though, is structural. The same CUDA monopoly that makes Nvidia indispensable to the AI economy now invites the same antitrust scrutiny that crippled Microsoft a generation earlier. The same dependence on TSMC that delivers Blackwell-class performance creates the single largest geopolitical exposure in modern technology. Huang's leather jacket has become the closest thing the chip industry has to a papal robe — and the discipline of a CEO who once cleaned dormitory toilets in rural Kentucky is now being tested against the responsibility of running, in effect, the world's most important utility. Watch what he does in the next decade. The 30-day rule has held for thirty years; whether it can survive its own success is the open question.