Nvidia Has a Secret

Key Points

▶ 0:42 Jensen was sent to a reform school in Kentucky at age nine, where he adapted by helping his 17-year-old roommate with math and learning weightlifting, showing early resilience.
▶ 2:35 At AMD during the 1985 chip downturn, Jensen absorbed a career-defining lesson: if you're not constantly innovating and expanding your mission, you're falling behind.
▶ 4:13 LSI Logic's record IPO return and its focus on application-specific integrated circuits (ASICs) convinced Jensen that specialization was the future of chip design.
▶ 5:29 The company was founded after Jensen met with two Sun engineers at Denny's; they pitched building a chip to offload 3D graphics from the PC's CPU, and Jensen committed fully.
▶ 6:59 When Jensen resigned from LSI Logic, CEO Wilf Coran was impressed by his answers and introduced him to Don Valentine, whose Sequoia funding was based on years of earned trust rather than Jensen's nervous pitch.
▶ 8:45 Nvidia's purpose was to make gaming a new storytelling medium, but realizing it required building APIs and convincing 2D developers to adopt 3D, which led to the early Sega partnership and a major technical blunder over geometric primitives.
▶ 10:56 Sega abandoned Nvidia's Dreamcast chip, and with memory prices halving and rivals benefiting from Moore's Law as followers, Nvidia was left hemorrhaging cash with only ~9 months of runway.
▶ 12:14 Jensen Huang's last-ditch plan was to get around Moore's Law by betting on unproven chip emulation software, compressing the normal 18–24 month design cycle into just 6 months with a skeleton team of 35.
▶ 13:31 The gamble produced the Revo 128, and by iterating every ~6 months, Nvidia was able to double performance per price by 1997—effectively running Moore's Law at 3x speed.
▶ 14:49 Nvidia launched the GeForce 256 in 1999, which Jensen Huang called the first GPU; it delivered up to 50%+ frame-rate gains and capitalized on slowing CPU improvements by leveraging parallel processing.
▶ 16:21 Nvidia went public on January 22, 1999 at $12/share, a $600M valuation; a $10,000 IPO investment would have grown to over $13 million by December 2021, giving the company capital and credibility.
▶ 17:12 A $500M Microsoft Xbox deal, powered by the GeForce 3 with programmable shaders, drove explosive growth: revenue doubled from $158M (FY1999) to about $1.4B by the Xbox's release, with new chip generations shipping every few months.
▶ 19:48 Jensen Huang wanted GPUs ready for uses beyond gaming, setting up Nvidia's potential reinvention.
▶ 20:16 A Stanford researcher used GeForce cards on a son's advice, cutting computation from two weeks to hours and beating his supercomputer by 10x.
▶ 20:48 Whether the anecdote is true matters less than its impact: it motivated Jensen to evolve Nvidia beyond gaming.
▶ 20:54 Nvidia's original mission was to enable new ways to tell stories, already succeeding in making games deeper and more immersive.
▶ 21:11 Gaming was not the end goal—it was simply the most obvious early use case for Nvidia's chips, not the ultimate purpose.
▶ 21:27 Jensen's ambition to help researchers make discoveries led to his biggest and boldest bet yet: pivoting toward scientific computing and building CUDA.
▶ 21:29 Nvidia's first-order question was the size of the addressable market, and from a quantitative standpoint it looked too small—customers were mostly universities and niche researchers in scientific domains.
▶ 22:04 Jensen was visionary and stubborn: despite market research, he fell in love with the concept, holding two convictions—innovation would expand use cases over time, and this play was the path to Nvidia's durable differentiation.
▶ 22:36 Jensen's logic inverted the usual saying: "If you don't build it, they can't come"—the best way to generate massive revenues is to create entirely new markets by building the foundational hardware first.
▶ 22:48 Creating new markets is the route to new revenues, and once the market size is set, the essential question becomes “What will it take to get where you want to go?”
▶ 23:03 Nvidia needed a new kind of technological platform—one that could rival the invention of the transistor in transforming computing.
▶ 23:11 The late-’90s GeForce, called the first GPU, enabled parallel execution of graphics tasks for faster, more dynamic rendering—and required extending the C programming language.
▶ 23:32 Nvidia introduces CG ("C for Graphics"), a high-level shading language developed with Microsoft for vertex and pixel shaders, but it was only a graphics-focused stepping stone toward a broader goal.
▶ 23:49 Jensen's actual vision required a full-stack API usable beyond graphics, not just a language for shaders.
▶ 23:56 Researcher Ian Buck's 2003 Brook model extended C with data-parallel constructs; by 2006 he joined Nvidia to build a general-purpose parallel computing platform for complex computational problems.
▶ 24:26 CUDA (Compute Unified Device Architecture) was introduced, with version 1.0 shipping in November 2006 as a "skeleton key" to new computing possibilities.
▶ 24:36 Many scientific and AI workloads have massive inherent parallelism, so CUDA offloaded complex functions from the CPU to GPUs, dramatically accelerating performance.
▶ 24:53 This accelerated power transformed fields including computational finance, fluid dynamics, weather forecasting, medical imaging, quantum chemistry, and electromagnetics.
▶ 25:02 CUDA's impact spans many scientific fields, ranking just behind the transistor and microprocessor among the greatest computing inventions.
▶ 25:12 The real breakthrough was the business model: CUDA is free, but it only runs on Nvidia's proprietary hardware.
▶ 25:26 This strategy mirrors Apple's ecosystem, though Nvidia's approach predates the iPhone — with both creating commercially viable platforms.
▶ 25:44 Parallel computing served as the foundation for both Apple and Nvidia, creating highly profitable business models.
▶ 25:51 Despite this, Nvidia's stock dropped about 50% in 2011 after missing earnings targets.
▶ 26:02 CUDA was invented before the market was ready for it, and it would take an unexpected twist to unlock its true value.
▶ 26:08 ImageNet’s emergence is framed as a matter of “fate” catching up with technology, with timing crucial for AI’s future.
▶ 26:10 In 2000, Fei-Fei Li began ImageNet, inspired by Princeton’s 1980s WordNet lexical database.
▶ 26:20 Her goal was millions of correctly labeled images for AI recognition, and she later launched a competition for researchers to use the database.
▶ 26:35 In 2012, University of Toronto researchers Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton entered the competition with their algorithm, AlexNet.
▶ 26:42 Their submission didn't just win—it dominated the competition.
▶ 26:47 AlexNet successfully recognized about 85% of the images it was presented with, a decisive and impressive result.
▶ 26:49 AlexNet's breakthrough: it correctly identified ~85% of images, proving the transformative power of a neural network.
▶ 26:58 Deep learning wasn’t new, but had been "basically impossible" in traditional computing due to extreme computational demands.
▶ 27:07 The "big bang moment" came when Toronto researchers trained AlexNet in CUDA on NVIDIA GPUs—marrying parallel power with deep learning and making CUDA the killer app; Jensen later argued this wasn't luck but years of GPU pioneering that made Nvidia the hardware supplier of choice.
▶ 27:46 Machine learning created entirely new revenue streams for Nvidia, transforming its business beyond traditional PC graphics.
▶ 28:05 Enterprise customers became the core of the opportunity, with top-tier GPUs powering massive data centers for Google, Microsoft, and Amazon.
▶ 28:22 Tesla's Dojo example showed Nvidia's market strength: a single customer using chips for one purpose still generated at least $50 million in revenue.
▶ 28:35 Rising enterprise revenue provides the concrete, measurable answer to the question "How big is this market?" that Wolf Coran and Don Valentine originally posed to Jensen Huang.
▶ 28:43 The market is definitively gigantic and growing every day, fueled by the world's biggest companies investing in deep learning for applications like fighting disease, autonomous vehicles, and natural sciences.
▶ 28:57 Nvidia GPUs have become the default infrastructure for these ambitious AI efforts, validating the enormous scale of the market.
▶ 28:58 Nvidia's GPU dominance became definitive, with no turning back after its inflection point.
▶ 29:03 Nvidia acquired Mellanox for $7 billion, and since then tripled its data center revenue.
▶ 29:11 Nvidia posted startup-like revenue growth: $11B, then $16.6B, then $26.9B (a 60% jump).
▶ 29:35 NVIDIA's core model is an open development ecosystem that runs exclusively on NVIDIA hardware, creating strong lock-in as applications built on CUDA can only be deployed on NVIDIA GPUs.
▶ 29:48 NVIDIA shows deep financial strength with $8 billion in annual free cash flow, over $20 billion in cash, and gross margins that climbed from 30% in the early '90s to 66% today.
▶ 29:58 NVIDIA has built and actively fortified its business, creating a durable moat around its CUDA ecosystem and GPU dominance in deep learning and high-performance computing.
▶ 30:02 NVIDIA wields pricing power by choosing where it operates on the price curve, a privilege usually reserved for monopolies.
▶ 30:12 It also holds innovation-driven power, forcing competitors to react to its new technologies in entirely new markets.
▶ 30:23 NVIDIA is rare in wielding both forms at once, achieved by pioneering graphics emulation, creating programmable GPUs, riding machine learning growth, and locking users into a proprietary ecosystem.
▶ 30:55 Nvidia is targeting a $300 billion addressable market in EV and autonomous vehicle software, with Jensen Huang still leading the push into new sectors.
▶ 31:10 The Omniverse is framed not as a consumer metaverse but as an enterprise/scientific simulation platform for testing products at 1:1 scale before real-world deployment.
▶ 31:19 Nvidia's Earth 2 aims to combine GPU computing, deep learning, and observed data to simulate ultra-high-resolution climate models, showing the limitless potential of simulation.
▶ 32:16 Nvidia’s 2006 investment in Keyhole (now Google Earth) shows Jensen Huang recognized simulation’s transformative potential early on.
▶ 32:30 Simulation promises to accelerate scientific discovery by an order of magnitude, and CUDA has made it a core scientific and industrial capability—not just a graphics tool.
▶ 32:36 Simulation already saved Nvidia: it let the company defy Moore’s law, develop its first consumer GPUs, and is the principal technology underlying CUDA itself—fulfilling Jensen Huang’s original promise to help people “tell better stories.”

Summary

Executive Summary

Key Points

Video Sections

Exact Transcript