Lex FridmanJensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494
Lex Fridman
Why extreme co-design is necessary 0:33
Jensen Huang explains that NVIDIA moved beyond building single GPUs because AI problems no longer fit inside one computer. To go a million times faster instead of just adding more machines, you have to break apart the algorithm, split the data, and shard the model across thousands of computers. This runs into Amdahl's Law, which says your overall speedup is limited by whatever fraction of the workload you actually accelerate, so if computation is only half the problem, speeding it up infinitely still only doubles total performance. That forces NVIDIA to treat the CPU, GPU, networking, and switching all as parts of one massive distributed computing problem, since Moore's Law has slowed as Dennard scaling stalled.
How NVIDIA organizes for co-design 4:31
Huang describes running a direct staff of 60 people, almost all with engineering backgrounds covering memory, CPUs, optical systems, GPUs, architecture, and algorithms. He avoids one-on-one meetings because the work itself is collaborative by nature. Instead, problems are presented to the whole group at once, so a discussion about cooling or power distribution is heard by everyone, and anyone whose area is affected is expected to speak up. He says a company's structure should reflect what it is actually trying to produce, rather than copying a generic organizational chart.
From accelerator to computing company 9:01
Huang traces NVIDIA's evolution from a narrow accelerator company, which had deep optimization but limited market reach, toward becoming a broader computing company without losing its specialization. The path went from inventing the programmable pixel shader, to adding IEEE-compatible FP32 support in shaders, to layering the C-based Cg language on top, which eventually led to CUDA. He frames this as a constant tension: becoming more general purpose as a computing company risks diluting the specialization that made the accelerator valuable in the first place.
The risky bet on CUDA and GeForce 14:00
Putting CUDA on GeForce was, in Huang's words, close to an existential decision, because install base determines whether developers adopt a computing platform, not elegance of design. He points to x86 surviving despite criticism of its architecture, purely because of its massive install base, while many well-designed RISC architectures failed. NVIDIA decided to embed CUDA into every GeForce GPU sold, reaching millions of gamers and students, and pushed it through universities and textbooks. The move raised GPU costs by 50 percent, crushed the company's gross margins, and helped drop NVIDIA's market cap to about one and a half billion dollars for a period, though Huang calls NVIDIA the house that GeForce built.
Manifesting the future through belief 17:01
Huang describes his decision-making style as reasoning his way to a conviction about what must happen, then spending sustained time shaping the beliefs of his board, management, employees, and industry partners before making any formal announcement. Rather than sudden top-down manifestos, he lays groundwork gradually, citing the Mellanox acquisition and the shift to deep learning as examples where by the time he announced a decision, people asked what took him so long. He notes that GTC keynotes serve this same purpose on an industry scale, shaping the beliefs of partners so that when NVIDIA's product is ready, the market has already been prepared for it.
Four scaling laws driving AI progress 22:33
Huang lays out four scaling laws he sees driving intelligence forward: pre-training, post-training, test-time, and agentic scaling. He addresses the fear that running out of human-generated data would end pre-training, arguing that synthetic data, generated and refined by AI itself, keeps that scaling going, so training becomes limited by compute rather than data. He also pushes back on the earlier assumption that inference would be computationally light, arguing that inference is really thinking and reasoning, which is harder than mere memorization, making test-time scaling intensely compute-hungry. The newest layer, agentic scaling, involves AI systems spawning teams of sub-agents that generate their own experiences and data, which then feeds back into pre-training and post-training in a continuous loop, meaning intelligence ultimately scales with compute.
Designing chips through co-design and listening 29:30
Jensen Huang explains that NVIDIA anticipates future needs partly through internal research, building its own models to gain hands-on experience, and partly by working with nearly every AI company in the world and listening to what challenges they face. He stresses that the CUDA architecture balances specialization, which allows for speed, with flexibility, which allows it to adapt to new algorithms. NVIDIA is already on CUDA version 13.2, and hardware evolves alongside software: when mixture-of-experts models emerged, NVIDIA responded with NVLink 72 instead of NVLink 8, letting a ten trillion parameter model run as if on a single GPU.
From LLM racks to agent racks 31:04
Huang describes how the Grace Blackwell rack was built entirely around running large language models, while a year later the Vera Rubin rack looks completely different, adding storage accelerators, a new CPU called Vera, and an additional rack called Rock. The shift happened because the earlier design served mixture-of-experts inference, while the new one is built for agents that need to use tools. He says this wasn't guesswork but reasoning: an LLM acting as a digital worker needs to access ground truth through a file system, do research, and use existing tools rather than reinvent them, much like a person would use a microwave instead of growing a laser hand.
OpenClaw as a milestone 34:35
Huang argues that OpenClaw effectively reinvented the computer by combining tool use, file access, and research capability into one agentic system, and that he had sketched a nearly identical schematic at GTC two years earlier. He credits the moment to a convergence of forces: models like Claude and GPT reaching sufficient capability, and someone finally building an open source project robust enough for everyone to use. He compares OpenClaw's impact on agentic systems to what ChatGPT did for generative systems, and notes that NVIDIA responded quickly with a security effort called NemoClaw, giving agents only two of three risky capabilities at a time, accessing sensitive data, executing code, or communicating externally, paired with enterprise access control and policy engines.
Power and efficiency as the real constraint 37:33
Asked about future blockers to scaling agents, Huang points to power, but says NVIDIA's answer is extreme co-design that keeps improving tokens generated per watt. He claims that while Moore's Law would have scaled computing about 100 times over the last decade, NVIDIA scaled it a million times, and that this efficiency push is what keeps token costs falling by roughly an order of magnitude each year even as computer prices rise.
Shaping the supply chain years ahead 39:04
Huang describes spending enormous effort informing CEOs across the semiconductor and infrastructure industries about where demand is heading, so they can make investment decisions early. He cites convincing DRAM CEOs three years ago to invest in HBM memory when it was still a niche supercomputer component, and persuading them to adapt low-power phone memory for data centers, moves that turned into record years for those companies. He also explains how NVIDIA shifted from assembling supercomputers inside data centers to effectively manufacturing them in the supply chain via dense NVLink-72 racks, which forced suppliers to build power and testing capacity into their own factories, changes he negotiates directly with partners rather than worrying about it at night.
Rethinking grid power and uptime demands 47:30
Huang suggests that power grids sit near their worst-case capacity only rarely, running around 60 percent of peak most of the time, and proposes that data centers could contractually accept reduced power during rare peak demand, shifting workloads elsewhere or simply running slower rather than requiring guaranteed uninterrupted supply. He frames the obstacle as a three-way problem: end customers demand near-perfect uptime, CEOs often don't realize how rigid the contracts their teams sign really are, and utilities could offer more flexible power tiers if data centers were built to gracefully degrade service instead of insisting on constant full power.
Elon Musk's approach to speed 52:36
Huang praises Elon Musk's construction of the Colossus supercomputer in Memphis in four months, crediting Musk's systems thinking, his habit of questioning whether each step is necessary or could be done differently, and his practice of being physically present at the point of a problem. Huang notes that this personal urgency spreads to suppliers, who then prioritize Musk's projects among their many commitments, and Lex adds an anecdote about Musk personally studying how cables get plugged into a rack to reduce errors.
Engineering against the speed of light 56:00
Huang describes a personal design philosophy he started thirty years ago called speed of light thinking, meaning every decision, memory speed, cost, manufacturing time, is measured against the absolute physical limit rather than against past performance. He contrasts this with continuous improvement, which he dislikes, preferring to ask why a task takes 74 days and to imagine building it from scratch, which might reveal it could take six days; only then does he view the gap between 74 and six as a meaningful, well-reasoned set of trade-offs.
The Scale of a Vera Rubin Pod 59:30
Jensen describes the sheer scale of the Vera Rubin pod: seven chip types, five purpose-built rack types, forty racks, 1.2 quadrillion transistors, nearly 20,000 NVIDIA dies, over 1,100 Rubin GPUs, 60 exaflops of compute, and 10 petabytes per second of scale bandwidth, all in a single pod. He notes NVIDIA may need to produce about 200 of these pods a week. Asked whether simplicity is a useful design goal amid such complexity, he says his guiding phrase is that things should be as complex as necessary but as simple as possible, and every layer of complexity must be tested and challenged to justify itself. He calls the resulting system the most complex computer the world has ever made, praising the engineering at NVIDIA, TSMC, and ASML as world class.
Why China Innovates So Fast 1:05:04
Lex asks how China built such a strong technology sector. Jensen points out that roughly half the world's AI researchers are Chinese, and China's tech industry rose at exactly the moment software mattered most in the mobile cloud era. Because China is made up of many provinces and cities competing internally, intense domestic competition breeds strong companies. He also points to a culture where family, friends, and company loyalties overlap heavily, so schoolmates and relatives often work at different companies together, making knowledge-sharing and open source contribution natural rather than something to guard. He adds that China's leaders are largely engineers who built the country out of poverty, unlike many Western leaders who come from law, and calls China a builder nation.
NVIDIA's Open Source Strategy 1:09:32
Turning to NVIDIA's own open source model, Nemotron 3 Super, a 120 billion parameter open-weight mixture-of-experts model, Jensen explains the reasoning. NVIDIA needs to understand how AI models evolve to design good computing systems, so basic research into architectures like transformers combined with state space models feeds directly into hardware design. He believes proprietary flagship models can coexist with open source models that let every industry, country, researcher, and student join the AI revolution, since NVIDIA has the scale and motivation to keep building and supporting these models indefinitely. He also stresses that AI extends beyond language into biology, chemistry, physics, and weather, so open models must push frontiers in every domain, giving as examples wanting every car company and drug company like Lilly to have access to the best AI systems.
What Makes TSMC Special 1:10:00
Reflecting on TSMC, Jensen says the common misunderstanding is that its edge is purely technological. The real strength is its ability to orchestrate the shifting demands of hundreds of companies, running high-yield, high-throughput manufacturing while keeping promises on wafer delivery. He credits TSMC's culture for balancing technological excellence with strong customer service, something few companies achieve simultaneously, and highlights trust as an intangible asset built over three decades of business with NVIDIA conducted without a formal contract. He confirms that in 2013 Morris Chang did offer him TSMC's chief executive role, which he declined out of commitment to NVIDIA's mission, despite deep respect and friendship with Chang.
CUDA as NVIDIA's Core Moat 1:15:00
Asked about NVIDIA's biggest competitive advantage, Jensen names the installed base of CUDA, built by 43,000 employees and trusted by millions of developers who have layered software on top of it over twenty years. Developers target CUDA first because they know performance will improve roughly every six months, reach hundreds of millions of machines across every cloud and industry, and can trust NVIDIA to maintain and optimize it indefinitely. The second advantage is NVIDIA's broad ecosystem, with one architecture integrated horizontally across Google Cloud, Amazon, Azure, CoreWeave, Nscale, supercomputers, enterprise systems, radio base stations, cars, robots, and satellites.
From Chips to AI Factories 1:19:03
Jensen explains that his mental model of NVIDIA's product has shifted from a single chip to a gigawatt-scale AI factory complete with power generation, cooling, and networking requiring thousands of engineers to bring online. He touches on space-based computing, noting NVIDIA GPUs already operate in space for satellite imaging, where AI must process data at the edge before transmission, though cooling in space remains a hard engineering problem solved mainly through radiation rather than convection. He says he prioritizes eliminating wasted power on Earth first while exploring space engineering challenges like radiation and graceful degradation in parallel.
Why NVIDIA's Growth Feels Inevitable 1:24:30
After the sponsor break, Jensen argues NVIDIA's continued growth toward figures like ten trillion dollars is close to inevitable, because computing has shifted from a retrieval-based, file-storage model to a generative, contextually aware model that requires far more processing. He describes computers becoming factories that generate valuable tokens rather than warehouses that merely store data, with token pricing beginning to segment much like iPhone tiers, and predicts paying $1000 per million tokens for specialized intelligence is coming soon. He expects this productivity shift to accelerate global GDP and dramatically raise the share of that GDP spent on computation, dismissing past predictions that NVIDIA could never exceed a billion or twenty-five billion dollars as failures of first-principles thinking.
Tokens as the New Product 1:32:02
Jensen describes a future built on token factories, where computing power is measured in tokens per second per watt, and every token carries value for someone. He points to agent tools like OpenClaw as the iPhone moment for tokens, calling it the fastest-growing application in history. Lex admits to already talking to his laptop to program, and Jensen predicts AI agents will soon be the entities people communicate with most, constantly reporting back on finished tasks and asking what to do next.
Carrying the Weight of NVIDIA 1:35:01
Asked how he handles the pressure of leading a company that nations and economies plan around, Jensen says he stays conscious of NVIDIA's role in generating tax revenue, national security, and re-industrialization in the United States. He deals with the weight by breaking every worry into a clear question of what changed, what is hard, and what he can do about it, then either acting himself or telling someone who can. Once he has voiced a concern to the right person, he considers the burden shared and lets himself move on.
Low Points and Forgetting 1:39:00
Jensen admits he has hit psychological low points often. His coping method involves systematic forgetting, similar to how AI learning requires discarding some information rather than holding onto everything. He decomposes problems, shares the load by telling others, and then deliberately turns attention to the next opportunity rather than dwelling on setbacks, the way great athletes focus only on the next point rather than the last one.
The Mind of a Child 1:41:32
Jensen explains that his approach to hard problems relies on approaching them with a child's mindset, asking simply how hard something can be rather than simulating every future setback in advance. He says you have to go into new experiences expecting things to be great, then rely on endurance and grit when disappointments surprise you anyway. As long as his underlying assumptions about the future stay valid, he trusts the outcome will still happen and keeps pursuing it.
Staying Humble Through Public Reasoning 1:45:00
Jensen says wealth and success have not made it harder for him to admit when he is wrong, largely because he does so much of his reasoning publicly, where mistakes are visible to everyone. In meetings he reasons out loud step by step so people can challenge his logic rather than just his conclusion, which lets the group collectively search for the right path forward. He credits part of his resilience to tolerance for embarrassment, built over decades, including from his first job cleaning toilets at Denny's.
DLSS 5 and Gaming Roots 1:49:31
Responding to gamer concerns that DLSS 5 produces AI slop, Jensen explains the technology is conditioned on ground-truth 3D geometry and artist-created textures, enhancing frames without changing the artist's intended structure or style. He calls GeForce NVIDIA's top marketing strategy, introducing people to the brand through games like Call of Duty before they later use CUDA or professional tools. He names Doom as the most culturally influential game for turning PCs into gaming devices, and Virtua Fighter as most important from a pure game-technology standpoint, while also praising fully ray-traced Cyberpunk 2077 and community mods for games like Skyrim.
Redefining AGI and Jobs 1:55:31
Asked whether an AI could run a billion-dollar company, Jensen says he believes AGI has essentially been achieved, though building something like NVIDIA repeatedly is still unlikely for agents alone. He argues that job tasks and job purpose are related but not identical, using radiology as his central example: computer vision became superhuman around 2019, yet the number of radiologists grew because diagnosing disease, not just reading scans, remains the job. He predicts software engineer numbers will keep growing at NVIDIA, and that coding, redefined as writing specifications, could expand from 30 million practitioners to roughly a billion, elevating carpenters, accountants, and plumbers into architects and analysts of their own trades.
Learning to Code Still Matters 2:03:03
Jensen argues that understanding traditional programming languages and design principles still has value even as AI changes software creation. He describes writing specifications as an art form that depends on the problem at hand. When setting corporate strategy, he deliberately under-specifies direction, giving enough detail for 43,000 employees to act on while leaving room for them to improve on his ideas. He says everyone will need to find their own place on a spectrum between highly prescriptive instructions and open-ended, exploratory collaboration with AI, and that this spectrum is the future of coding.
Job Anxiety and Using AI 2:05:31
Jensen acknowledges real pain and anxiety around job loss from automation, saying compassion is needed for people and families affected. His practical advice is to separate what can be controlled from what can't, then act on the controllable part. Given a choice between two new hires, one AI-fluent and one not, he would always choose the AI-fluent one, whether the role is accountant, marketer, lawyer, farmer, pharmacist, carpenter, or electrician. He urges every student and worker to become expert in AI to elevate their own field rather than be replaced by it, while noting that jobs defined purely by a single automatable task face the highest risk. Lex adds that chatbots can function like a practical life coach, walking a beginner through unfamiliar tasks, from learning AI itself to planning a trip to Taiwan, and Jensen jokes that Lex should ask the AI for his favorite Taipei restaurants.
Human Feeling Versus Computation 2:11:01
Asked whether something about human consciousness is fundamentally non-computational, Jensen says he doesn't believe a chip will ever get nervous, even if AI can recognize and understand emotions like anxiety or excitement. He argues that identical circumstances presented to different people produce wildly different performance because of subjective feeling, something he doubts a computer experiences even when producing statistically varied outputs. Lex reflects on the richness of subjective human experience, love, heartbreak, fear of death, and grief, and says he remains open to being surprised by how far scaling can go.
Intelligence Versus Humanity 2:13:30
Jensen insists intelligence is a functional commodity, not something mystical or equivalent to humanity itself. He describes being surrounded by 60 direct reports more intelligent and better educated than he is in their specific domains, yet he still orchestrates them, because character, compassion, generosity, life experience, and tolerance for pain are separate qualities from raw intelligence. He believes the commoditization of intelligence through AI should inspire people rather than cause anxiety, since it frees humanity to be celebrated more, not less.
Mortality and Passing On Knowledge 2:17:31
Asked about his own mortality, Jensen says he doesn't want to die, describing his work at NVIDIA as a once in a humanity experience rather than once in a lifetime. He explains he doesn't believe in succession planning, not from any sense of immortality, but because the real answer to that anxiety is to continuously pass on knowledge, insight, and skill to his team every single day, so nothing he learns sits unshared for long. His hope is to die on the job, instantaneously, with no prolonged suffering.
Hope for the Future 2:20:03
Jensen expresses deep confidence in human kindness, generosity, and compassion, saying people's desire to do good is proven right again and again, even when he's occasionally taken advantage of. He says the reach of what can now be solved, curing disease, drastically reducing pollution, and even traveling at the speed of light over short distances, makes this an exciting time to be alive. He describes plans to send a humanoid robot into space that keeps improving during flight, and to eventually transmit his own uploaded consciousness, built from his lifetime of writing and communication, to catch up with it at light speed. The conversation closes with mutual thanks between Jensen and Lex, followed by Lex signing off with a quote from Alan Kay: the best way to predict the future is to invent it.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
