Jensen Huang is the co-founder and CEO of NVIDIA, the world’s most valuable company and the engine powering the AI computing revolution. Thank you for listening ❤ Check out our sponsors: https://lexfridman.com/sponsors/ep494-sc https://lexfridman.com/sponsors/ep494-sc See below for timestamps, transcript, and to give feedback, submit questions, contact Lex, etc. Transcript: https://lexfridman.com/jensen-huang-transcript https://lexfridman.com/jensen-huang-transcript CONTACT LEX: Feedback – give feedback to Lex: https://lexfridman.com/survey https://lexfridman.com/survey AMA – submit questions, videos or call-in: https://lexfridman.com/ama https://lexfridman.com/ama Hiring – join our team: https://lexfridman.com/hiring https://lexfridman.com/hiring Other – other ways to get in touch: https://lexfridman.com/contact https://lexfridman.com/contact EPISODE LINKS: NVIDIA: https://nvidia.com https://nvidia.com NVIDIA on X: https://x.com/nvidia https://x.com/nvidia NVIDIA AI on X: https://x.com/NVIDIAAI https://x.com/NVIDIAAI NVIDIA on YouTube: https://youtube.com/@nvidia https://youtube.com/@nvidia NVIDIA on Instagram: https://www.instagram.com/nvidia/ https://www.instagram.com/nvidia/ NVIDIA on LinkedIn: https://www.linkedin.com/company/nvidia/ https://www.linkedin.com/company/nvidia/ NVIDIA on Facebook: https://www.facebook.com/NVIDIA/ https://www.facebook.com/NVIDIA/ NVIDIA on GitHub: https://github.com/NVIDIA https://github.com/NVIDIA Nemotron: https://developer.nvidia.com/nemotron https://developer.nvidia.com/nemotron
Huang says AI workloads no longer fit inside one computer, so NVIDIA co-designs chips, systems, software, and data center infrastructure together. The hardest part is optimizing across many interacting bottlenecks like networking, memory, power, and cooling.
He frames the core challenge as distributing algorithms so performance improves faster than simply adding computers. Once compute is sped up, networking and other non-compute parts dominate and limit end-to-end speedups.
Huang describes a company design that mirrors the product, with many domain experts engaged in shared problem-solving. He avoids one-on-ones and prefers group discussions where experts cross-check trade-offs across the stack.
NVIDIA’s evolution is described as carefully widening from narrow acceleration to broader computing while retaining specialization. Steps included programmable shaders, FP32, CG, and then CUDA as a general computing layer.
He recounts putting CUDA on consumer GeForce GPUs to build a massive install base for developers. The move increased costs sharply, crushed margins, and temporarily hurt valuation, but seeded long-term platform adoption.
Huang argues architecture success is driven more by install base and developer reach than elegance. He claims CUDA’s large install base and continuous improvement make it the default target for developers.
He explains major bets as the result of long, gradual alignment rather than sudden reorganizations. Huang says he “lays bricks” publicly and internally so big decisions feel obvious when announced.
He emphasizes NVIDIA designs and optimizes vertically but opens the stack for integration by others. This depends on convincing partners and ecosystems ahead of product readiness so adoption is primed.
He outlines pre-training, post-training, test-time, and agentic scaling as distinct growth mechanisms. He argues the loop between them keeps advancing capability and increases the centrality of compute.
Huang says data limits are eased by synthetic generation and augmentation, shifting constraints toward compute. He frames much human knowledge exchange as effectively synthetic and iterative.
He rejects the idea that inference will be simple or cheap, calling it reasoning, planning, and search. Test-time scaling, in his view, makes inference highly compute intensive.
He describes agents spawning subagents to form large teams that use tools, databases, and workflows. This “AI multiplying AI” becomes a new scaling axis that generates more data and capabilities.
He notes model architectures evolve faster than hardware cycles, forcing long-range prediction. NVIDIA tries to hedge with internal research, broad partner exposure, and flexible architectures like CUDA.
Huang says the unit of compute has shifted from GPU to cluster to AI factory. He describes modern systems like NVL72 and Rubin-era pods as massive, component-dense, supply-chain-built supercomputers.
He highlights power availability and efficiency as the major constraint for continued scaling. NVIDIA targets orders-of-magnitude improvements in tokens per second per watt to drive token cost down.
He proposes exploiting the grid’s typical spare capacity rather than building only for peak extremes. That requires flexible contracts, data centers that gracefully degrade, and utilities offering tiered guarantees.
He describes constant work with upstream and downstream partners on capacity and investment timing. He claims NVIDIA guides suppliers with forecasts and first-principles reasoning to unlock multi-billion-dollar buildouts.
Huang praises Musk’s systems thinking, urgency, and minimalist questioning that compresses timelines. He suggests presence at the point of action and prioritization across suppliers enables rapid buildouts.
He contrasts first-principles optimization against incremental improvement as a starting point. The approach benchmarks designs against physical limits, then makes explicit trade-offs for real systems.
He credits China’s talent base, intense internal competition, and rapid knowledge sharing. He also points to strong math education, software-era timing, and open-source culture as accelerants.
He frames open models as necessary for broad adoption, research, and non-language domains like biology and physics. He says NVIDIA open-sources weights, data, and recipes while also maintaining proprietary products.
Huang praises TSMC’s ability to manage dynamic global demand while sustaining yields, cost, and service. He highlights trust and execution reliability as an intangible advantage beyond transistor technology.
He argues computing is shifting from retrieval and storage toward generative production. Tokens become a segmented commodity with growing willingness to pay, making “factories” directly tied to revenue generation.
He calls agentic systems a breakthrough application that drove rapid adoption and attention. The idea is that tool-using agents make AI directly useful, changing how people work with computers.
Huang says he manages stress by breaking problems into actionable parts and delegating. He also emphasizes systematic forgetting, public humility, and staying focused on the next opportunity.
He argues tasks will change faster than job purposes, citing radiology as an example where AI adoption coincided with rising demand. He advises everyone to become proficient with AI tools to stay competitive.
He predicts “coding” becomes primarily specifying goals and architecture in natural language. That expands the pool of effective software creators dramatically, while preserving value in engineering principles.
He suggests AGI depends on definition and claims systems today could create a short-lived billion-dollar company via virality. He also warns that the chance of agents building an NVIDIA-like enduring company remains near zero.
He separates intelligence as functional and increasingly commoditized from humanity and character as deeper values. He argues AI may push society to elevate compassion, generosity, and purpose over raw cognition.
He notes GPUs are already used in space for onboard imaging to reduce downlink data. He sees space as a longer-term exploration while prioritizing near-term gains like eliminating power and infrastructure waste.
He says he fears death because the work feels historically important, but focuses on continuously transferring knowledge instead of formal succession planning. He describes constant reasoning in public and in meetings as institutional memory-building.