No PriorsNo Priors Ep. 13 | With Jensen Huang, Founder & CEO of NVIDIA
CHAPTERS
- 0:05 – 2:55
Jensen’s path from Oregon State to AMD and LSI—and meeting future co-founders
Jensen recounts how early recruiting and chip-design passion took him from Oregon State to AMD, then to LSI at the dawn of EDA. He highlights the formative exposure to legendary architects and systems thinkers that later shaped NVIDIA’s approach.
- •Campus recruiting and why AMD appealed to him as a chip designer
- •Move to LSI to work on early computer-aided chip design (EDA)
- •Learning from renowned computer architects and industry leaders
- •Early network that later connects to NVIDIA’s founding team
- 2:55 – 4:43
Founding NVIDIA around a contrarian bet: accelerated computing over general-purpose CPUs
Jensen explains that starting NVIDIA wasn’t initially his idea, but he joined after being persuaded by his future co-founders. The company formed around a mission to solve problems that CPUs couldn’t, betting on acceleration when most of the Valley favored general-purpose computing.
- •Chris Malachowsky and Curtis Priem recruit Jensen to start the company
- •The 99% vs 1% split: general-purpose computing vs acceleration
- •Mission focus: tackle ‘barely solvable’ or ‘unsolvable’ problems on CPUs
- •How that mission naturally leads to robotics, climate, biology, and AI
- 4:43 – 9:59
Expanding GPUs beyond graphics: CUDA, platform compatibility, and the ‘walk the line’ tradeoff
NVIDIA broadened the GPU from graphics into a more general accelerator to support post-processing, physics, and scientific workloads. Jensen details the tension between being too general (losing acceleration) and too narrow (insufficient market/R&D), and why CUDA compatibility across generations was essential to attract developers.
- •Why GPUs needed to become more general-purpose (image effects, physics)
- •Core tradeoff: generality vs domain acceleration performance
- •Economic constraint: funding R&D in a small market while competing with CPU R&D scale
- •CUDA as a new computing model and the importance of cross-generation compatibility
- •Early gross-margin pain from investing in CUDA before it had paying applications
- 9:59 – 11:35
Organic adoption: gaming cards power early AI, crypto, and scientific computing workloads
The conversation covers how a true platform gets adopted in unexpected ways, with GeForce GPUs becoming the early workhorses for deep learning labs. Jensen and Elad discuss how communities discovered GPUs for parallel compute and scaling, often ahead of formal go-to-market efforts.
- •Why platforms eventually ‘market themselves’ once capability is there
- •Early targeted domains: molecular dynamics (e.g., NAMD) and seismic processing
- •How research centers surfaced ‘beyond reach’ problems (quantum chemistry/physics)
- •GeForce GPUs in Hinton’s lab: gaming hardware becomes deep learning infrastructure
- •Scaling via interconnect and parallelism as a durable GPU advantage
- 11:35 – 15:23
The 2012 inflection: ImageNet/AlexNet and outreach from Ng, Hinton, and LeCun
Jensen pinpoints 2012 as the moment NVIDIA recognized deep learning as a major wave, driven by simultaneous signals from multiple leading labs. ImageNet served as the ‘big bang’ that validated neural networks’ scalability and impact.
- •Andrew Ng’s request to port neural nets from thousands of CPUs to a few GPUs
- •Parallel interest from Geoff Hinton and Yann LeCun’s labs
- •ImageNet/AlexNet as a field-wide attention catalyst
- •Early belief forming: deep learning could scale and generalize beyond vision
- 15:23 – 20:15
AI as a platform shift: deep learning becomes a new way to write software
Jensen describes how NVIDIA moved from seeing deep learning as a vision algorithm to recognizing it as a new computing paradigm. The emergence of transformers and later systems (RLHF, retrieval, guardrails) culminates in natural language becoming a programming interface for computers.
- •From ‘better vision’ to ‘new software creation model’
- •Rapid model evolution: CNNs, RNNs/LSTMs, transformers, BERT
- •Why transformers matter: parallel training and scalable sequence learning
- •ChatGPT moment: RLHF + retrieval + dialogue/guardrails complete the experience
- •Programming disruption: human language as the new ‘coding language’
- 20:15 – 25:04
NVIDIA’s strategy up the stack: go as far as developers need (and no further)
Jensen explains NVIDIA as a computing platform company that climbs the software stack to enable real users, which can mean libraries, algorithms, or even domain-specific tooling. He emphasizes helping industries build proprietary and domain-specific foundation models rather than becoming a general AI model company.
- •Redefining ‘developer’ across eras: drivers → solvers → foundation-model users
- •Why algorithms must be redesigned to match GPU/data-center architecture
- •Domain libraries as leverage: cuDNN for deep learning, RTX for ray tracing
- •When NVIDIA might build foundation models (robotics, 3D/virtual worlds)
- •Principle: ‘as little as possible, as much as necessary’ to empower customers
- 25:04 – 29:40
Long-term conviction under public-company pressure: funding the future while staying sustainable
Jensen argues that investing in long-term bets (like CUDA) is compatible with present-day sustainability if the organization develops operational skill. He frames profitability and efficient execution as learnable skills in service of pursuing the company’s core beliefs.
- •Long-term bets aren’t inherently in conflict with near-term performance
- •Running efficiently and making money are skills, not just conviction
- •The institution’s job: pursue core beliefs and build what customers buy
- •Accelerated computing as a 30-year foundational conviction
- •‘Barely possible’ problems are where NVIDIA expects to be called in
- 29:40 – 35:09
CEO self-doubt and organizational design: why not every company should look like the US military
Prompted about management and CEO fit, Jensen shares an unconventional org philosophy: avoid generic hierarchies and tailor structure to the company’s function and leadership style. He describes NVIDIA’s unusually wide executive span and discipline around a single architecture/instruction set.
- •Regularly gut-check whether you’re right for the CEO role
- •Critique of traditional org charts mirroring military hierarchy
- •NVIDIA’s model: ~40 direct reports, minimal one-on-ones/career coaching at exec level
- •Discipline where it matters: ‘you can only afford one’ core computer architecture
- •Org should be bespoke to the company’s purpose and delivery model
- 35:09 – 36:50
Two operating modes: perfecting complex systems vs skunkworks exploration
Jensen outlines NVIDIA’s dual ‘motions’: one organization focused on building extremely complex computers reliably, and another designed for experimentation and pivots. These modes operate side-by-side to balance refinement with invention.
- •Refinement organization: build complicated systems ‘perfectly’
- •Skunkworks: invent 10-years-out bets with frequent adaptation
- •Resource mobility: stop what isn’t working and redeploy quickly
- •Need to run both systems simultaneously to sustain innovation at scale
- 36:50 – 38:44
Hopper (H100) technical breakthroughs: FP8 quantization and the Transformer Engine
Jensen highlights Hopper’s major architectural advances: embracing lower-precision formats to massively increase AI supercomputer throughput, and adding specialized transformer acceleration. He frames much of the rest (size, speed, memory, interconnect) as powerful but more ‘brute force’ evolution.
- •Quantization insight: many AI workloads don’t need 64-bit FP
- •FP8 as a step-change lever for performance and efficiency
- •Transformer Engine: pipeline shaped for training/inference of transformers
- •Scaling via fastest memories and connecting many chips efficiently
- 38:44 – 42:33
What’s next: robotics, video understanding, multimodality, and generative science
Looking forward, Jensen points to robotic foundation models, autonomous driving progress, and video as a key next training modality for learning physical structure and articulation. He also discusses excitement around diffusion-adjacent generative approaches and applying generation to proteins, chemicals, and other domains.
- •Robotics foundation model as a likely breakthrough within ~5–10 years
- •Video as the next major unstructured dataset to learn physical structure
- •Trajectory from GANs → style transfer/VAEs → diffusion model family
- •Generative models expanding from images/3D to proteins and chemicals
- •Multimodal learning as a major driver of future capability
- 42:33 – 48:02
Advice for founders: ignorance, resilience, and balancing conviction with agility
Jensen gives candid founder guidance: startups are uniquely painful, and ‘ignorance’ can be a necessary superpower to begin. He stresses resilience (forgetting pain) and the paradoxical mindset of holding strong conviction while believing you might be wrong—enabling learning and pivots.
- •Building companies is hard enough that doing it twice is irrational (but common)
- •Ignorance as a founder advantage—don’t wait too long to start
- •Balance: conviction without stubbornness; agility to learn and adapt
- •Resilience through ‘forgetting’ setbacks and moving to the next challenge
- •Anecdote: crypto demand swings and the pain of missing by billions
- 48:02 – 50:57
10–20 year hopes: AI for healthcare/drug discovery and climate modeling (Earth-2, Clara)
Jensen describes a planning technique—stand in the future and look back—to decide what to do now. He highlights two impact areas NVIDIA aims to advance: computational drug discovery/biology and climate science via multiphysics foundation models to dramatically reduce simulation costs.
- •Future-back planning as a decision method
- •Drug discovery complexity: massive combinatorial search spaces
- •Learning the ‘language’ and meaning of proteins/chemicals via AI
- •Climate multiphysics foundation models to make long-horizon earth predictions feasible
- •NVIDIA initiatives: Earth-2 (climate) and Clara (healthcare)
- 50:57 – 52:11
Closing fun: the origin of Jensen’s leather jackets
The episode ends with a light audience question about Jensen’s signature leather jackets. He credits his wife and family for finding them, joking that many are too fashion-forward unless you’re truly cool enough to wear them.
- •Jensen doesn’t manage the jacket sourcing—his wife does
- •Family ‘hunts’ for jackets; many are too bold for regular wear
- •Playful wrap-up and thanks to close the conversation
