CHAPTERS
Behind the scenes & why this Jensen interview matters
Ben and David set the stage: after hundreds of hours researching NVIDIA, they travel to NVIDIA HQ to interview founder/CEO Jensen Huang. They frame the central tension—NVIDIA’s extraordinary AI-era position versus the uncertainty of whether dominance and the AI wave will persist.
- •Hosts’ context: deep prior research and on-site interview at NVIDIA
- •NVIDIA’s scale and strategic position at the time of recording
- •Core question: sustainability of NVIDIA’s lead as AI becomes a trillion-dollar wave
- •Promise of stories: survival, pivots, and founder lessons
Riva 128: the near-death bet that reset NVIDIA (1997)
Jensen recounts NVIDIA’s do-or-die moment shipping Riva 128 with ~6 months of cash left. The chip shipped with missing DirectX blend modes, but the company decided to embrace DirectX, build the fastest possible product, and force success through execution under extreme constraints.
- •NV1/NV2 architectural dead-end (curves/forward texture mapping) and incompatibility with DirectX
- •Riva 128 as a full company reset amid intense competition and low morale
- •Strategy: build the biggest/fastest chip, use fastest memory, and target enthusiasts willing to pay
- •Reality check: only 8/32 blend modes worked; NVIDIA had to market and manage around constraints
One-shot tapeout: emulation, simulation, and pulling risk forward
Jensen explains how NVIDIA used hardware emulation to virtually prototype and run the software stack before silicon arrived—because a normal iterate-and-fix cycle would have killed the company. The lesson he draws isn’t ‘bet the company,’ but ‘prefetch the future’ by simulating and de-risking as much as possible before committing.
- •Bought an emulator (Icos) to run drivers and software before physical prototypes
- •Extreme constraint: only one tapeout possible before running out of money
- •Mindset: if you must bet, make it knowable by moving risky work earlier
- •General principle: simulate/prefetch everything you can to turn conviction into informed commitment
CUDA’s roots: from graphics abstraction to general-purpose parallel compute
Rather than a sudden leap, Jensen frames CUDA as a continuation of earlier efforts (e.g., Cg) to create higher-level abstractions over programmable GPUs. NVIDIA tested early non-graphics workloads and recognized GPUs as uniquely parallel, massively threaded processors—creating a plausible path to general-purpose computing before AI demand exploded.
- •CUDA preceded by Cg and early experiments in GPU abstraction layers
- •Early use cases: CT reconstruction, imaging, and scientific compute
- •GPU characteristics: highly parallel and massively threaded processing
- •Long-horizon platform investment (thousands of person-years) before mainstream ML
Post-AlexNet reasoning: why deep learning would scale and matter everywhere
After AlexNet’s breakthrough, Jensen describes returning to first principles: deep learning looked like a scalable ‘universal function approximator’ and potentially a teachable universal computer. This reframed NVIDIA’s opportunity from niche acceleration to a foundational shift in how software would be built across industries.
- •AlexNet as a discontinuity forcing NVIDIA to ask “why did this leapfrog decades?”
- •Inference: deep learning scales with more data, depth, and compute
- •Shift from causality to predictability for many valuable real-world problems
- •Conclusion: software and computing architectures would change around ML
Working the research frontier: researchers as the early feedback loop
NVIDIA leaned on its CUDA-era relationships with universities and scientists to find and support the earliest deep learning users. Jensen describes engaging directly with leading researchers (Hinton, LeCun, Ng, Goodfellow, Sutskever) and watching progress accelerate from quarterly papers to daily breakthroughs.
- •CUDA’s existing footprint in scientific domains made researcher outreach natural
- •NVIDIA actively sought “every AI researcher on the planet” to help them advance
- •Early GANs and other prototypes looked toy-like but appeared obviously scalable
- •Observation: research cadence and capability growth became visibly exponential
OpenAI’s founding and DGX: supplying the early ‘AI supercomputer’
Jensen explains he wasn’t a founder of OpenAI but knew many of the people involved and understood their need for next-generation compute. He delivered an early DGX system to OpenAI, reflecting NVIDIA’s strategy of building purpose-built AI computing systems and placing them with frontier labs.
- •Not involved in founding OpenAI, but connected to the key people
- •OpenAI needed the kind of compute NVIDIA was building (DGX lineage)
- •Jensen personally delivered the first DGX to OpenAI
- •NVIDIA’s early success: aligning systems + software stacks to researcher needs
Language models: self-supervision, scaling, and emergent reasoning
Jensen reflects on the elegance of masked/next-token prediction (BERT-style self-supervised learning) and why scaling would unlock more capability. He argues that because text encodes reasoning and common sense, large-scale compression of language can naturally yield emergent reasoning behaviors—still miraculous even if explainable.
- •Self-supervised objective (mask/predict) as a clever way to learn from vast text
- •Scaling intuition: more parameters/data should encode more knowledge and structure
- •Text contains implicit reasoning; models can learn reasoning-like behavior from it
- •‘Sensible but still amazing’: understanding doesn’t remove the wonder
NVIDIA’s leadership architecture: 40+ direct reports and ‘mission is the boss’
Jensen describes NVIDIA as organized like a computing stack rather than a traditional hierarchy. Information is shared broadly and quickly, authority comes from reasoning and enabling others, and cross-functional teams assemble around concrete missions (e.g., “build Hopper”), producing a neural-network-like organization.
- •Org model: computing stack of modules/layers, not military command-and-control
- •‘Mission is the boss’: assemble best people across the company for a specific deliverable
- •Trade-off: high pressure on leaders because power doesn’t come from privileged info
- •Meetings include new grads and execs receiving the same information simultaneously
High-velocity product cadence: learn broadly, don’t imitate blindly
Asked about NVIDIA’s impressive shipping cycle, Jensen emphasizes learning from many sources—competitors, adjacent industries, and business books—without copying them. The goal is to translate lessons into strategies suited to NVIDIA’s environment, capabilities, and objectives.
- •Business books and competitive intel are for inspiration, not replication
- •Core practice: ask “what does this mean to us, right now?”
- •Execution speed as a performance dimension (time-to-market matters)
- •Continuous learning culture as an input to strategy, not a substitute for it
Journey to the data center: separating compute from the viewing device
Jensen traces NVIDIA’s data center move back ~17 years to an insight: being tethered to monitors would cap growth. Early work on streaming frame buffers foreshadowed cloud gaming (GeForce Now) and remote graphics, ultimately evolving into data center supercomputing and positioning NVIDIA for the AI boom.
- •Key constraint: GPUs plugged into desktop PCs next to monitors limits TAM
- •Early prototype: capture frame buffer, encode, and stream—compute/viewing separation
- •GeForce Now as NVIDIA’s first data center product; remote graphics as the second
- •Strategic rule: pave the way early so you’re ‘near the tree’ when the apple falls
Mellanox acquisition: why networking became the AI scaling bottleneck
Jensen explains that being a true data center company requires more than processors—it requires networking and infrastructure. He contrasts hyperscale’s many-users-per-machine model with AI training’s one-job-across-many-processors model, making high-performance interconnects (InfiniBand) essential and validating Mellanox as a pivotal acquisition.
- •Data centers are defined by networking/infrastructure, not just CPUs/GPUs
- •AI training is distributed computing at massive scale (inverse of classic hyperscale)
- •Commodity Ethernet works for many workloads but not for tightly-coupled training at scale
- •Mellanox brought world-class HPC networking talent and became a top strategic decision
Founder strategy lessons: zero-billion-dollar markets, platforms, and moats-as-networks
Jensen advises companies to position in markets that don’t exist yet and to build platforms by enabling ecosystems, not by defending castles. He argues NVIDIA has always been developer-oriented, with architectural compatibility (UDA/CUDA lineage) as an ‘unnegotiable’ rule enabling a massive installed base and durable platform effects.
- •Early-stage enemy: non-consumption; later: design for emerging needs before others arrive
- •‘Zero-billion-dollar markets’: invest a decade early to earn asymmetry when markets emerge
- •Platforms form by enabling ecosystems; moats emerge as networks of developers/customers
- •UDA→CUDA continuity and strict compatibility across generations as the foundation of NVIDIA’s platform
AI’s societal impact: safety, human-in-the-loop, and jobs via productivity-led prosperity
Jensen outlines multiple dimensions of AI safety—physical safety in robotics/autonomy, information safety (bias, truth, creators’ rights), and governance via human-in-the-loop workflows. On jobs, he argues productivity tends to create prosperity and expansion, leading to more hiring overall, though roles will shift and individuals can be displaced.
- •AI safety domains: robotics/automotive functional safety and info safety (bias/misinformation)
- •Preference for human-in-the-loop and controlled model improvement (avoid wild self-modification)
- •Jobs thesis: prosperity from productivity leads companies to pursue more ideas and hire
- •Advice: learn to use AI to avoid being displaced by other humans augmented by AI
Lightning round & personal founder reflections: time, fear, support systems, and ‘don’t start a company’
In rapid-fire questions, Jensen shares personal tastes (Star Trek, cars, favorite business authors) and reveals deeper founder psychology: there’s ‘plenty of time’ with correct prioritization, and his enduring fear is letting employees down. He closes with a striking view: if he knew how hard it truly was, he wouldn’t start a company again—highlighting the necessity of support systems and the ‘How hard can it be?’ entrepreneur’s self-trick.
- •Personal notes: Star Trek fandom, daily driver (Mercedes EQS), business authors (Christensen, Grove)
- •Core fear: letting employees down; leadership as responsibility to others’ hopes and careers
- •Perspective shift: ‘there’s plenty of time’ if you prioritize (don’t let Outlook run you)
- •Entrepreneurship is harder than anyone expects; success requires unwavering support from team, family, and long-term backers
