CHAPTERS
- 0:00 – 2:59
Behind-the-scenes teaser and why this Jensen interview matters now
Ben and David set the stage: after years of research on NVIDIA, they traveled to HQ to interview Jensen Huang at a pivotal moment for the company and the AI boom. They frame the core question: will AI be a trillion-dollar wave and can NVIDIA sustain its dominance?
- •Acquired hosts reflect on NVIDIA’s production setup and the “insanely cool” interview
- •Context: NVIDIA at ~$1.1T market cap and central to the AI explosion
- •Big open questions about durability of NVIDIA’s lead and AI’s long-term magnitude
- •What to expect: memory lane, near-death experiences, founder advice, emotional reflections
- 2:59 – 10:50
Riva 128: the bet-the-company chip, built in simulation under extreme constraints
Jensen recounts NVIDIA’s 1997 crisis: six months of cash, a one-shot tape-out, and a decision to simulate everything rather than iterate in silicon. The launch wasn’t perfect (8/32 DirectX blend modes), but the company made a sequence of “extraordinarily good decisions” when it mattered most.
- •NV1/NV2 missteps and the architectural reset required by DirectX
- •Desperation-driven decision-making: one tape-out, go straight to production
- •Use of an emulator to virtually prototype hardware and run the software stack early
- •“If it’s not perfect, we’ll be out of business” as an organizing principle
- •Market positioning: build the biggest/fastest chip for performance-hungry enthusiasts
- 10:50 – 12:07
What ‘betting the farm’ really means: pulling risk forward through simulation
The hosts probe whether Jensen’s lesson is simply to make bold, company-defining bets. Jensen reframes it: you only push chips in after you’ve de-risked via exhaustive pre-work—simulate, prefetch the future, and validate as much as possible before committing.
- •Bold bets aren’t faith-based; they’re preceded by intense validation and preparation
- •Key idea: “pull risky things from the future into the present” via simulation
- •Why traditional iterate/tape-out/fix loops were impossible for early NVIDIA
- •How this mindset persists in NVIDIA’s culture and execution today
- 12:07 – 13:29
From CG to CUDA: building a general-purpose abstraction layer above the GPU
Jensen explains CUDA’s roots in earlier developer-facing work like Cg and programmable shaders. He describes why GPUs’ massively parallel, threaded nature suggested broader computing potential well before ML demand exploded.
- •Cg as a precursor: higher-level expression over graphics hardware
- •Early non-graphics use cases: CT reconstruction, imaging, scientific workloads
- •Programmable shaders as “processors” with unique parallel characteristics
- •CUDA as a long-horizon platform investment that preceded mainstream demand
- 13:29 – 18:22
Post-AlexNet reasoning: why deep learning was scalable and economically huge
After AlexNet, Jensen goes back to first principles: deep learning looked like a universal function approximator—and potentially a teachable universal computer. He connects the technical breakthrough to massive commercial applicability where prediction often matters more than causality.
- •Deep learning’s leap over decades of CV work prompted a ‘why does it work?’ reset
- •Scalability logic: more data + deeper models → better performance
- •Prediction vs causality: many industries just need outcomes, not explanations
- •Implications: most software could eventually be ‘programmed’ via ML methods
- 18:22 – 22:21
Researchers, conferences, and the early AI flywheel (plus OpenAI’s first DGX)
Jensen describes how NVIDIA leaned into universities and the AI research community as early adopters of CUDA and deep learning infrastructure. He recounts delivering an early DGX system to OpenAI, aligning NVIDIA’s systems roadmap with frontier research needs.
- •CUDA’s research footprint across sciences created a natural distribution channel
- •Direct engagement with key figures (LeCun, Ng, Hinton) and early conferences
- •Tracking rapid paper progress as evidence of exponential improvement
- •OpenAI: Jensen wasn’t a founder but supported with early DGX delivery
- 22:21 – 24:56
Language models: why scaling works, emergent reasoning, and ‘still a miracle’
The conversation shifts to large language models and the surprise of scale. Jensen highlights the elegance of self-supervised learning (mask/predict), explains why reasoning can be learned from text, and emphasizes that even understandable systems remain astonishing.
- •BERT as a key moment: clever self-supervision and a clear scaling path
- •LLMs as information compression/encoding that can capture reasoning patterns
- •Emergent capabilities as a natural consequence of learned reasoning
- •Engineering awe: even fully understood systems can feel miraculous
- 24:56 – 27:17
Sponsor break: Statsig and the experimentation layer behind ML deployment
Ben and David explain how experimentation platforms helped companies like Google and Facebook deploy and improve ML models through A/B testing. They present Statsig as an accessible, best-in-class experimentation and analytics platform used by modern AI teams.
- •Experimentation de-risks releases and speeds iteration for ML-driven products
- •A/B testing as the deployment bridge between research and production outcomes
- •Statsig’s offering: experimentation, feature flags, product analytics
- •Adoption by AI companies including OpenAI and Anthropic
- 27:17 – 34:16
NVIDIA’s org design: 40+ direct reports, ‘mission is the boss,’ and information equality
Jensen describes an unconventional organization built like a computing stack rather than a command-and-control hierarchy. With fast information dissemination and cross-cutting missions, leadership is earned by reasoning and enabling others—not by controlling information.
- •Org chart mirrors product architecture: layers/modules like a computing stack
- •‘Mission is the boss’: teams wire up like a neural network to accomplish concrete goals
- •High pressure on leaders: fewer power advantages from privileged information
- •Meetings include new grads and execs learning decisions simultaneously
- 34:16 – 39:31
Journey to the data center: cloud gaming as the wedge and ‘separating compute from viewing’
Jensen traces NVIDIA’s data center evolution back ~17 years, starting with the insight that GPUs tethered to monitors capped growth. GeForce Now and remote graphics pioneered decoupling compute from the endpoint, laying groundwork for today’s AI-centric data center dominance.
- •Strategic constraint: limited desktop GPU attach rate and physical proximity to displays
- •Core insight: separate computing from viewing to expand market opportunity
- •GeForce Now as NVIDIA’s first “cloud/data center” product and long latency battle
- •Progression: remote graphics → CUDA + GPU as supercomputer → modern data center stack
- 39:31 – 43:42
Mellanox acquisition: networking as the real data center differentiator for AI scale-out
Jensen explains why Mellanox was essential: data centers are defined by infrastructure and networking, not just processors. AI training flips hyperscale’s paradigm—one job distributed across huge clusters—making high-performance networking (e.g., InfiniBand) critical.
- •Data centers differ from desktops primarily through networking/infrastructure
- •AI training = distributed computing across massive clusters (inverse of hyperscale virtualization)
- •Ethernet is fine for many workloads; AI scale-out needs HPC-class networking
- •Mellanox: strategic fit, proven HPC collaboration, and strong Israel team
- 43:42 – 45:30
Sponsor break: Crusoe and clean-energy AI cloud built around H100 clusters
The hosts introduce Crusoe as an AI-focused cloud provider partnered with NVIDIA, emphasizing H100 availability, InfiniBand interconnects, and performance-per-dollar advantages. They highlight Crusoe’s model of powering AI workloads with stranded/wasted or clean energy to reduce emissions.
- •AI GPU scarcity and Crusoe’s early ability to offer H100s at scale
- •Optimized AI cloud stack: H100 clusters + 3,200Gb InfiniBand + block storage
- •Clean/stranded energy as a cost and environmental advantage
- •Operational impact example: methane capture and CO2 emissions avoidance
- 45:30 – 54:44
Company-building strategy: ‘zero billion dollar markets,’ moats as networks, and platform DNA
Jensen’s advice centers on positioning early in markets that don’t exist yet, building ecosystems so competitors can’t easily swarm later. He reframes “moats” as networks, then explains NVIDIA’s platform orientation going back to UDA/DirectNV and developer relationships from day one.
- •Competing by serving needs before they emerge (the ‘zero billion dollar market’ approach)
- •Sustaining lead via ecosystem/platform formation—moat as a network of developers/customers
- •NVIDIA platform roots: UDA → CUDA, and early developer-relations-first strategy
- •Compatibility as an unbreakable rule: every NVIDIA chip runs CUDA; long-term architectural continuity
- 54:44 – 59:55
Luck, skill, and near-death experiences: Riva lessons and the Carmack/Sweeney inflection
Jensen reflects on how pivotal moments mix execution excellence and external luck. He credits Riva 128 for shaping lasting chip-development practices, and notes that gaming’s acceleration “killer apps” (Quake/Unreal) helped tip the market in NVIDIA’s favor.
- •Riva 128 as a template for modern chip execution: do it once, do it right
- •Luck factors: market timing and developer adoption of hardware acceleration
- •John Carmack’s shift to OpenGL acceleration for Quake as a major catalyst
- •Tim Sweeney/Unreal as another early accelerant for consumer 3D
- 59:55 – 1:08:32
AI safety, job displacement, and why productivity tends to create more work (not less)
The discussion moves to AI’s societal effects: safety in robotics and information integrity, and the economics of automation. Jensen argues productivity usually leads to prosperity and expansion, creating more jobs overall—even if it shifts which individuals hold them.
- •Safety domains: functional safety (robots/cars) and information safety (bias, rights)
- •Preference for human-in-the-loop and controlled retraining/validation cycles
- •Automation thesis: people lose jobs to other people using AI more than to AI itself
- •Historical analogy: productivity increases ambition and expands industries
- 1:08:32 – 1:24:51
Lightning round and personal reflections: time, fear of letting employees down, and founder pain
Jensen answers quick personal questions (Star Trek, cars, business books, Don Valentine) before turning introspective. He shares fears about letting employees down, describes NVIDIA as his last job, and explains why he wouldn’t start a company again knowing the emotional toll—underscoring the importance of unwavering support systems.
- •Personal picks: Star Trek, Mercedes EQS, Christensen and Grove, Don Valentine story
- •Advice to younger self: prioritize—don’t let Outlook control your life
- •Core fear: letting employees down; responsibility to those who bet their lives on the mission
- •Founder reality: entrepreneurship is far harder than expected; needs deep support from family/colleagues/investors
- •Public-market drawdowns and conviction required to persist through volatility
- 1:24:51 – 1:30:00
Closing thoughts: NVIDIA’s opportunity scale and Acquired outro
Jensen frames NVIDIA’s evolution from ‘chip company’ to ‘manufacturing intelligence,’ expanding the addressable market into the trillions. Ben and David wrap with community calls-to-action and credits to sponsors.
- •Opportunity sizing: AI as ‘manufacturing of intelligent work’ with trillion-dollar scope
- •Reframing value: systems/services can dwarf per-device chip economics
- •Hosts’ closing: email list, ACQ2, LP program, Slack, merch
- •Sponsor thanks and episode sign-off
