Huang, Jensen
▶ Watch on YouTubeCore ideas
Huang frames NVIDIA's current era as "extreme co-design": because AI workloads no longer fit on one computer, NVIDIA must jointly engineer GPUs, CPUs, memory, networking, power, and cooling as a single system, since Amdahl's law means accelerating only one component (say, compute) yields diminishing returns if everything else becomes the bottleneck. This is reflected organizationally — his direct staff is ~60 people, deliberately too large for one-on-ones, because problems are attacked collectively in the room ("no conversation is ever one person") rather than delegated down a hierarchy. He traces NVIDIA's strategic arc as a narrow path between being a highly specialized accelerator (deep but narrow market) and a general-purpose computing company (broad but unspecialized) — pixel shaders, then IEEE-compliant FP32, then Cg, then CUDA were each incremental steps widening that aperture without sacrificing acceleration.
He tells the CUDA-on-GeForce decision as NVIDIA's defining existential bet: adding CUDA to consumer GPUs raised costs ~50% and crushed gross margins (market cap fell to ~$1.5B) with no immediate consumer payoff, but was necessary because a computing architecture lives or dies by install base, not elegance (he contrasts inelegant-but-dominant x86 against beautifully designed but failed RISC architectures). That decade-long bet — millions of GeForce cards seeding CUDA into universities and labs — became the substrate the deep learning revolution was built on, and Huang argues install base plus execution velocity plus developer trust remains NVIDIA's core moat today, ahead of its second moat: one architecture spanning clouds, cars, robots, satellites, and edge devices. On leadership, he describes a deliberate practice of "manifesting" strategic shifts (CUDA, deep learning, Mellanox) by pre-seeding belief across the company and industry for months or years before any formal announcement, so that by the time a decision is declared, it feels inevitable rather than surprising — the same practice he applies externally to persuade memory makers (HBM, repurposed low-power phone memory) and supply-chain partners to make multi-billion-dollar bets years ahead of demand.
He lays out four scaling laws (pre-training, post-training, test-time/inference, and now agentic scaling — spawning teams of sub-agents), arguing each supposed blocker (data scarcity, "inference is easy," compute limits) has been repeatedly wrong because thinking/reasoning is fundamentally harder and more compute-intensive than memorization. On power, his proposed fix isn't just efficiency gains (NVIDIA has scaled compute ~1,000,000x in a decade versus Moore's Law's ~100x) but restructuring contracts so data centers use the grid's normally-idle excess capacity (~40% headroom held for rare peak/emergency conditions) and gracefully degrade rather than demanding "six nines" uptime. He credits Elon Musk's Colossus build (200,000 GPUs in four months) to relentless first-principles questioning ("is this necessary, does it have to take this long") and personal presence at the point of action. On AGI and jobs, he claims "I think we've achieved AGI" by a functional definition, argues (using radiology's growth despite superhuman computer vision since ~2020 as his central case study) that automation historically expands rather than eliminates professions because a job's purpose and its tasks are related but not identical, and separates "intelligence" (which he calls an increasingly commoditized, functional capability — he's "surrounded by 60 people smarter than me" and still orchestrates them) from "humanity" (character, compassion, resilience), which he argues AI will make more valuable, not obsolete.
Excerpts
On why NVIDIA bet the company on CUDA despite near-term financial devastation:
“A computing platform is all about developers... a developer, like anybody else, wants to develop software that reaches a lot of people. So the install base is, in fact, the single most important part of an architecture... no architecture has ever attracted more criticism than x86... yet it is the defining architecture of today." The bet cost so much that "our market cap went down to like one and a half billion dollars," but "NVIDIA is the house that GeForce built, because it was GeForce that took CUDA out to everybody.”
On why he doesn't do succession planning, reframing mortality as an argument for constant knowledge transfer:
“If you're worried about succession planning... the most important thing you should do today, if you care about the future of your company, post you, is to pass on knowledge, information, insight, skills, experience as often and continuously as you can... I'm constantly passing knowledge, empowering people, elevating the capability of everybody around me, so that the outcome that I seek... is that I die on the job.”
On separating "intelligence" from "humanity" as a response to AI-driven job anxiety:
“Intelligence is a functional thing. Humanity is not specified functionally. It's a much, much bigger word... I'm surrounded by intelligent people more intelligent than I am in each one of the spaces that they're in... and somehow I'm sitting in the middle orchestrating all 60 of them... don't let this democratization of intelligence, this commoditization of intelligence, cause you anxiety. You should be inspired by that.”