Jensen Huang: The Mindset That Built NVIDIA

Why does it matter? Because the most consequential company in AI was built on wrong assumptions and a trip to Fry's Electronics.

NVIDIA raised money to build technology that was completely wrong, bought textbooks at a big-box electronics store to fix it, and ended up inventing modern computer graphics and the infrastructure for the entire AI revolution. Jensen Huang's conversation at YC Startup School 2026 isn't about GPU hardware — it's about the specific mental models that compound over decades when applied consistently.

• NVIDIA's founding algorithm was entirely wrong — Jensen fixed it with three OpenGL textbooks from a Fry's Electronics, handed them to the engineers, and that was that • Jensen saw AlexNet in 2012 not as an image classifier but as a universal function approximator — a reframing he made 15 years before the rest of the market understood it • AI is growing radiology jobs by 20% and software engineering jobs by 10%; the narrative that AI destroys employment is factually backwards • Controllability — not model accuracy — is the actual bottleneck for agents, and the breakthrough nobody is prioritizing

NVIDIA raised money, then bought textbooks to learn the technology it was supposed to already have

By 1995, two years into building 3D graphics chips, Jensen realized the algorithm NVIDIA was founded on was completely wrong — and none of the engineers knew how to do it the right way either. With 35–40 competitors already in the market, he drove to Fry's Electronics, bought three textbooks on OpenGL pipeline design, and handed them to the team.

"We actually started the company, raised money, and bought textbooks when you think about it."

The company that is now synonymous with GPU supremacy learned computer graphics from a bookstore run. What Jensen draws from this isn't about grit — it's about a very specific willingness: "Technology is changing all the time. And so long as you're able to confront the reality, so long as you are able to learn, the technology itself actually doesn't matter." Companies fail because they can't say the founding idea is wrong, out loud, while there's still time. Jensen said it immediately. That gap — between acknowledging failure and rationalizing it — is most of what separates the companies that make it.

Jensen saw AlexNet as a universal function approximator in 2012 — a reframing that explains NVIDIA's AI dominance 15 years later

Everyone else looked at AlexNet and saw a better image classifier. Jensen, whose lens was always scanning for important algorithmic domains to accelerate, saw something foundational.

"The breakthrough for us was realizing that AlexNet was not AlexNet, that AlexNet was an approach with deep learning that allows you to learn any function."

For most important problems in the world, the underlying function is imprecise and unknowable analytically. A universal function approximator doesn't just improve image recognition — it opens up that entire class of problem. "15 years ago, I was telling everybody that we just discovered the universal function approximator."

From that single realization, NVIDIA immediately began working on computer vision, robotics, and self-driving cars — not because those markets were trending, but because the algorithm was general enough to reach all of them. The question Jensen asked wasn't "what does this do?" — it was "what does this do to the computing stack? What happens to software? What industries does this touch?"

First-principles reasoning about a new technology, applied 15 years early, built a position that compounded in silence until the rest of the world caught up.

Radiology jobs up 20%. Software engineering jobs up 10%. The dominant story about AI and employment is factually wrong.

"The narrative about AI destroying jobs is exactly backwards."

Jensen isn't speculating — he's pointing at labor market data. AI has automated the core task of reading radiology scans, and radiology employment rose 20%, because automating the bottleneck revealed the enormous patient backlog underneath it. Law firms equipped with Harvey can take on more cases, so paralegal roles are growing "like crazy." Software engineering jobs are up 10% year over year despite AI generating code.

The mechanism is the same across all three: automate the constraining task, and unmet demand floods in faster than jobs disappear. "The backlog of ambition and aspiration is so high that if we can automate away the task of programming, we could hire more software engineers to do more things. We could be more ambitious."

Jobs exist because there's a purpose behind them. Automate some of the tasks; the purpose and the backlog remain. Founders building products premised on AI-displaces-workers anxiety are misreading the actual opportunity — which lives in the backlog, not the shrinkage.

The real unlock for agents isn't accuracy — it's being able to change one word in a plan and get a precisely different output

An agent running at 80% quality is already deployable. Jensen's argument isn't that the quality bar is low — it's that the missing ingredient isn't capability at all.

"The single biggest breakthrough that we need for agents at every single level — controllability."

Right now, human intervention in agentic systems is blunt. Prompts and conditioning are coarse. Jensen's target: change a single word in a plan file and get a delta — one pixel different, one triangle, one connection in a CAD design. Not a completely different output. A surgically specific one.

"We don't need the agents to be 100% accurate in order for us to use it. It could literally be 80%. And then we help it the rest of the way."

That collaboration model — human directing with precision, agent executing — changes the economics of deployment entirely. The gap between agent demos and real production is mostly a controllability gap, not a model quality gap. Build the fine-grained override mechanisms first.

NVIDIA is an algorithm-acceleration company — the chip has always been the means, never the strategy

"It's not about building a great chip. It's about accelerating an algorithm domain."

This is how NVIDIA kept winning across completely different eras — graphics, scientific computing, deep learning — without changing what it fundamentally is. The founding insight was that CPUs can be augmented with accelerators to solve otherwise intractable problems. Which algorithm they applied that insight to varied. The identity didn't.

Jensen describes his lens as always scanning for important algorithmic domains: OpenGL, VASP, SQL, deep learning. When AlexNet appeared, he didn't pivot — he ran the same question he'd been asking for 20 years: what is the algorithm, why does it matter, and what happens when we scale it? "Our realization is everything to do with algorithm, not the chip, turns out to be exactly right."

Define your company by the class of problem you solve, not the technology you use to solve it. Technology changes across decades; a clear problem-space identity lets you adapt and expand without losing direction.

Jensen told Sega he couldn't deliver on a $12M contract — and then asked to keep the money anyway. Sega said yes.

The technology didn't work. Jensen flew to Japan, sat down with Sega CEO Irma Jirisan, explained that NVIDIA couldn't build what they'd been contracted to build, advised Sega to hire someone else — and then said he'd still need the money.

"What you're telling me is, what I contract you to do, you can't do, but you would like all the money on the contract. And I said, you got it. That's exactly right."

Sega gave him $5 million. It kept NVIDIA alive long enough to discover the right approach. Jensen's read on why it worked: Irma Jirisan had trusted Jensen enough to write the contract in the first place. Radical honesty about failure — no spin, no softening — deepened that trust rather than destroying it. "You don't invest in companies, you invest in people."

When things go catastrophically wrong with a key partner, the instinct is to manage the narrative. Jensen argues the opposite move is what actually keeps you solvent.

Systems thinking is what AI cannot replace — and its value is already compounding

Coding, synthesis, low-level implementation — all of it is being automated. Jensen is direct: "Most of the low-level things that have to be done are going to be done agentically anyway. So you have to be much more able to think abstractly about systems."

What remains is the ability to reason about constraints, bottlenecks, and information flow: Where is the input? Where is the output? Is the limit memory, processor, or network? What breaks when volume scales by 10x? At NVIDIA, chip designers are already mostly systems designers — the transistor-level work is synthesized. Software is heading the same direction.

"The better you are at systems thinking, so that you could orchestrate millions of agents solving problems autonomously, the better off you are."

The hard sciences and hard engineering disciplines don't disappear — they move to the center. Learn to reason about constraints and architecture. That abstract layer is what AI cannot yet replace, and the premium on it is only growing.

'How hard can it be?' isn't optimism — it's a deliberate trick to keep anxiety from stopping you before the work begins

Jensen's founding psychology comes down to two interlocking pieces. First: learning is the primary superpower. If it's important and anyone else has figured it out, he can too. Second: a very specific way of approaching difficulty.

"You want to imagine in your head, how hard can it be? And let the suffering come to you a little bit at a time. Don't imagine how hard it's going to be and let all of that turn into anxiety."

It always turns out to be much harder than expected. But accurately forecasting the difficulty before starting doesn't reduce the difficulty — it converts future suffering into present paralysis. The phrase is not a prediction. It's a permission slip.

Paired with it: a compressed time horizon. "You don't have to overcome life in one day. You just have to overcome that morning."

The framework doesn't make hard things easy. It keeps you from stopping before the hard things begin — which is the only prerequisite for everything else.

The AI era rewards people who reframe the algorithm first — not people who implement it fastest

The throughline in everything Jensen described: the transition from "who builds it best" to "who reasons about it most clearly" is already well underway. Systems thinking, first-principles analysis, and the willingness to confront a wrong assumption the moment it surfaces — these compound across decades. Implementation skill gets commoditized in years.

NVIDIA is what happens when you apply that reasoning discipline consistently across 30 years and multiple technological eras. The founders who last in the AI era won't be the ones who coded fastest. They'll be the ones who saw what the algorithm actually was — and what it implied — long before the market did.

---

Topics: NVIDIA, Jensen Huang, founder mindset, AI strategy, agentic AI, accelerated computing, robotics, physical AI, entrepreneurship, systems thinking, AI and jobs, open source AI, YC Startup School