Jensen Huang: The Mindset That Built NVIDIA

TL;DR

At Startup School 2026, Garry Tan interviewed Jensen Huang about the mindset behind NVIDIA's 33-year run from a failed 3D-graphics algorithm to the center of the AI economy. The through-lines: confront reality immediately when your technology is wrong, believe you can learn anything (NVIDIA relearned graphics from three textbooks bought at Fry's), and anchor a company on a unique, deeply-held perspective about the world rather than on any particular technology. Looking forward, Jensen argues the durable human skill is systems thinking — because low-level work is being automated by agents — and names fine-grained controllability as the single biggest breakthrough agents still need. He closes with the advice he'd telegraph to his younger self: resilience, one day at a time, and the attitude "how hard can it be?"

Key takeaways

• NVIDIA's founding technology was flat-out wrong. The 3D-graphics algorithm the company was built on failed by 1995, with 35–40 competitors already in the market — and nobody on the team knew the right way to do it.
• The fix was three textbooks. Jensen bought OpenGL texts at Fry's, handed them to his engineers, and NVIDIA rebuilt its graphics pipeline from them. His lesson: if you can confront reality and learn, the specific technology you started with doesn't matter.
• The idea that was right: accelerate algorithm domains, not chips. Augmenting CPUs with accelerators for whole classes of problems — graphics, molecular dynamics, image processing, and eventually deep learning — is the perspective the entire company is built on.
• Great companies come from a unique perspective you deeply believe in — more than from the technology or even the market — ideally one that is hard to pursue.
• The Sega story is a masterclass in honest failure. Jensen flew to Japan, told Sega's CEO the contracted technology didn't work, advised him to pick another vendor — and asked to keep the money anyway. The $5M kept NVIDIA alive; Sega's stake was worth $15M at the 1999 IPO.
• AlexNet wasn't AlexNet — it was a universal function approximator. Jensen's reframe 15 years ago: deep learning is a new way of doing software, which meant the entire industrial stack — processor, middleware, algorithms, applications — would be reinvented. NVIDIA started on vision, robotics, and self-driving almost immediately.
• Coarse recursive self-improvement already exists. Agents improve their own markdown files and long-term memory with every use. What's missing is fine-grained controllability — change one word in a plan and get a one-word delta, one pixel, one via, one CAD component — which Jensen calls the single biggest breakthrough agents need.
• Agents don't need to be 100% right to be useful. 80% plus a human finishing the job is already game-changing; controllability is what makes that collaboration work.
• AI automates tasks, not jobs. Coding is automated yet software-engineering jobs grew ~10% year over year; radiology reading is automated yet radiology jobs grew ~20%; paralegals are growing despite legal AI. Backlogs are so deep that productivity growth drives hiring.
• Physical AI is next, and it's already a ~$10B business for NVIDIA. The robotics "ChatGPT moment" happened a couple of years ago; self-driving was deliberately chosen as the first market big and standardized enough to spin the flywheel.

Lessons learned

Confront reality, then buy the textbook

The founding story is a template for handling being wrong. NVIDIA didn't rationalize its failing algorithm — Jensen stood in front of the company and said there would be no company unless they admitted it didn't work. Then came the part most founders skip: admitting nobody there knew the right way either, and going to learn it from scratch. "We actually started the company, raised money, and bought textbooks," as he tells it. The same posture — if it's important, we'll go learn it — is how NVIDIA has entered every domain since, always paired with the deliberately naive attitude of "how hard can it be?" (It's always harder. That's the point of the framing: keep anxiety from stopping you before you start.)

A company is a perspective, not a technology

The technology NVIDIA started with was wrong, but the perspective — general-purpose CPUs can be augmented by accelerators to solve otherwise-impossible problems, and the value is in accelerating an algorithm domain, not building a great chip — was exactly right, and it has absorbed graphics, HPC, deep learning, and now robotics. Jensen's generalization: what makes great companies is a high-level vision about the future of something important, a unique take on it you deeply believe, ideally one that's hard to pursue.

Fit the car to the driver

On organization: Jensen rejects adopting conventional management structure for its own sake. The founder is an F1 driver and the company is the car — you adapt the car to yourself, because the race is competitive and any change that doesn't fit you slows you down. When asked what happens after him: the next CEO reshapes the company to fit them. His day-to-day is constant tweaking of business processes so he can be more effective — founder mode, sustained for 34 years, zero to five trillion.

Stay in the weeds to read the waves

Jensen's first-principles habit isn't a management technique; it starts as personal curiosity, then becomes service — learn a domain deeply so you can break it down for the company and empower everyone else. In a fast-changing field, without a "tactile sensation" of what's actually happening, everything looks like chaos; with first principles, it looks like waves a surfer can read. It's also how NVIDIA lives five to ten years ahead: systems take ~3 years to design, years to ramp, and a decade in service, so you must reason from workloads and algorithms — agents, memory, sandboxes, MCP, Amdahl's-law bottlenecks — long before they're mainstream.

See the implication early, act a decade before it's obvious

A repeated NVIDIA pattern: observe a primitive, reason "if this, then what?", and commit early. Chain-of-thought out of Stanford → reasoning from prior knowledge → cars shouldn't need billions of miles → Alpamayo, a "thinking" self-driving car trained on a couple million miles. Video generation in their labs → if a model can generate a hand picking up a glass, it can drive a robot's articulation → physical AI, world foundation models, and the real-to-sim / simulate / sim-to-real training loop (Isaac Sim, Cosmos).

Tips & tricks

• When your technology is wrong, say so out loud and immediately — to your team and to your customers. The Sega negotiation worked because honesty made trust the asset when the product failed. People invest in people.
• Adopt "how hard can it be?" as an operating stance. Not because it won't be hard — it always is — but because imagining all the difficulty up front converts into anxiety and inaction. Let the suffering arrive a little at a time.
• Look at every breakthrough through the algorithm lens. Jensen saw the same AlexNet everyone else did; the differentiated move was asking what the algorithm is, why it works, what else it can do, and what it means for the stack if it scales.
• Practice systems thinking deliberately — inputs, outputs, information flow rates, constraints (processor, memory, networking). The low-level work is going agentic; the person who can specify and orchestrate systems, eventually millions of agents, is the one who stays valuable.
• Let a thousand flowers bloom on tooling. Inside NVIDIA, Claude Code runs autonomously in sandboxes company-wide alongside Codex, Cursor, and Cognition — people pick their tools, and the company learns from all of it.
• Use the cloud AIs, but build your own where your domain lives. Jensen's framing of OpenClaw as a "Linux moment": open-source harnesses (OpenClaw, Hermes, LangChain, DeepAgent) mean every company can and should build domain-specific AI — that's where the innovation and the alpha are.
• Don't dismiss markdown-file memory as trivial. The exchange with Garry landed on why agent memory compounds: as Jensen put it, "words are thoughts."

Notable moments

• The Fry's textbook run. With a couple hundred dollars in his pocket after realizing the company's algorithm was wrong, Jensen bought three OpenGL textbooks and handed them to the engineering team — the starting point of NVIDIA becoming the world leader in computer graphics.
• The $5M honesty ask. Telling Sega's CEO Irimajiri-san that NVIDIA couldn't deliver Dreamcast's graphics, recommending a competitor, and still asking for the contract money — and getting it, because trust survived the failure.
• "All of NVIDIA's engineers are your engineers." Jensen's message to OpenClaw's creator after seeing it as the operating system for LLMs, followed by the same offer to the Hermes team — plus a sandboxing toolkit when people worried the harness was unsafe.
• Jensen's first post on X, in 2026. Self-described introvert, he says he overcame the shyness only because open-weights models mattered too much: without Linux, Kubernetes, PyTorch and their lineage, modern AI wouldn't exist.
• The telegram to his 18-year-old self: you'll never stop not-knowing things — he barely knew how to answer investors' questions then and "barely" does now — so bet on your ability to learn, and get through today. Stick with it long enough and NVIDIA happens.

Caveats & advice for builders

Jensen's automation take is deliberately uneven: whole categories of cognitive tasks — including sitting down and writing code — go the way of long division, but the hard problems don't. His advice to students is to run at the hard sciences (physics, chemistry, biology, computer science and engineering) and especially the intersections between domains, because AI raises ambition rather than eliminating the work — a trillion-transistor chip stopped being remarkable because the scale of the task stopped mattering. The skills he'd still learn the old way: systems thinking above all, plus reading the intersection of technology with social issues and market gaps. And his macro claim for this room: the computer — "the single most important technology in human history" — has been completely reset, making this the best time in 60 years to start a company.