In 1995, Jensen Huang flew to Japan to tell one of his company’s biggest clients that NVIDIA could not build the chip it had promised. Then he asked Sega’s CEO to send the $5 million payment anyway. Shoichiro Irimajiri, the Sega president on the other side of the table, said yes — not because the chip would work, but because he trusted the team and wanted to see them survive.
That confession kept NVIDIA alive. Today the company is worth roughly $5 trillion, and Huang is telling a new kind of truth: the physical AI business — robots, autonomous vehicles, systems that move and manipulate the world — is already generating about $10 billion a year, and he expects it to reach $100 billion. Speaking on the Y Combinator Startup Podcast, the NVIDIA CEO laid out a vision for the next decade that connects the company’s near-death founding to the architecture of tomorrow’s AI economy.
The $5 Million Confession
NVIDIA was founded in 1993 on a premise that turned out to be exactly wrong. The company’s original 3D graphics algorithm was fundamentally flawed, and by 1995 roughly 35 to 40 competitors were chasing the same market with similar designs. Huang did not pretend otherwise. He went to Fry’s Electronics, bought three OpenGL textbooks, and handed them to his engineers.
“Everybody would have thought that NVIDIA started out as world leaders in 3D graphics, and we learned it from a textbook,” Huang recalled. “The big lesson is that technology is changing all the time, 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.”
The Sega episode was the ultimate test of that philosophy. NVIDIA had contracted to supply the graphics chip for the Dreamcast console. Huang knew his team could not deliver. Rather than bluff, he flew to Japan and told Irimajiri they would fail — but still needed the cash. Sega paid. When NVIDIA went public in 1999 at a $300 million valuation, Sega sold its stake for $15 million.
The lesson Huang took from that moment was not about luck. It was about intellectual honesty. Being wrong about the core technology at founding is survivable if you acknowledge it and reinvest in learning. That pattern — confront reality, decompose hard problems, build infrastructure to learn faster — became the operating system for everything that followed.
The specifics of that reinvention were more brutal than the anecdote suggests. NVIDIA’s first chip, the NV1, drew 3D scenes using quadratic texture mapping — curved surfaces — while the rest of the industry standardized on triangles. When Microsoft committed DirectX exclusively to triangle primitives, the approach became a dead end, and the NV2 part being built for the Dreamcast was cancelled around 1996. NVIDIA laid off roughly half of its 100 employees and aimed everything that was left at a triangle-based replacement. That chip, the RIVA 128, shipped in August 1997 with the company down to a single month of payroll — the origin of what is still quoted internally as NVIDIA’s unofficial motto: “Our company is thirty days from going out of business.”
Why NVIDIA Has to Live in the Future
Designing a new computing platform takes roughly three years, ramping production takes another two, and customers expect to use the systems for a decade. That means Huang has to place bets on workloads that do not yet exist.
“We kind of have to live in the future five to ten years because it takes three or so years just to build a system, takes a couple years to ramp it up, and you would like them to be able to use the computer for 10 years after,” he explained.
This is not abstract strategy. The decision that transformed NVIDIA from a graphics card company into the dominant force in artificial intelligence was made roughly 15 years ago, when Huang recognized that a breakthrough in image recognition called AlexNet was not just a better classifier — it was something far more fundamental. Deep learning, he realized, was a “universal function approximator,” a method that could learn virtually any function given enough data.
“And so 15 years ago I was telling everybody that, hey guess what, we just learned the universal function approximator,” Huang said. “We can give it the answer for almost any function and it could learn what the function is.”
That insight triggered an immediate reorientation of NVIDIA’s entire stack — processor design, middleware, algorithms, applications. The company bet that accelerated computing would reorganize entire industries, and it spent the next decade and a half building the infrastructure for a revolution that most people did not see coming until ChatGPT arrived in late 2022.
The timeline explains why a $10,000 investment in NVIDIA shares made a decade ago was worth roughly $1.52 million as of July 2026, according to an analysis by The Motley Fool — a total return of roughly 15,090%. It also explains why the stock currently trades at a forward price-to-earnings ratio of about 22 to 23, below the Nasdaq-100 average of roughly 25, even as the company projects year-over-year revenue growth of approximately 95% in the current quarter. The market has grown accustomed to NVIDIA outperforming, but the company’s next act is already being built.
MetricFigureMarket capitalizationAbout $5 trillionPhysical AI revenue, annualizedAbout $10 billionPhysical AI target$100 billion, in 3 to 10 yearsEstimated GPU market shareAbout 85%Forward price-to-earnings ratioAbout 22 to 23 timesNasdaq-100 average forward P/EAbout 25 timesProjected revenue growth, current quarterAbout 95% year over yearValue of $10,000 invested ten years agoAbout $1.52 million as of July 2026Total ten-year returnAbout 15,090%Big-tech AI infrastructure spending, 2026 estimateAbout $730 billionCapex-to-semiconductor-index 30-day correlation+0.78 in April 2026, near zero by mid-year
Figures cited by Huang on the Y Combinator Startup Podcast, except the ten-year return calculation (The Motley Fool) and the 2026 infrastructure spending estimate (Charles Schwab).
The Linux Moment for AI Agents
Huang sees agentic AI — software that can reason, plan, use tools, and act autonomously — as the next operating system. When he first encountered OpenClaw, an open-source framework for building AI agents, his reaction was immediate.
“OpenClaw to me was a very Linux moment,” Huang said. “And now everybody can build their own AI.”
The comparison is deliberate. Just as Linux democratized access to operating system infrastructure and enabled an entire generation of software innovation, open-source agent frameworks lower the barrier for every company to construct domain-specific AI systems. NVIDIA responded by offering its entire engineering workforce to support the OpenClaw project. The company also pointed to Hermes, an open-weight model, as another building block that lets organizations create custom AI without starting from scratch.
“If not for open source, how would we have modern AI?” Huang asked, framing the question as rhetorical. The argument is that proprietary stacks would have concentrated AI development inside a handful of companies, while open platforms let the entire industry experiment in parallel — exactly the pattern that let NVIDIA pivot from a failed graphics algorithm to market leadership three decades ago.
The precedent is measurable. Linux, released by Linus Torvalds in 1991, has run every one of the world’s 500 fastest supercomputers since November 2017, according to the TOP500 project — a clean sweep no proprietary operating system has ever managed. That is the scale of adoption Huang is invoking when he calls an agent framework a Linux moment.
Internally, NVIDIA is running multiple AI coding agents — Claude Code, Codex, Cursor, and Cognition — in sandboxed environments, encouraging what Huang described as a “thousand flowers bloom” approach. The goal is not to pick a single winner. It is to learn from every tool available and accelerate the company’s own development velocity.
Here is how the components of an agentic AI system fit together:
The Controllability Problem
For all the excitement around agents, Huang identified a specific technical bottleneck that he believes is holding the entire field back: controllability.
Current prompting is too coarse. You give an agent a text instruction and hope it produces something useful. Huang envisions a world where humans and agents collaborate at a much finer grain — adjusting one parameter in a plan, modifying one pixel in a generated image, then letting the agent regenerate the rest around that edit. He does not need agents to be perfect.
“Controllability is probably the single biggest breakthrough that we need for agents at every single level,” Huang said. “We don’t need the agents to be a hundred percent accurate; it could be 80% and we help it the rest of the way.”
This framing is a departure from the race toward fully autonomous systems that dominates much of Silicon Valley’s AI discourse. Huang is arguing that the practical path to reliable agents runs through better human interfaces, not through chasing the last 20% of accuracy that might take years of research to achieve. If you can steer an agent precisely when it veers off course, 80% reliability becomes a usable tool rather than a frustrating toy.
The implication for software engineering is profound. Huang predicted that “most software is going to be done agentically anyhow,” which means the most valuable skill in the future will not be writing code line by line. It will be systems thinking — the ability to design architectures, organize components, and think abstractly about how information flows through complex systems.
“This is probably going to be the most useful skill in the future,” Huang said. “Systems understanding, systems awareness, system design, system organization, but systems thinking.”
The $10 Billion Robot Business Nobody Talks About
NVIDIA’s physical AI business — encompassing autonomous vehicles, warehouse robots, agricultural machinery, and delivery systems — is already “really, really big,” Huang said, estimating it at roughly $10 billion in annual revenue. The foundation was laid with self-driving cars: Waymo, Tesla, and Mercedes all use NVIDIA chips, software, or both.
But the automotive market is only the beginning. NVIDIA recently open-sourced its entire autonomous driving stack under the name Alpamayo, a strategic decision that reveals how Huang thinks about market structure. The individual markets for agricultural robots, mail delivery vehicles, and warehouse autonomous mobile robots are each too small to justify the cost of a proprietary software stack. Combined, however, they represent a substantial opportunity — but only if the development cost is shared across them. By releasing Alpamayo as open-source software, NVIDIA makes it economically viable for every niche robotics application to build on a common platform, while the company sells the hardware and simulation tools underneath.
The “ChatGPT moment” for robotics, in Huang’s view, has already happened — arriving quietly around 2024 to 2025. Robots that can walk, manipulate objects, and perform reinforcement learning grounded in real physics became demonstrable. The remaining challenge is building the full training pipeline: simulation environments like Isaac Sim and Cosmos for learning and evaluation, followed by the notoriously difficult sim-to-real transfer.
“The moment I saw us generating video of a hand picking up a glass, why can’t I cause a robot to do the same?” Huang recalled. “The ChatGPT moment of robots happened a couple of years ago already.”
The $100 billion target is not a random round number. It is an extrapolation from where the physical AI business stands today, growing inside a global industrial automation market that measures in the trillions. Huang frames the timeline as “3 to 10 years,” acknowledging both the pace of technological progress and the inertia of physical-world deployment.
Why AI Won’t Steal Your Job
Huang directly challenged what he called the “exactly backwards” narrative that artificial intelligence will destroy employment at scale. His argument rests on a distinction between automating a task and automating a job.
“The narrative about AI destroying jobs is exactly backwards,” Huang said. “AI eliminates tasks. AI automates tasks away, but it doesn’t necessarily eliminate jobs.”
He offered three data points to support the claim. In radiology, the task of reading scans has been increasingly automated for years. Yet the number of radiology jobs has increased roughly 20% over the last several years. The mechanism: faster scan reading reduces backlogs, which lets hospitals process more patients, which requires more radiologists and supporting nurses. The productivity gain expands demand faster than it contracts labor.
In software engineering, the task of writing code is being automated at an accelerating pace. But the number of software engineering jobs has grown approximately 10% year-over-year. Automation lowers the cost of implementation, which means more ideas get attempted, which means more engineers are needed to design, test, and integrate those systems. Paralegals, Huang added, are “growing like crazy” for the same reason — law firms that can handle more cases per paralegal take on more cases.
FieldTask being automatedEmployment trend Huang citedWhy headcount still risesRadiologyReading scansUp about 20% in recent yearsFaster reads clear backlogs, hospitals process more patients, more radiologists and nurses are neededSoftware engineeringWriting codeUp about 10% year over yearCheaper implementation means more ideas get attempted, and more engineers to design, test and integrate themParalegal workCase and document handlingGrowing sharply, per HuangFirms that handle more cases per paralegal take on more cases
Employment figures as cited by Huang on the podcast; he did not attach a specific number to the paralegal trend.
The jobs data Huang cited lines up against enormous spending numbers that suggest the AI buildout has years to run. Alphabet, Microsoft, Amazon, and Meta are projected to invest roughly $730 billion in AI infrastructure in 2026, according to research from Charles Schwab, with analysts already penciling in further increases for 2027. That spending flows directly into the semiconductor supply chain where NVIDIA occupies a commanding position — an estimated 85% share of the GPU market.
The relationship between hyperscaler spending and chipmaker revenue has recently decoupled, however. The 30-day correlation between big tech capital expenditure commitments and the Philadelphia Semiconductor Index plunged from roughly +0.78 in April 2026 to near zero by mid-year. Investors are beginning to distinguish between companies that book revenue immediately when orders arrive — like NVIDIA — and companies placing bets that may take years to generate returns.
The CEO Who Refuses to Step Back
Huang has been in what Silicon Valley now calls “founder mode” for 34 years, and he shows no sign of delegating the deep technical work. He reads research papers directly, talks to principal scientists, and insists on maintaining what he described as a “tactile sensation” of where the technology frontier actually sits.
“Part of having to be near the ground and be in the weeds is because oftentimes the technology is complicated or it’s changing really fast,” Huang said. “Unless you have a tactile sensation of what is actually happening, it could either feel like it’s just moving way too fast to understand, but if you understand the first principles of it over time then everything kind of makes sense.”
He compared the CEO to a Formula 1 driver who builds the car — the company — to fit their own driving style, not to some abstract management ideal. The organization can be constantly tweaked. This is not a conventional view of how to run a five-trillion-dollar enterprise, but Huang argues there is no evidence the approach breaks at scale. He has been doing it for more than three decades.
His advice to founders distills to a single heuristic: when you face an impossibly hard problem, tell yourself “how hard can it be?” and start working on it that morning.
“You want to imagine in your head, ‘How hard can it be?’ And then you get going on working on it,” Huang said. “You don’t have to overcome life in one day, you just have to overcome that morning, that morning. If you stick with it long enough, NVIDIA happens.”
The non-obvious advice he offered to aspiring founders was to invest in the hard sciences — physics, chemistry, biology, computer engineering, systems thinking. “The simple stuff is going to get automated away,” he said. “Those hard problems will never go away.”
The next act for NVIDIA is already being assembled inside those hard problems: agentic software that anyone can customize, robots that learn in simulation before touching the physical world, and a computing platform designed for workloads that will not become mainstream for another five to ten years. The company’s history suggests that being right about the destination matters more than being right about every step of the journey. Huang’s $5 million confession to Sega proved that admitting you are wrong, and learning fast enough to become right, is a strategy that scales.