In 2006, Nvidia shipped CUDA — software that let a graphics chip do general-purpose computation, well beyond games. The catch was that no market existed to buy it. Back then “AI” was a word academia had half given up on, and deep learning was stuck in a long winter. When a company that sold gaming chips poured serious R&D money into general-purpose computing software that earned nothing in the near term, the market called it a drag on margins.
Jensen Huang has a name for places like this: a “Zero Billion Dollar Market.” The market is worth nothing today, but he’s convinced the direction is right. This is where his strategy splits from everyone else’s. Most companies pile into a market that’s already large and fight over share. Huang does the opposite — he walks into a market that doesn’t exist yet and grows the market itself.
Six years with no market
Nvidia laid down CUDA and then had to wait six years.
Through those six years, Nvidia pushed CUDA hard into universities and research labs. It gave the software away free, built course material, and supported researchers writing papers on GPUs. Until then, doing general computation on a GPU meant a grotesque workaround: you had to disguise an ordinary computation as a “drawing problem” and trick the graphics API into running it. To run a physics simulation, you’d con the chip into thinking “this is a job about coloring pixels.” CUDA stripped away the disguise and let a developer hand the GPU real computation directly, in C.
The turning point came in 2012. A team at the University of Toronto trained a neural network (later known as AlexNet) on two consumer gaming graphics cards and won ImageNet — an image-recognition contest — by a crushing margin. In that moment everyone understood the same thing: training deep learning at any realistic scale needed GPUs, and the only chip that came with the software to actually drive those GPUs was Nvidia’s.
This is where Nvidia’s money machine reveals itself. CUDA was free, but a program written in CUDA ran only on Nvidia GPUs. Give the software away to lock in the ecosystem, then make your money selling the GPU chips that the software has to run on. The invisible toll laid down for free six years earlier only began collecting once the AI boom finally arrived.
Was it luck?
Stop here and you could call it luck. But the same thing recurs across Huang’s career, and that recurrence unsettles the easy reading.
In 1999, Nvidia launched the GeForce 256 and coined the term GPU — Graphics Processing Unit — itself. Until then a graphics chip was a helper part that assisted the CPU. Huang pulled transform and lighting calculations into the chip and declared, “This isn’t a helper chip; it’s a processor in its own right.” A single word reframed how the market saw the category. 3dfx, once the undisputed king of 3D graphics, collapsed the next year and was absorbed by Nvidia.
CUDA in 2006 was the next bet. And right now, Huang is making the same move a third time: physical AI — robots that move in the real world. The idea is to drag AI out of on-screen text and images and into a physical body. He sees humanoid robots and labor automation as an enormous future market, and he’s laying the groundwork in advance: a simulation platform (Omniverse) that trains robots millions of times in virtual space, and a general-purpose robot brain (GR00T) that drops onto any company’s robot. All of this while the robot market is still worth roughly zero.
All three share the same shape. Lay the infrastructure before the market exists, then wait for that market to arrive.
What made it possible
The first thing is betting on a direction. Huang didn’t foresee AI. He bet on the hypothesis that “parallel computing will matter someday,” and AI was one of several outcomes that arrived along that vector. He didn’t call the result; he fixed a direction and refused to be knocked off it.
The second is a structure built to outlast the wait. Huang has been the same company’s CEO for more than thirty years, since founding it in 1993 — statistically rare, and the foundation that makes the first thing possible. An investment like CUDA, which bears fruit only a decade later, can’t be pushed through by anyone but an owner who’s still accountable a decade out. A hired executive chasing quarterly numbers structurally can’t make that bet.
The third is a paranoid sense of crisis. Huang’s internal mantra is famously “Our company is always thirty days from going out of business.” Nvidia did, in fact, nearly fold several times early on. Instead of exhaling once the danger passed, he preserved that tension and mounted it as a permanent management tool.
Just don’t forget that these three are reverse-engineered from a success. Companies that placed the same bet on the same direction, only for the market never to arrive, don’t get to leave a story behind. We only see the bets that survived and turned out right. So this is less a formula — “do this and it works” — than a set of conditions the winning bets happened to share.
What zero really means
Laying something down in a zero-dollar market is, for now, pure cost. The cost of getting mocked, the cost of thinner margins, the time no one credits you for. But the moment the direction proves right and the market shows up, that cost turns into a toll no one can catch up to overnight.
So the biggest bets are often already settled before the market is even born. The number zero may not mean “worthless” — it may mean “no one has priced it yet.” The eye that tells those two apart is what let one man play the same move three times across thirty years.
— tomte
Read next · Is the Pick-and-Shovel Play Still Safe in the Age of AI: Where the Analogy Breaks