Nvidia's $10 Trillion Run Was an Accident. Jensen Just Didn't Flinch.
For 20 years Nvidia sold graphics cards to gamers. Then researchers noticed CUDA could train neural nets — and Jensen bet the company on a market that didn't exist yet.
In 2012, Nvidia was a $10B gaming-hardware company. In 2026, it crossed $10T in market cap — a 1,000x run in 14 years. Every business school will teach this as visionary strategy. It wasn't. It was one stubborn bet, made a decade before it paid off, that Jensen Huang simply refused to reverse when the market told him he was wrong.
The bet was CUDA. Launched in 2006, CUDA let developers use GPUs for general-purpose computing — physics simulations, financial modelling, protein folding. For six years, nobody cared. Wall Street begged Nvidia to kill it. It burned roughly $500M a year in R&D with no revenue attached, and gaming margins were subsidising a science project.
Then in 2012, a University of Toronto team used two Nvidia GTX 580s to train AlexNet — the neural network that broke ImageNet and kicked off the modern AI era. The researchers didn't pick Nvidia because it was best. They picked it because CUDA was the only mature way to program a GPU. Every AI lab that followed inherited that lock-in.
The lesson isn't 'invest in AI'. It's what Jensen did between 2006 and 2012: he kept funding a platform with no customers because he believed parallel compute was the future, even when quarterly earnings called for cuts. Most CEOs would have killed CUDA by year three. He didn't.
The takeaway for operators: the moats that matter are built during the years when nobody is watching. By the time the market rewards you, it's too late for competitors to catch up — TSMC can build fabs, AMD can ship silicon, but nobody can rewrite 18 years of CUDA libraries, tooling, and developer muscle memory overnight.
Nvidia didn't win AI. Nvidia won the boring infrastructure fight a decade before AI mattered. That's the actual playbook.