In 2006, Jensen Huang announced CUDA — a programming framework that let developers write general-purpose software for NVIDIA’s graphics chips.
The analysts were puzzled. NVIDIA made GPUs for gamers and game developers. Why would a gaming chip company invest hundreds of millions building a software platform for… scientists?
For five years, almost nobody outside of a few physics departments used it. Jensen kept funding it anyway.
Then came September 30, 2012.
Alex Krizhevsky and Geoffrey Hinton submitted AlexNet to the ImageNet Large Scale Visual Recognition Challenge. Their neural network won by a margin of 10 percentage points — an unheard-of gap in a field where improvements were measured in fractions of a percent.
It ran on two NVIDIA GTX 580 GPUs using CUDA. The deep learning era had started, and NVIDIA was the only company with the infrastructure already in place to serve it.
That was the moment. Everything after is just math.
Today NVIDIA supplies roughly 95% of the chips used to train frontier AI models. H100 GPUs sell for $30,000–$40,000 each. At peak demand, delivery times stretched to 52 weeks. Hyperscalers were flying executives to Taiwan to personally negotiate allocation.
NVIDIA’s market cap hit $3.3 trillion in 2024 — surpassing Saudi Aramco, Apple briefly, and the entire GDP of France. It is the most important company in the world that most people couldn’t have named five years ago.
The lesson isn’t just about chips.
Jensen held the CUDA bet through years with no return. The bet wasn’t “GPUs will be used for AI.” The bet was: parallel computation will become the limiting resource for the next wave of computing, and we’re going to be the only ones ready.
Being early and being wrong look identical for a long time. The only way to tell them apart is to still be there when the world catches up.
Jensen was still there.