Nvidia started out making graphics chips for games, and almost died early on. Around 2006 Huang pushed CUDA, a way to use those chips for general computation. For years it cost money and Wall Street didn't see the point.
Then deep learning turned out to need exactly that kind of parallel computing. Researchers were already using Nvidia cards, and the software was already there.
He's run the company for over thirty years with a very flat structure and a lot of direct reports. He talks openly about how hard and painful the early years were.