Nvidia

The GeForce 256 was marketed in 1999 as the world's first GPU. The genuinely decisive move came in 2007 with CUDA, which let developers run general parallel computation on the graphics card rather than only graphics. The investment looked unjustified for a long time.

It paid off in 2012, when AlexNet was trained on Nvidia cards and showed that deep learning is, in practice, large-scale matrix multiplication — precisely what a GPU does. Today the CUDA ecosystem is the company's strongest defence against competition, more so than the silicon itself. It designs but does not manufacture; production sits with TSMC.