The first GPU

On 31 August 1999 NVIDIA announced the GeForce 256 as "the world's first GPU", a term the company coined for a chip that handled the whole graphics pipeline, transforming and lighting 10 million polygons a second, without the CPU's help. It was built for Quake. Its architecture, many identical units doing the same arithmetic to different pixels and vertices at once, was the single-instruction, multiple-data parallelism that Danny Hillis had proposed for the Connection Machine, shrunk onto a consumer chip that sold for $300.

Nobody involved was thinking about neural networks, but a neural network is mostly matrix multiplication, and matrix multiplication is exactly what a chip for lighting polygons is good at. By the mid-2000s researchers were writing shader programs to run learning algorithms on graphics cards; CUDA in 2007 made that respectable; and in 2012 two GeForce GTX 580s trained AlexNet.

The consequence is that the compute for the deep-learning era was paid for by gamers. Moore's law had delivered transistors; the games industry delivered the architecture that could use them for learning, and NVIDIA, which in 1999 was one graphics-card maker among several, became in 2025 the first company worth four trillion dollars.