CUDA

Researchers had been coaxing graphics cards into doing scientific arithmetic since the early 2000s by disguising their problems as textures and shaders. Ian Buck's Brook project at Stanford showed it could be made systematic, and NVIDIA hired him. CUDA, announced in November 2006 and released as version 1.0 on 23 June 2007, let a programmer write a function in C, mark it as a kernel, and launch it across the hundreds of cores of a GeForce card as thousands of parallel threads. The GPU became a computer.

Its importance for this timeline is hard to overstate. A neural network's forward and backward passes are matrix multiplications, and a GPU does those tens of times faster than a CPU of the same price. Rajat Raina, Anand Madhavan and Andrew Ng showed in 2009 that CUDA trained deep belief networks seventy times faster; Alex Krizhevsky wrote AlexNet's training code directly in CUDA in 2012. Every framework since, TensorFlow, PyTorch, JAX, generates CUDA kernels under the hood.

CUDA is also the moat. Fifteen years of libraries, drivers and habits meant that when demand for AI compute exploded in 2023, NVIDIA had the only hardware most people could use, and the company's value rose past four trillion dollars in 2025.