NVIDIA

NVIDIA designs GPUs. For twenty years that meant graphics cards for gamers. Around 2012 researchers discovered those same chips were ideal for training neural networks, and NVIDIA leaned in hard: it built CUDA, the software layer that lets programmers use its GPUs for any kind of math, and it became the default for AI.

Today its data center chips (H100, B200 and successors) are what OpenAI, Anthropic, Google and Meta use to train and run models. Demand is so high that access to NVIDIA GPUs is a geopolitical topic and a line item in national budgets.

NVIDIA does not manufacture the chips itself: it designs them and TSMC in Taiwan fabricates them. Its moat is less the silicon than the software ecosystem everyone already wrote on top of CUDA.