When explaining why AI is a high friction industry with a few gatekeepers, we often focus on the high-level side of things (only a fee companies have both enough data and computing power to train large models etc.).
And we often forget the absolute monopoly that rules anything close to the silicon.
>95% of the ML models today are trained on Nvidia GPUs. And Nvidia GPUs have an absolute monopoly over both the hardware (the GPU itself) and the middleware (CUDA) in the form of proprietary tech without a single competitor.