The biggest irony in AI right now is that $NVDA basically enabled the AI boom, and now vibecoding in diminishing the moat around CUDA.
@semidoped had an interview with $META's head of ad infra who talked about LLMs rewriting their kernels so that the same models run across GPUs, ASICs, etc. The main advantage of cuda was that it was so much easier to optimize your model to Nvidia's hardware.
Still think cuda is excellent and has a deeper base, but writing on the wall is in? Jensen mostly talks about TCO these days as their moat.