The counter point is they could make a higher level version of CUDA which wouldn't necessitate all the other supporting libraries. The draw of cuBLAS is that CUDA is a confusing pain. It seems reasonable to think they could write a better, higher level language (in the same vein as triton) and not have to write as many support libraries

100% valid - Nvidia is trying to address that now with cuTile and the new Python front-end for CUTLASS.