More generally any GPU computation, not necessarily graphics. The above argument could go that there is no point in high level languages for CPUs either and everyone should just always use assembly, which is obviously false. Same can go for GPUs.
It's just that gpus have no general purpose usage .. for ai, it's basically all about perf.
Of course some cross platform stuff can be made for modest acceleration, but the state of the art is always going to be out of reach.