really? epyc used a lot for host cpu to work with nvidia gpus, which is totally cuda, nobody uses rocm for ai.
All AMD has to do is offer lots of petaflops on the cheap, and software that technically works. Once the hardware is there, and research groups have won grants to work on it, it's up to grad students (who actually do the research work) to figure out how to get anything done on ROCm.
From H100 Nvidia tries to use its own host cpu, of course AMD now wants to have its own accelerator the MI300, soon they will each have their turnkey systems, it's just that AMD's software stack is still way behind.