So, HIP at a raw level is as performant as CUDA. The real problems come from higher level stack (BLAS, LAPACK libraries for example). But not all software need higher level stack. So, then it becomes a cost benefit analysis.

A 15k AMD part vs a 60k nvidia part. For 100 Nvidia GPUs, you can buy 200 AMD GPUs and at least 2-3 engineers for 3 years at 300k to fix the specific library for that GPU. If you can make that work for a lower level library right now, then it makes to sustain it in future.

rocBLAS/hipBLAS are pretty solid. You're on the money with AMD's implementations of LAPACK not being up to snuff, though.

HIP is a thing. A lot of existing CUDA code can be translated with minimal effort. And machines get better all the time at guiding that rewrite.

I’m skeptical (without evidence). Companies that have never been good at software can’t just suddenly be good at software overnight.

Amd did several acquihires in recent past also , and now amd have more money than has ever had and stated that ai software is one top priority of the company since 1y .. results are surfacing..

Near constant releases of ROCm.