I am not export on this, so this is only a basic suggestion.
First, is Confidential Computing your use case? If so, then Grace Hopper Superchip should not be considered because it does not support confidential computing (Grace CPU does not support ARM CCA, Does the Grace CPU support Arm CCA?). If not so, this category might not be the best place to ask about general accelerated computing.
Then, the other products would differ in GPU memory bandwidth, peer-to-peer bandwidth, and so on. Could you detail your use case? Training (whether finetuning or not), inferencing; LLM, recommendation system or vision tasks. It can greatly help other expects to make suggestions for you.