Training chat AI models demands high interconnect GPU clusters for parallel computation, while inference servers pr