I think the key AI improvement is that these chips double the memory bandwidth and this has been the main limit for AI performance recently, became LLMs do not fit into cache.