It makes sense to ASIC-ify the thing to get lower latencies and make the whole thing cheaper, so MS can run GPT-(n+1) cheaper. But this bet only pays off if the LLM industry gets into the mature stage where costs dominate, not innovation.