I hope Intel, as the underdog, can do mid-tier cards, but with lot more VRAM. I can see myself using multiple LLM models at once, not necessarily running at once, but interleaved.
No one's catching up to the topline perf or even power efficiency (Intel actually suck at perf/power right now) but they might be able to compete with raw GPU costs at least. Give us a $2k card with 256gb ram and we'll probably take it over the $40k 80gb cards despite the power efficiency and topline perf that the $40k card has.