Given that models are only going to get larger, and the sheer amount of compute required, I think the endgame here is dedicated "inference boxes" that actual user-facing devices call into. There are already a couple of home appliances like these - NAS, home automation servers - which have some intersecting requirements (e.g. storage for NAS) - so maybe we just need to resurrect the "home server" category.
I agree, and if you want to have the opportunity to build such a product, then you need a computer whose specs today are what a home server would have in four years. If you want to build the future you have to live in the future. I'm proud to make stuff most people can't even run yet, because I know they'll be able to soon. That buys me time to polish their future and work out all the bugs too.
