I think the scenario being described may have a lot of validity: a lot of folks, for a wide variety of reasons, will want to have their own local processing power for inference models rather than cloud LLM inference. And that it appears that the memory and processing power and power efficiency of doing that will be in reach of a very large population. If not quite now, then very soon. I for one will switch to that configuration as soon as I can get a local model that will do local image analysis and tagging without undue hassle. I've already looked into it, but the hassle hurdle is high for now.

I'm not as confident as the author is that the current generation of devices unambiguously shows Apple to be in the lead in this regard and that they will maintain that lead. Maybe, but that's a much bigger leap.

I think a standalone box that is both NAS for my data and compute for my LLM(s) would suit me fine--I don't immediately see why it would have to be integrated with my PC as long as it's suitably integrated with my data. In that scenario, anybody with a source of appropriate chips could make the box. And, by extension, I don't see why the maker of that box would necessarily be getting a really great business with high barriers to competition. Maybe four years from now I get that box from Synology who makes a 4% margin on the things, and I get my LLM for it from Gitee. Who knows?

Jim