> In the future if LLMs are more efficient and eventually commoditized, needing less infraestructure to be run (think on-device as chip efficiency increases).

You speak as if this is given. Sure, LLMs are gonna get effective but frontier ones are always going to be hosted.

Also, on-device LLMs are gonna have a 'cutoff' for training data. You cant ask a gpt-oss4 about "Who won Arsenal x Athetico Madrid game". It has to go to the internet, and do the 'search'. You certainly cant ask the local model "What was Google's Q1 2026 earnings".

The self-hosted / on-device LLMs are going to do a lot, but not all and the moment 'search' is involved, people will reach for Google.

--

Lots of commercial queries need to be 'fresh' ("Build me an itinerary for Paris"). While a self-hosted LLM can do it, you might not want to trust it, because its info might be stale.

I encourage you to go beyond group think (HN is guilty of that) and really evaluate your position is actually valid or not.