The frontier of how good models are also shifts and will remain ahead of local models unless we hit some dead end limitation in the algorithms themselves. A ceiling so to speak on how good LLM can get before the law of diminishing returns starts to apply.

1. Is it cheaper for me to buy hardware and electricity than to call an API? (doesn't seem like it right now)

2. The best models are still worth it, unclear when this changes

3. Average person doesn't have the skill to do this. They are afraid to run even simpler things

Young people have had even the concepts of filesystems conditioned out for files to live in a 'folder' of an APP.

Local sovereignty isn't a pressing need for most users.

I do believe this is gonna get commodatized like the internet has. Hardware obviously keeps getting better and cheaper as the time goes by. Sofware in this case is already free/open-weights.

The moats these companies might end up having in near future:

1. Government and enterprise contracts;

2. Even better private models not released to public and only accessible through long-term/exclusive contracts;

3. Gatekeeping the access to millions of their users, especially the non-technical ones, and charging premium for the same;

4. Becoming more and more as the full-stack OS'es to build on top of them.. By proving ready-made foundational layers like knowledge, memory, search/research, sandboxes, deployments, etc...

5. Data/network effects from large-scale usage and feedback loops.

...

They won't survive. AI-as-a-service for frontier models will be relegated to military and research – if that. We're already at diminishing returns on model improvements. Latest improvements are on surrounding architecture, harnesses, agent systems, etc. Consumer hardware will be running the equivalent of ChatGPT 5.2 and IMO most interaction with personal computing devices will be done via natural language LLM personal assistants.

Maybe it takes a bit longer than 5 years but that's where we're going. Already the only reason you're not interacting via personal assistant for everything isn't really LLM capability but the lack of tooling.