That ignores the possibility that local inference gets good enough to run without a subscription on reasonably priced hardware.

I don't think that's too far away. Anthropic, OpenAI, etc. are pushing the idea that you need a subscription but if opensource tools get good enough they could easily become an expensive irrelivance.

There is that, but the way this usually works is that there is always a better closed service you have to pay for, and we see that with LLMs as well. Plus there is the fact that you currently need a very powerful machine to run these models at anywhere near the speed of the PaaS systems, and I'm not convinced we'll be able to do the Moore's law style jumps required to get that level of performance locally, not to mention the massive energy requirements, you can only go so small, and we are getting pretty close to the limit. Perhaps I'm wrong, but we don't see the jumps in processing power we used to see in the 80s and 90s, due to clock speed jumps, the clock speed of most CPUs has stayed pretty much the same for a long time. As LLMs are essentially probabilistic in nature, this does open up options not available to current deterministic CPU designs, so that might be an avenue which gets exploited to bring this to local development.

Local inference is already very good on open models if you have the hardware for it.

My concern is that inference hardware is becoming more and more specialized and datacenter-only. It won’t be possible any longer to just throw in a beefy GPU (in fact we’re already past that point).