The inference part requires training the models to keep up to date as far as I am aware(I am open to being convinced otherwise if you have data showing so). Everything I’ve read about it implies that within a few years of reality updating or even just software if we constrain it to models for coding, won’t be able to handle all of the updates within the context windows available.

Given that I don’t know how you(royal you, not you specifically) can constrain an analysis of inference profitability to just the literal running of inference and not include training costs. That’s where it all breaks down for profitability.