> Capital expenditures also have their own utility curve

This a function of the use-cases. For simple use cases like document summarization, or speech-to-text, we might already be there. There are already applications where using something like gpt-3-turbo or a local-llm vs gpt-4 make more economic sense. The marginal improvement in reasoning or extra knowledge don't provide benefits relative to the inference costs. And you can see this playing out with Mistral, Anthropic and Google also providing models at different cost/complexity tradeoffs.

But .... if we get _new_ use cases, emergent use cases, from bigger models, there will be continued value in chasing that frontier. Reasoning is still poor in many areas. The ability to learn outside of application specific memory patterns that store state is largely unexplored. We are still far from AGI and there is a ton of value to be minded between current systems and a superhuman capabilities