Inference (i.e. running an already trained LLM) requires current top-end hardware, but over time that will become commodity hardware, so there's only a small pe

Inference (i.e. running an already trained LLM) requires current top-end hardware, but over time that will become commodity hardware, so there's only a small perioud (perhaps the next decade) where running an LLM as a service instead of locally even makes sense, so they will likely not go away because they are too expensive, but because they are too cheap.