Model inference compute outweights training compute by 10:1 and more for frontier LLMs. "LLM depreciation" is an expense, but not a dealbreaker.