The hyperscaler companies have acknowledged among themselves that they probably shouldn't all be spending this much money and don't need five large language models, but each is terrified of being left out and losing competitive position to competitors, creating a prisoners' dilemma dynamic that forces overinvestment.