There is significant redundancy of supply in AI data center capacity and chip production, similar to the redundancy in railroad track miles and fiber optic capacity in prior cycles, and because multiple hyperscalers are building duplicate LLM capabilities and infrastructure, they will eventually compete on price, which will compress margins below acceptable levels.