“Some of my hedge fund friends were talking about this exact topic. And the logic goes something like this. New capacity is going to come online slower than we previously thought. This disadvantages neoclouds. This theoretically advantages the existing hyperscalers who have already built tons of data centers. … If there's less compute to go around and less inference to go around, then they don't have to spend as much money.”

This is actually why. So some of my hedge fund friends were talking about this exact topic. And the logic that they were using goes something like this. New capacity is going to come online slower than we previously thought, right? This disadvantages neoclouds. This theoretically advantages the existing hyperscalers who have already built tons of data centers, whether they fully retrofit that for new AI use cases or not. In some ways, it directly benefits the competitive dynamic of the frontier companies in the aggregate. Because if there's less compute to go around and less inference to go around, then they don't have to spend as much money. Their commitments are lower. If it affects everybody equally, it's a good thing. It's almost like the pacing that they were asking for, but in a fair, more uniformly distributed way. And so theoretically, it like reduces their like CapEx burden and it allows them to like continue delivering their models, but without having to spend the bajillion dollars that they've committed.