Compute constraints are a double bind: On the inference side you need to either (a) raise prices, (b) ration use, and/or (c) serve worse models. This hurts current growth On the training side, you can't train the next gen of models to stay competitive. This hurts future growth