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