There's limited global manufacturing capacity. AI companies have been buying up all they can get to store their datasets on. They're not fast enough for the models to run off but they still need them to store the training data which is in the thousands of petabytes.