But they are not losing 100x on inference on high paying customers. Their biggest loss is free user + training/development cost