“Among AI modalities, text and audio generation are relatively inexpensive, whereas video and world models are very expensive to both train and serve.”