“Nikesh Arora believes that developing physical AI will be more difficult than developing generic frontier models because physical AI is a "depth" use case without a broad consumer application to drive training data.”