“There is a "compute overhang," with the global supply of compute power far exceeding what is currently used for training AI models.”