“Based on analysis of constraints like energy and GPU production, there are likely only three to four more orders of magnitude of compute scaling possible for training AI models.”