“In the near term, a key metric for measuring the efficiency of AI models is joules per token.”