Wouldn't the exact same considerations apply to AI training/inference shops, seeing as gigawatts are usually the key constraint?