Anyone care to have a go at back of the envelope number for training energy use amortized per query for ChatGPT's models? Is the training or the inference going to dominate?