Is training the bigger producer of CO2 or inference for the whole world using it? I suspect that in the long run, it is the latter.