I do wonder how much of their inference is running on 10-year old TPUs. I’d imagine not much, if only because 10 years ago the justification for massive compute was quite different than today.