With a typical transformer and a GPU the batch size that saturates the compute is at least hundreds. Otherwise (including typical size of 1 for local inference) you're memory bound.