Inference is run on shared hardware already, so they're not giving you the full bandwidth of the system by default. This most likely just allocates more resources to your request.

Could well be running on Google TPUs.

The models are running on Google TPUs.