Tensorflow in production with AWS Lambda

I'm curious: How much difference would having a GPU make in this situation?

I also wonder how many people think of saving TensorFlow state to disk?