Reducing the Cost of Inference by a Tenth