“In 2012, Google's speech team developed a model that outperformed humans but could not afford to deploy it due to high inference costs.”