A low-power chip that runs AI models using analog rather than digital computation shows comparable accuracy on speech-recognition tasks but is more than 14 times as energy efficient.