In production on millions of boxes and the payoff is a 25% reduction in machines needed for some inference workloads ...