“Increasingly specialized hardware will enable much lower latency and more capable AI models at affordable prices.”