“Future AI models and their underlying systems will achieve 20x to 50x lower latency than what is available today.”