but if a GPU is waiting for data 75 percent of the time because of the inefficiency of communication, reducing that compute delay by hiding latency and scheduling tricks like L3 caches do for CPU and GPU cores, then maybe, if DeepSeek can push that computational efficiency to ne...