“The theoretical minimum for chip-to-chip latency is the speed of light, which is around 2-3 nanoseconds, compared to the approximately 4,000 nanoseconds on current NVIDIA products.”