The significance of CXL is not to replace GPU memory, but to turn memory into a schedulable resource—thereby opening up a second growth curve for AI infrastructure.

The key is to decouple Prefill and Decode, separating the compute and memory bottlenecks, so as to maximize GPU utilization.
CXL 4.0 is an important leap in high-speed interconnect technology. It reduces latency to the nanosecond level, boosts data rates to 128 GT/s, and introduces the concept of native x2 width to support higher fan-out capabilities for platforms—thereby improving link distance and enabling longer signal transmission compared with CXL 3.0.