Co-packaged optics (CPO) mounts optical modules right next to the processor die instead of using pluggable transceivers at the edge of a server. As AI clusters scale to tens of thousands of GPUs, this cuts the power and latency lost in electrical signaling and becomes a key lever for how far a rack-scale AI system like AMD's Helios or Nvidia's NVL72 can scale before power and bandwidth become bottlenecks.