Zigong and Tsinghua University Propose ZCube Networking Architecture: 15% Increase in Large Model Inference Throughput, One-Third Reduction in Network Costs

Large model inference is driving changes in AI infrastructure, with networking architecture innovation becoming key to unlocking hardware potential. Zigong, Yuxun Network, and Tsinghua University presented research on the ZCube networking architecture at ACM SIGCOMM 2025, and successfully deployed it in the GLM-5.1 coding production environment in May 2026. Benchmark tests show that, under unchanged GPUs, software stacks, and applications, the ZCube architecture significantly reduces capital expenditures on switches and optical modules.