Citi's AI Infrastructure Summit notes indicate that as AI workloads become more complex, infrastructure must address issues such as interconnection efficiency, memory capacity, bandwidth, power consumption, and cost simultaneously. Broadcom believes Ethernet will become the preferred interconnection solution for large-scale AI infrastructure, potentially saving over 30% in five-year total cost of ownership (TCO). Samsung disclosed its HBM roadmap, projecting that the new generation zHBM could achieve a 4 to 8 times single-GPU performance improvement compared to HBM5, while reducing thermal resistance by 75% to 90%. Marvell, meanwhile, is focusing on memory hierarchy optimization for KV caching, and AWS is enhancing memory bandwidth economics through custom chips.