Physical Layer Standardization for AI Data Centers: Challenges, Progress and Perspectives

The rapid advancement of artificial intelligence (AI) is driving unprecedented demand on datacenter infrastructure, especially within the physical layer. With the parameter size of large language models doubling every four months in the recent years, physical layer connectivity faces immense pressure to meet new demands for reliability, bandwidth, power efficiency, and latency. This paper begins by examining the architecture of AI datacenters, highlighting key network segments, detailing specific requirements for critical functions such as model training and inference. Next, we review recent progress in input/output (I/O) interconnect technologies for both electrical and optical links, along with the major challenges that remain. Finally, we explore ongoing standardization efforts for electrical and optical interconnects, emphasizing the potential for achieving 400G per lane and assessing the feasibility of convergence across various standards.