AI Clusters and the New Economics of Data Center Optics

AI infrastructure demands a fundamental shift in optical networking, as power efficiency and interconnect reliability become critical bottlenecks for scaling GPU clusters. Optics now represent a significant portion of the Bill of Materials, necessitating transitions from 400G to 800G and 1.6 terabit speeds to maintain performance. Link reliability is paramount, as errors trigger costly GPU synchronization restarts, unlike traditional network traffic. While electrical interfaces remain viable for short-range scale-up, physics constraints drive the industry toward optical solutions, including co-packaged and near-packaged optics to reduce power consumption and signal integrity issues. Cisco’s Bill Gartner highlights that routed optical networking and coherent technologies, traditionally reserved for long-haul wide area networks, are increasingly essential within data centers to manage the massive bandwidth demands of modern AI training and inference workloads.