AI infrastructure moves toward scale-across; demand for coherent optics supply chain heats up

DIGITIMES notes that scale-across, which is essentially a data center interconnect (DCI), is gradually becoming the core of the next wave of AI infrastructure. Because a single data center faces limits in power supply, floor space, and cooling, hyperscalers are distributing GPU clusters across campus data centers 2-20 kilometers (km) apart, then linking them with high-speed optical networks so that dispersed computing resources can work together for AI model training and inference. As AI interconnect bandwidth rises to 800G and 1.6T, traditional intensity modulation/direct detection (IMDD) technology is increasingly insufficient for high-capacity, long-distance data transmission, making coherent optics the mainstream technology for scale-across.

However, coherent optics were originally designed for long-haul backbone networks, and optical modules and line systems come at a relatively high cost. If they are applied directly to large-scale AI scale-across deployments, costs would rise sharply. As a result, optical communications equipment vendors are rolling out more cost-effective solutions tailored to AI data center needs, such as coherent lite, which lowers module costs by using simplified digital signal processors (DSPs) and optical component specifications; and multi-rail, which increases line system density, shares cabinets, power, and cooling infrastructure across multiple rails, and reduces the cost of deploying repeater facilities, all of which make coherent optics better suited to large-scale AI data center rollouts.

This report explores how scale-across is gaining traction amid increasing AI adoption, its key components, and the supply chains preparing for its arrival.