Bandwidth demand in AI intelligent computing scenarios is growing exponentially. Ten-thousand-scale intelligent computing clusters are driving the upgrade of optical modules to speeds of 1.6T and above. Constrained by the rate bottlenecks and cost disadvantages of laser chips, emerging technologies such as silicon photonics and CPO are accelerating their market penetration.Clusters with millions of GPUs are advancing optical interconnection from intra-data center connectivity to cross-data center and cross-regional collaboration, fostering differentiated adaptation solutions. Scale out focuses on interconnection between supernodes, emphasizing standardization and interoperability. Owing to high standardization and strong compatibility, pluggable optical modules remain the mainstream solution at present. Scale up centers on intra-cluster and intra-super-node interconnection, pursuing ultimate bandwidth, ultra-low latency and low power consumption. It will be anchored by NPO/CPO solutions, which boast significant advantages in bandwidth density, latency and power consumption. Among them, NPO features better ecosystem compatibility and is expected to be deployed in end-user businesses earlier than CPO. The industry forecasts that 2027 will likely mark the first year of commercial adoption for NPO.

Despite the rapid development of optical interconnection driven by AI services, it still faces multiple core challenges. Technically, it is necessary to balance multi-dimensional performance including bandwidth density and energy efficiency, while technological breakthroughs for Scale up scenarios remain difficult. The packaging processes of CPO are immature, and OCS has insufficient compatibility. High-speed optical modules face bottlenecks in power consumption and heat dissipation. In addition, there is a lack of unified standards for emerging technologies like NPO and CPO, and core components maintain high costs. Optical modules account for over 50% of the total network cost, restricting large-scale deployment.

In end-user applications, AI intelligent computing scenarios feature unique traffic characteristics that traditional solutions can hardly adapt to. The implementation of CPO requires changes to operation and maintenance models, and the construction of a cross-domain slicing management and control system is highly complex. In terms of localization, the core processes and components of CPO rely on imports. Key upstream materials (such as silicon photonic chips and Faraday rotators) suffer from temporary supply shortages, while domestic mass production capacity is limited, posing immense pressure for domestic substitution.

Based on this background, this forum aims to gather resources from industry, academia and research sectors. CIOE collaborates with internet service providers, data center equipment vendors, and core manufacturers of optical modules, components and chips to host the Optical Interconnect Technology for AI Computing Clusters Forum. The event will facilitate in-depth discussions on key next-generation optical interconnection technologies, innovative applications and future development directions in the AI computing era.