As a critical component in the Artificial Intelligence (AI) era, intelligent computing power is projected to reach 1037.3 EFLOPS by 2025, with a CAGR (Compound Annual Growth Rate) of 46.2% from 2023 to 2028. The optical network serves as the fundamental infrastructure for the digital economy. Optical interconnects facilitate long-distance, high-bitrate, and large-capacity transmission for inter-AI datacenter (AIDC) connections across national, regional, and edge AIDC networks. Various optical modules and connections are employed within different scales of intra-AI datacenters, utilizing either electronic packet switching-based architecture or optical switching-based architecture, to establish computing power pools that deliver abundant computing resources to end-users efficiently and flexibly.

This review comprehensively examines recent developments and key technologies in optical networks for AI datacenter interconnects. The discussion begins with the overall architecture of AI datacenter interconnects, encompassing inter-connections, intra-connections, and intelligent systems, while analyzing the general requirements of optical networks.

For inter-AI datacenter connections, high-bitrate and large-capacity transmission is achieved through C+L bands 400 Gbit/s coherent techniques based on 128 Gbaud DSP and components. The review explores wavelength band extension, addressing challenges and recent developments in S band and other bands. The multi-core fibre (MCF) transmission advances with ITU standardization focusing on G.65x single mode fibre back-compatible MCF. Hollow core fibre (HCF) demonstrates significant advantages despite substantial challenges. The review compares various protection schemes to achieve high availability and reliability in optical networks. Lossless transmission emerges as a new challenge for optical networks supporting distributed datacenter training.

For intra-AI data center connections, the primary objectives include reducing cost per bit and power consumption per bit while achieving lower power usage effectiveness (PUE). The review analyzes and compares different module types for electronic packet switching-based architecture, including retimed modules, linear-drive pluggable optics, co-packaged optics, and optical I/O. Optical switching is being evaluated and tested for new large-scale datacenter architectures, offering significant advantages. Given the presence of various massive modules in datacenters, intelligent operation and maintenance becomes essential.AI-native capabilities are crucial for optical networks, with particular attention given to digital twin and multi-agent applications in network operation and maintenance.

The review summarizes the primary applications and associated requirements of AI datacenter interconnects, including computing power access, model training, deployment, and inference. Additionally, it presents current industrial practices for these applications.

The rapid expansion of AI applications demands substantial computing power, creating significant challenges for AIDC optical interconnects. Optical networks provide the essential infrastructure for inter-AIDC and intra-AIDC connections, offering advantages in sustainability, ultra-broadband capability, intelligence, agility, low latency, high reliability, and cost-effectiveness. Despite these advances, optical networks face ongoing challenges in multi-bands transmission, new fibre-based transmission, convergent optical communication and sensing, and hybrid optical and electronic switching. Future research should focus on optimizing overall architecture, increasing transmission capacity, enhancing automation and intelligence, improving awareness and sensing capabilities, and achieving environmental sustainability in optical networks for AIDC, including both cluster and distributed datacenters, to deliver ubiquitous AI services.