Compared to cloud computing data centers, AI computing centers experience more frequent technological iterations in computing infrastructure. Since 2023, the commercial use of the ChatGPT large-scale model has sparked global enthusiasm for AI development. The continuous expansion of the parameter size of large AI models has led to explosive growth in computing resource demand, necessitating greater computing power for training and inference, as well as extensive memory resources. It is practically impossible for a single GPU to process all the data for large models. The global AI computing industry has introduced supernode technology to address this issue.

Large AI models involve both scale-up and scale-out networks, and AI computing supernode technology aims to integrate hundreds of GPUs/NPUs to form high-density computing units, such as the NVIDIA GB200 NVL72, Huawei CM384, and ODCC ETH-X. For the optical communications industry, supernode technology means AI computing clusters will usher in a new wave of optical communications technology and market opportunities. According to ICC’s industry discussions, the deployment of supernode technology in AI computing clusters will drive increased bandwidth requirements, technological architecture innovations, network performance breakthroughs, and a reshaped industry ecosystem.

Increasing Demand for Optical Bandwidth

The bandwidth upgrade for AI computing interconnects will drive a structural increase in the optical module ratio. The ratio of next-generation AI chips (such as the B300) to optical modules will increase from 1:3 (H100) to 1:4.5 or even 1:8 (specific ASIC architectures), directly driving demand for high-speed optical modules. ICC predicts that demand for 800G optical modules will reach 22-25 million units in 2025, while Goldman Sachs projects demand for 800G modules at 33.5 million units in 2026. In addition, the interconnection within and across super nodes needs to support TB-level data exchange, which will drive the upgrade of optical modules from 400G/800G to 1.6T. The current 800G pluggable modules will be commercialized on a large scale in 2025. The 1.6T CPO and switch series were unveiled at the NVIDIA GTC conference and are expected to be introduced for commercial use in the second half of 2025. ByteDance also launched the LPO/LRO solution based on 200G/channel this year, which is expected to achieve a 50% reduction in power consumption.

Innovation in Optoelectronic Technology Architecture

AI computing power will accelerate the commercialization of new optoelectronic integration solutions, evolving from pluggable to integrated optoelectronics. On one hand, CPO (co-packaged optics) is expected to become a critical requirement for supernodes. This technology co-packages optical engines with ASICs/GPUs, addressing electrical interconnect bandwidth bottlenecks and energy consumption issues. For example, NVIDIA’s Quantum-X switch will utilize CPO to achieve 1.6T ports, tripling GPU deployment at the same power consumption. Yole Développement predicts that the CPO market will explode from US$46 million in 2024 to US$8.1 billion in 2030 (a CAGR of 137%). Meanwhile, companies like TSMC and Broadcom are integrating lasers, modulators, and waveguides onto a single chip through InP-SiN heterogeneous integration, reducing size by 70% and latency by 50%, effectively driving the large-scale deployment of silicon photonics technology. Even more exciting is the extension of optical I/O (OIO) technology to inter-chip interconnects, opening up new avenues for the deployment of next-generation supernodes in AI computing clusters.

All-optical network performance breakthroughs

Supernodes have extremely high requirements for low latency. The advantages of all-optical switching technology in compressing latency are opening up new application opportunities. For example, Huawei’s AIOTN uses all-optical switching (WSS components are 55% smaller) to achieve “one-hop” connections between metropolitan areas, reducing latency to a level that is imperceptible to services. In real-world cases, for example, Shanghai’s 1ms computing network supports a 30% increase in medical AI model training efficiency. At the same time, supernode optical interconnection requires 99.999% reliability. Huawei has proposed a laser array redundancy solution, which automatically compensates for a single laser failure. Meta-data shows that optical module failures can cause AI cluster efficiency to plummet by 40%. LPO/CPO further reduce failure rates through simplified design.

Reshaping the Optoelectronics Industry Ecosystem

AI computing supernodes will co-evolve with optical communications. Supernodes are more than just hardware stacks; they represent a systemic project integrating “optics, computing, and intelligence.” As Jensen Huang stated, “Over the next decade, the computing bottleneck will be determined by optical power”—and CPO (Computer-Optical Optics) holds the ultimate key to unlocking this bottleneck. AI computing supernodes are driving the rise of the CPO industry chain. On the one hand, upstream silicon photonics chips have become a strategic advantage. The silicon photonics industry has evolved from a battle between EML and silicon photonics solutions to a path characterized by mature silicon photonics processes, low-cost CMOS manufacturing, and customized process platforms. On the other hand, while AI computing supernodes currently deploy mainstream traditional pluggable 400G and 800G modules, NVIDIA’s announcement that its Quantum-X switches will fully incorporate CPO (co-packaged optics) technology by 2025 signals the beginning of a revolution sweeping the optical communications industry chain. CPO, by co-packaging the silicon photonic engine with the AI chip, reconstructs and restructures traditional pluggable optical modules. Furthermore, switch manufacturers are deeply tied to chip manufacturers, bypassing the intermediate optical module stage. For example, Cisco and Arista abandoned their own pluggable module development and instead formed a CPO alliance with NVidia and AMD. As optical engines become the “optical I/O organs” of AI chips, the core of industry competition has shifted from port speed to optoelectronics collaborative design capabilities. CPO not only changes the form factor of optical modules but also completely dismantles the traditional linear division of labor in the supply chain. The future winners will be those “optoelectronics converged system providers” who master the trinity of silicon photonics chips, advanced packaging, and system architecture.

Summary

From the development of supernodes in AI computing clusters to the reshaping of the optical communications supply chain by silicon photonics and CPO, the rise of AI may have a greater impact on the optical communications industry than previous applications such as backbone metropolitan area networks, fiber-optic broadband, wireless access, and cloud computing. Supernode deployment has driven a continuous increase in the scale and computing performance of AI computing clusters, laying the foundation for more intelligent AI large-scale model applications. Furthermore, the deployment of supernode technology in intelligent computing centers requires lower energy consumption, lower latency, and higher density, creating opportunities for CPO applications based on silicon photonics integration and driving changes in the industry chain.