研报 | 受谷歌高速互连架构带动，预估2026年800G以上光收发模块占比将突破60%

Core Insights - The article discusses the advancements in high-speed interconnect technology driven by AI demands, particularly focusing on Google's new Ironwood cabinet system and its integration with Apollo OCS optical networks, which is expected to significantly increase the global shipment share of 800G optical transceivers from 19.5% in 2024 to over 60% by 2026, establishing it as a standard in AI data centers [2][5]. Group 1: Technology Advancements - Google's Ironwood cabinet system utilizes a combination of 3D Torus network topology and Apollo OCS full optical networks to create a high-speed interconnect architecture [2]. - The architecture allows for short-distance interconnects using high-speed copper cables within the cabinet, while long-distance data transmission between cabinets is handled by optical networks [5]. - The design emphasizes energy efficiency, with Apollo OCS technology using micro-electromechanical systems (MEMS) to connect data fibers directly, reducing power consumption from approximately 3000 watts for traditional switches to about 100 watts for a single OCS switch, a reduction of about 95% [5]. Group 2: Market Implications - TrendForce estimates that Google's TPU shipments will reach nearly 4 million units by 2026, leading to a demand for over 6 million units of 800G optical transceivers [5][6]. - The transition from 800G to 1.6T bandwidth can be achieved by simply replacing high-speed optical modules, making upgrade costs more competitive without the need for a complete system overhaul [6]. - Key suppliers like Innolight and Eoptolink are expected to capture nearly 80% of Google's orders for 800G optical modules, while Lumentum plays a crucial role in the OCS system and MEMS supply chain, impacting the implementation pace of Apollo OCS [6]. Group 3: Future Outlook - As computational power continues to grow, the demand for data transmission between cabinets and clusters will also increase, making the evolution and supply of high-speed optical modules and laser components a critical factor in determining the speed and cost of computational expansion, alongside GPUs and memory [6].