CITIC Construction Investment: The demand for 800G optical modules is expected to continue growing rapidly by 2026, and the shipment scale for 1.6T will also increase significantly.

robot
Abstract generation in progress

CITIC Securities Research Report points out that with the rapid upgrade and iteration of GPUs and ASICs, computing power performance continues to improve, while data transmission demands have also increased significantly. In AI data centers, more and more customers prefer to choose higher bandwidth network hardware. The larger the bandwidth, the lower the cost per bit transmission, the lower the power consumption, and the smaller the size. The rapid growth of 800G optical modules already reflects AI’s urgent need for bandwidth. We believe that by 2026, demand for 800G optical modules is expected to continue growing rapidly, while shipments of 1.6T modules will also increase significantly, and R&D for 3.2T modules has officially begun.

Full Text Below

Prioritizing Computing Power, Focusing on New Technological Breakthroughs

Optical Modules: Continuous High Growth in Demand for High-Speed Optical Modules, Scale-up Expected to Open New Market Opportunities

As GPUs and ASICs rapidly upgrade and iterate, computing power performance continues to improve, and data transmission needs have surged. In AI data centers, an increasing number of customers prefer higher bandwidth network hardware. Larger bandwidth means lower cost per bit, reduced power consumption, and smaller size. The rapid growth of 800G optical modules already indicates AI’s pressing demand for bandwidth. We expect that by 2026, demand for 800G optical modules will maintain high-speed growth, with shipments of 1.6T modules also expanding significantly, and development of 3.2T modules has officially started.

From NVIDIA’s GPU roadmap, the company’s product upgrade cycle has been compressed to roughly one generation per year, with continuous improvements in speed and bandwidth. From 2023-2024, the company will launch the Blackwell platform, equipped with 1800GB/s NVLink 5 switches, 800G CX8 network cards, and 51T Spectrum 5 Ethernet switches; by 2025-2026, the Rubin platform will be introduced, featuring 3600GB/s NVLink 6 and NVLink 7 switches, 102T Spectrum 6 CPO switches, and 1600Gbps CX9 network cards; then in 2027-2028, the Feynman platform will feature NVLink 8 switches, 204T Spectrum 7 CPO switches, and CX10 network cards.

Looking at Google’s TPU product evolution, from the V2 product launched in 2018 to the V7 Ironwood announced in 2025, bandwidth has continued to grow. TPU V2 uses a 2D Torus architecture, with a single superpod containing 256 chips and ICI bandwidth of 800GB/s; Ironwood chips adopt a 3D Torus network topology, with ICI bandwidth of 1200GB/s, using 800G OSPF optical modules, with a single optical channel rate reaching 200Gbps, connected via OCS.

Currently, industry demand for optical modules mainly comes from scale-out networks. Scale-out generally refers to horizontal expansion, increasing the number of compute nodes to enhance overall processing capacity to handle high concurrency or large data volumes. Most newly built large clusters exceed 10,000 nodes, with some reaching hundreds of thousands, highlighting the importance of scale-out. Besides scale-out networks, future demand for scale-up networks’ optical modules will be very broad. Scale-up refers to vertical expansion, increasing the number of GPUs/XPU within a single compute node to boost computing power. Initially, this is mainly in eight-card server configurations, but as the scale-up domain expands to 36/64/72 cards, the single-node product form shifts to rack-level.

From current industry XPU products, the bandwidth for scale-up significantly exceeds that for scale-out. Comparing NVIDIA’s GPUs and various ASICs, the average scale-up bandwidth for XPUs is 10Tbps, while scale-out bandwidth is 800Gbps, with a ratio of 12.5. According to an interview with Broadcom CEO Hock, the company believes the market for scale-up is 5-10 times larger than that for scale-out. We interpret this as referring to network hardware, including switches, optical modules, copper cables, and PCBs. Therefore, all subfields of the network sector are expected to benefit.

As the scale-up domain gradually expands and signal transmission bandwidth increases, the transmission distance of electrical signals in metallic media is greatly limited, and a large amount of heat is generated during transmission, increasing power consumption and affecting GPU and switch chip performance. Therefore, optical interconnects are likely to become the mainstream solution in the scale-up field, solving distance and bandwidth bottlenecks. Besides OIO solutions, Google, Meta, and Huawei have already begun building scale-up networks using optical modules. Analyzing mainstream network architectures in the industry, the demand space driven by scale-up optical modules is broad. For example, NVIDIA’s Blackwell platform offers a bandwidth of 7.2Tbps, nine times that of scale-out bandwidth. As the scale-up domain continues to grow, using a two-layer fat-tree architecture, the ratio of single GPU to 800G optical modules could reach 1:36, offering significant incremental space. Looking at overseas CSP network architectures, if future scale-up entirely adopts optical modules, the market could be 5-8 times larger than current. For example, AMD’s MI400 series has higher-than-average bandwidth for both scale-up and scale-out, with a high proportion of optical modules. We continue to recommend the optical module/optical device/optical chip sector, focusing on: 1) leading optical module companies; 2) companies with potential breakthroughs in market and customer acquisition; 3) upstream optical chip/laser and optical component companies; 4) companies entering the field through acquisitions.

Risk Warning: Changes in the international environment affecting supply chain security and stability; unexpected tariffs; rapid growth in AI demand leading to supply shortages and underperformance; AI industry development below expectations, capital expenditure below expectations affecting cloud computing and computing power industry chain companies; intensified market competition causing rapid decline in gross margins; exchange rate fluctuations impacting foreign exchange gains and gross margins of export-oriented companies, including ICT equipment, optical modules/optical components; slower-than-expected development of digital economy and Digital China; telecom operators’ cloud business growth below expectations; operators’ capital expenditure below expectations; high holdings causing market volatility.

View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • Comment
  • Repost
  • Share
Comment
0/400
No comments
  • Pin