In the previous article, we broke down the compilation rules and sample structure of the ChiNext Compute Power Infrastructure Index (970083), and classified all 50 constituent stocks according to the four major stages of a "compute factory." Among them, the transmission compute power segment contains 13 related targets with a combined weight of 28.12%, making it the second-largest block among the four stages. Starting from this article, we will unfold each stage one by one, beginning with transmission compute power: what hardware builds this "highway" of computing, and why does everyone say that "optical is essential"?
Breaking Down Transmission Compute Power: What Hardware Makes Up the Compute Highway
Three Major Hardware Categories Build the Compute Transmission Network
If we compare a compute cluster to a large factory, with chips as the production workshops, then the entire optical interconnect system corresponding to transmission compute power is the blood vessels and highways running through the whole plant. Transmission compute power is mainly composed of three sub-sectors: optical modules, switches, and optical components and connection assemblies. Figure 1: Breakdown of the transmission compute power industry. The core unit of transmission compute power is the optical module, which is further broken down into active optical components, core electrical chips, and structural and packaging parts; switches handle traffic scheduling; optical components and connection assemblies include various connectors, optical fibers, and passive optical components, together forming the entire transmission compute power hardware system.
Each Plays Its Role: Understanding the Division of Labor Among the Three Major Hardware Categories
Optical modules handle photoelectric signal conversion and are the "translators" for communication between GPUs — electrical signals inside servers must be converted into optical signals for long-distance, high-speed transmission, and then converted back into electrical signals at the receiving end; switches are responsible for data distribution and scheduling, serving as the traffic hubs in the cluster, completing data packet routing and traffic allocation between compute nodes; optical components and connection assemblies belong to the physical link, providing transmission channels for optical signals. All links work together to complete data interaction within compute clusters, across servers, and across racks. In the construction of AI intelligent compute clusters, optical modules account for the highest value proportion among the entire transmission compute power hardware, and are the core components determining cluster communication bandwidth and latency performance, while optical components and switches play supporting and coordinating roles. A shortfall in any link will constrain overall interconnect performance. (Note: An intelligent compute cluster is a compute cluster purpose-built for AI large model training, networking a large number of GPUs for collaborative computation, with very high requirements for internal data transmission speed.)
Why Is Interconnect More Important as AI Develops Further?
Single-Card Compute Power Has a Natural Ceiling; Clusters Are a Necessity
A single GPU chip has a natural ceiling in compute power, and large model training tasks cannot be completed by a single card — they must rely on parallel computation across compute clusters of thousands or even tens of thousands of cards. Relevant brokerage reports note that in large-scale cluster training scenarios, multiple GPUs need to exchange model update information with each other, and data exchange volumes rise exponentially. The continuous rapid expansion of domestic intelligent compute power scale in recent years also indirectly confirms the major trend of cluster-based construction. Figure 2: China's intelligent compute power scale (EFlops) and year-on-year growth rate (%). Data source: Wind, statistical period: 2021 to June 30, 2026. As shown in Figure 2, from 2021 to the first half of 2026, domestic intelligent compute power continued to grow at high speed, with year-on-year growth rates reaching 223% and 177% in 2025 and the first half of 2026 respectively. The construction scale of compute clusters has risen rapidly, and the interconnect pressure within clusters has amplified correspondingly, driving demand for high-speed optical interconnect hardware higher.
If Bandwidth Can't Keep Up, Even the Strongest GPUs Will "Slack Off"
Once communication bandwidth is insufficient and latency is too high, GPUs will sit idle waiting for data, the theoretical compute power of the hardware cannot be fully released, and overall cluster utilization will decline significantly. Simply put, no matter how strong the chip performance is, if the "highway" is congested, a large amount of compute power will be wasted. As model parameter counts continue to rise and cluster scale expands from hundreds of cards to tens of thousands of cards, the compute bottleneck has gradually shifted from single-card chips to the interconnect network within clusters. Institutional estimates show that for a ten-thousand-card large model training cluster, if interconnect bandwidth is insufficiently configured, effective GPU utilization may even fall below 50%, with a large amount of hardware compute power consumed for nothing. In the compute formula, equipment efficiency is largely determined by interconnect capability, which is why optical interconnect has upgraded from a supporting role to a core necessity in AI compute power construction.
The Underlying Business Logic of Transmission Compute Power: Making Money by Following Generational Upgrades
Speed Iteration: Product Upgrades Open Up Value Space
First, the industry is in a rapid upgrade cycle transitioning from 800G to 1.6T. Optical module speed iteration is not just about doubling bandwidth numbers — it also drives an increase in the number of optical modules per server and a rise in unit value per product. New product iteration has become the core source of corporate profit growth. Relevant institutions point out that 800G remains the current mainstream, while 1.6T products are entering a large-scale volume ramp-up cycle, and future 3.2T products are also in the sampling and verification stage. The industry maintains a generational iteration rhythm of once every 2 to 3 years. Compared with the previous generation 800G products, the unit value of 1.6T optical modules increases by about 60% to 80%, and product iteration brings significant value increments. For manufacturers, relying solely on price competition with old products makes it difficult to maintain profitability. Whether they can keep up with the generational rhythm and stably produce qualified high-speed new products is the key to widening the profit gap between companies. Leading manufacturers can rely on the higher value of new-generation products to offset the pressure of price cuts on old-spec products, achieving maintenance and improvement of profit levels — this is also the core logic of "iteration equals dividend" in the optical interconnect sector.
Simultaneous Volume and Price Resonance: Dual Drivers of Cluster Expansion and Product Upgrades
Second, the industry has the characteristic of "rising volume and price together." Continuous cluster expansion brings growth in demand quantity, while generational upgrades bring increases in unit prices, and the dual dividends drive sector revenue and profit upward. According to LightCounting statistics, the global optical module market size reached US$26.8 billion in 2025, a significant year-on-year increase of 74%; in the first quarter of 2026 alone, the market size already reached US$10 billion, with a year-on-year growth rate exceeding 90%, showing exponential growth in the industry. (Note: Data communication optical modules are mainly used for data transmission between servers in AI compute clusters and cloud data centers.) By speed structure, 800G and 1.6T high-speed products are the main growth drivers. In 2026, the combined market size of 800G and 1.6T is expected to reach US$14.6 billion, accounting for about 64% of the overall data communication optical module market. In the 800G and above high-speed data communication optical module segment, domestic manufacturers hold a global market share of over 70%, fully enjoying the industry dividends brought by speed iteration.
Global Supply Chain: Domestic Manufacturers' Cost and Response Advantages
Third, domestic manufacturers hold advantages in the global supply chain. Domestic manufacturers collectively occupy more than 70% of the global market share in the high-speed optical module field. According to customs statistics, China's optical module export value reached 5.778 billion yuan in August 2026, a year-on-year increase of about 120%. Relying on manufacturing capabilities, cost advantages, and rapid customer response capabilities, they continue to secure orders from overseas cloud manufacturers while also deeply participating in the construction of domestic intelligent compute centers, and their voice in the global optical interconnect supply chain continues to rise. Figure 3: China's monthly optical module (85177950) export value over the past three years (100 million yuan). Data source: Customs Statistics Data Platform, statistical period: August 2023 to August 2026. In addition to manufacturing cost advantages, domestic manufacturers respond faster in iteration and can quickly complete sample iteration in line with the customized needs of overseas cloud manufacturers, which is an important reason why domestic companies can capture global share.
High Industry Prosperity: Opportunities Under the Resonance of Domestic and Overseas Demand
Domestic and Overseas Demand Resonate, High-Speed Interconnect Demand Continues to Be Released
From the overseas market perspective, leading North American cloud manufacturers continue to raise capital expenditure plans, with AI infrastructure becoming the core direction of capital deployment, directly driving procurement scale for high-speed optical modules, switches, and other interconnect hardware. Institutional statistics show that in the second quarter of 2026, the combined capital expenditure of the four major North American cloud manufacturers reached US$171.2 billion, a year-on-year increase of 78.6% and a quarter-on-quarter increase of 30.1%. The median estimate for the four major cloud manufacturers' total capital expenditure in 2026 is about US$732.5 billion, with the vast majority of resources directed toward AI compute cluster construction, and demand for high-speed interconnect hardware continues to spill over. Goldman Sachs estimates that, driven by both AI server architecture upgrades and an increase in the number of optical modules per server, the global optical module market size will continue to rise substantially from 2026 to 2028, with the 2026 market size expected to reach US$67.7 billion, among which 800G and 1.6T high-speed products are the main growth drivers. Order feedback from the industry chain also verifies the prosperity. A number of overseas communications manufacturers have disclosed that order schedules for high-end optical modules have extended to 2028, with downstream customers locking in capacity through prepayments, reflecting the high certainty of hardware demand. Domestic intelligent compute center construction has also entered an accelerated implementation stage, forming a demand pattern driven by both overseas and domestic factors. According to Shuzhi Qianxian statistics, in the first seven months of 2026, the number of winning bids for domestic data centers and intelligent compute centers at the hundred-million-yuan level and above reached 158, with a total winning bid amount exceeding 113 billion yuan, and both the number of projects and total amount increased by more than 18% year-on-year. More than twenty provinces across the country are simultaneously advancing intelligent compute cluster construction. With the successive implementation of domestic ten-thousand-card-level domestic AI compute clusters, procurement demand within intelligent compute clusters for 800G, 1.6T high-speed optical modules, and high-speed switches is being rapidly released. The China Commercial Industry Research Institute predicts that the domestic optical module market size is expected to reach 56.4 billion yuan in 2026, a substantial leap compared with previous years. Domestic optical interconnect hardware is deeply participating in domestic intelligent compute infrastructure construction, opening up local incremental space.
Industry Landscape: Head Concentration, Generational Switching Ushers in Share Window
The current high-speed optical interconnect sector overall presents a competitive landscape of head concentration and structural opportunities brought by generational switching. Institutions point out that during the key cycle of iterating from 800G to 1.6T products, customers place higher demands on product yield, delivery stability, and rapid iteration response capability. Leading manufacturers that have passed certification by major overseas customers, relying on supply chain management and mass delivery capabilities, preferentially take on incremental orders, and the industry's Matthew effect continues to intensify. From the supply structure perspective, 1.6T products are at a key inflection point moving from sample verification to large-scale mass delivery. LightCounting predicts that global 1.6T optical module shipments are expected to reach the ten-million level in 2026. The ramp-up of new-generation products also opens a share reshuffling window for second-tier manufacturers with technical reserves. Relying on a complete manufacturing industry chain and rapid customized iteration capabilities, domestic companies have become the core supply force for global high-speed optical interconnect hardware. They are deeply bound to the supply chains of overseas cloud manufacturers while also fully benefiting from the wave of domestic intelligent compute center construction, opening up growth space in both domestic and overseas markets.
How Does the ChiNext Compute Power Index Allocate to Transmission Compute Power?
The transmission compute power-related targets in the ChiNext Compute Power Index (970083) cover sub-sectors such as optical modules, optical components, and switches. Related constituent stocks include Linktel Technologies, Eoptolink Technology, Zhongji Innolight, T&S Communications, Changxin Bochuang, Afr Fibertech, Tfc Optical Communication, Zhishang Technology, Kingsignal, Zhaolong Interconnect, Wanma Technology, Ruijie Networks, and Feiling Kesai (data source: Wind, as of October 18, 2026). From the perspective of the index weight structure, transmission compute power-related targets together account for 28.12% of the index weight, among which the optical components sub-industry accounts for the highest proportion. Relying on the characteristics of ChiNext-listed companies, the transmission compute power companies selected by the index participate in global AI compute power construction while also benefiting from the wave of domestic intelligent compute cluster construction, covering the core links of the transmission compute power industry chain. Figure 4: ChiNext Compute Power Index: weight distribution of constituent stocks in the transmission compute power segment. Data source: Wind, as of October 18, 2026.
Summary
In order to produce more compute power, optical interconnect continues to iterate and evolve, which also places new requirements on the "base" that carries them, prompting synchronous improvement of the base — and that "base" is PCB. In the next article of the Decoding Compute Power series, we will focus on the PCB segment that carries compute power and continue to break down the underlying logic of compute hardware. If you are interested in compute power investment, welcome to continue following the "Decoding Compute Power" series! Also welcome to follow GF ChiNext Compute Power ETF (158049)! GF ChiNext Compute Power ETF (158049) tracks the ChiNext Compute Power Infrastructure Index (970083), pure AI hardware, focusing on core targets in compute infrastructure. Its top ten constituent stocks include Xiechuang Data, T&S Communications, Kexiang Co., Softcom Power, Wangsu Science & Technology, Changxin Bochuang, Jingjia Micro, Hongjing Technology, Linktel Technologies, and Afr Fibertech (data source: Wind, as of September 30, 2026). Data sources: Wind, Customs Statistics Data Platform; Huatai Securities "AI Compute Interconnect In-Depth: Optical Interconnect Is the Bottleneck of Cluster Efficiency"; Goldman Sachs "Optical Module Industry Update: 1.6T Demand Expectations Revised Upward"; CICC "Optical Communications In-Depth: Optical Expo Review, High Prosperity Continues"; CITIC Securities "AI Compute Prosperity Continues, Holdings Hit New Highs, Cloud Manufacturers' Capital Expenditure Continues to Increase"; Orient Securities Communications Industry Report; CITIC Securities Communications Industry Research Report; East Money Research Center "Optical Module Equipment Industry In-Depth: AI Development Drives Explosive Demand for Optical Modules"; LightCounting "Ethernet Optics," LightCounting June 2026 "Quarterly Market Update Report"; Ministry of Industry and Information Technology "High-Speed Optical Module Industry Development White Paper"; Shuzhi Qianxian Industry Project Statistics. The above individual stocks are only index constituent stocks and are not investment recommendations. The above data are point-in-time data only. The index company may subsequently adjust the index compilation plan and sample stocks, and the composition and weight proportion of index constituent stocks will change accordingly. Investors should pay attention to the risks that such adjustments may bring. Funds involve risks, and investment must be cautious. The sales fees for GF ChiNext Compute Power ETF are as follows: when investors subscribe for fund shares, the subscription agency broker may charge a commission of no more than 0.30%; when redeeming fund shares, the redemption agency broker may charge a commission of no more than 0.50%. On-exchange trading fees are subject to the actual charges of the securities company. For details, please refer to the prospectus and other legal documents and fund announcements. Risk warning: The above fund invests in the securities market. Before investing in this fund, investors need to fully understand the product characteristics of this fund and bear various risks arising from fund investment. The above fund is issued and managed by GF Fund Management Co., Ltd., and the distribution agency does not assume responsibility for the product's investment and redemption. Before investing, please carefully read the fund contract, prospectus, and other legal documents of the above fund to fully understand the details and risk characteristics of this fund. The above fund is an equity fund, with risks and returns higher than mixed funds, bond funds, and money market funds. The specific risk rating results are subject to the rating results provided by the fund manager and sales institution. Investors should choose products that match their risk tolerance and investment objectives. Funds involve risks, and investment must be cautious. A MACD golden cross signal has formed, and these stocks are rising well!
Comments