万亿Marvell背后：中国光通信巨头的效率霸权

This article outlines China's global competitive position and current development status of the optical module industry amid the AI computing power era, with key takeaways as follows:

1. Competition in the AI industry has shifted from a race for individual GPU performance to interconnection networking for 10,000-card training clusters. Data transmission speed has become the core bottleneck, and high-end high-speed optical modules are the key component to improve cluster computing power utilization. A single 10,000-card cluster requires 100,000 optical modules.

2. U.S. export controls primarily target high-end GPUs and do not restrict optical module packaging. After two decades of accumulation in China's communications industry, three Chinese companies have built a global leading position in optical module packaging, collectively holding 62% of the global market share for 800G and higher-end optical modules.

3. The industry still has notable weaknesses: high-end DSP chips are heavily reliant on imports, exposing the sector to potential regulatory risks. Domestic companies are bypassing this dependency via the LPO technology path, while advancing independent R&D for upstream components to gradually achieve full industrial chain self-sufficiency.

This article sorts out shifting demand and the competitive landscape of the optical module industry amid the AI boom, with core takeaways for optical communication brands as follows:

1. Industry demand trend: After large AI models surpassed the trillion-parameter mark, 10,000-card training clusters have become widespread, pushing demand for high-end high-speed optical modules from the millions to tens of millions level. The market is seeing explosive growth, making it a clear high-growth track with broad development space.

2. Product R&D direction: Brands need to proactively forward-position cutting-edge technology roadmaps. For example, Eoptolink gained a first-mover advantage by early布局 of LPO (linear direct-drive) optical modules. At the same time, LPO can only cover medium- and short-distance scenarios, and DSP-equipped high-end optical modules will remain the market mainstream for the next three to five years, so brands need a balanced布局 of product lines.

3. Building brand competitiveness: A brand's core moat is not just technical product specifications, but more importantly stable mass production delivery capacity. Once approved by leading cloud providers, it creates stable purchasing inertia that is hard for new entrants to disrupt. Long-term process accumulation delivers more value than pure price competition.

The optical module track is seeing a historic growth opportunity in the AI era, but also carries potential risks, with key takeaways for industry practitioners as follows:

1. Growth opportunity: More than half of global high-end optical module supply is controlled by Chinese companies. The AI computing power boom has driven rapid demand expansion. Any player with solid process accumulation and stable delivery capacity can secure a core supplier quota from top cloud providers and enjoy enormous dividends from industry growth.

2. Risk warning: The localization rate of high-end digital signal processing (DSP) chips is currently less than 5%, heavily reliant on overseas suppliers such as Marvell. While export controls have not yet extended to this sector, there is future policy risk. If controls are implemented, the entire industry will be thrown into a passive position, so practitioners need to stay vigilant in advance.

3. Mitigation strategies: Practitioners can proactively布局 the LPO technology path to bypass reliance on imported DSP chips integrated in optical modules. They should also follow the mass production progress of 1.6T DSP from domestic players such as Phylite, and diversify supply chain布局 in advance to hedge against potential policy and supply chain risks.

This article offers key insights for optical module and related manufacturing facilities, with core takeaways as follows:

1. Production requirements: Optical module packaging has extremely high precision requirements: micron-level optical chip coupling accuracy must reach nanometer levels. Mass production requires long-term process iteration to control the impact of deviation, temperature drift and vibration on yield. This process accumulation is the core competitiveness for manufacturing, and cannot be obtained directly from academic papers.

2. Business opportunity: The AI computing power boom has driven explosive growth in demand for high-end optical modules. After two decades of accumulation in China's communications industry, Chinese players already hold over half of the global high-end optical module market. Factories with precision manufacturing capabilities can enter this track to capture enormous growth space.

3. Development insights: Do not underestimate the value of manufacturing. Long-term engineering iteration builds barriers that are hard for other players to surpass. Even in high-tech industries, manufacturing advantages can become an unassailable strategic stronghold. Factories need to prioritize deep alignment between R&D and manufacturing, and continuously refine process capabilities.

This article sorts out development trends and core pain points of the optical module industry, with core takeaways for optical communication-related service providers as follows:

1. Industry development trend: The second half of AI computing power competition has shifted from individual GPU performance to 10,000-card cluster networking. Demand for high-end high-speed optical modules is growing explosively, and global supply is increasingly concentrated among leading Chinese manufacturers. Industrial chain self-reliance is the core long-term development direction for the industry, creating huge space for domestic substitution.

2. Core customer pain points: Domestic optical module manufacturers currently face heavy reliance on imports for high-end DSP chips, high-speed optical chips and specialized packaging equipment, exposing them to potential supply chain control risks. Meanwhile, the commercialization of the new LPO technology path still leaves unmet needs across multiple links, including driver adaptation and parameter tuning.

3. Business opportunities: Service providers can布局 along the domestic substitution direction, filling gaps in upstream equipment, chip design testing and other fields. They can also support the commercialization of new technologies such as LPO by offering technical adaptation, testing verification and other related services, to capture incremental business brought by industry transformation.

Shifting dynamics in the optical module industry offer these key insights for cloud services, computing platforms and other related platform operators on operations and supply chain management:

1. Changing supply chain requirements: In a 10,000-card GPU cluster, the delivery timeline of optical modules directly determines when the cluster can go live. Delivery delays will leave large amounts of expensive GPU hardware idle, causing substantial revenue losses. As a result, platforms prioritize delivery certainty far more than product price.

2. Sourcing and supplier management direction: Three leading Chinese optical module manufacturers collectively hold 62% of the global high-end optical module market, and their delivery capabilities have been validated by the world's top cloud providers. Platforms can include these manufacturers in their core supplier systems to lock in stable supply and guarantee on-time project launch.

3. Risk mitigation: Platforms should proactively diversify their technology roadmaps, pilot LPO optical modules to reduce reliance on imported DSP, and follow the mass production progress of domestic high-end DSP. They should build backup supply chains in advance to mitigate the risk of supply chain disruption if export controls are extended to this sector in the future.

This article outlines new trends and challenges in the optical communication industry in the AI era, with key insights for industry researchers as follows:

1. New industry trends: The development of large AI models has shifted the focus of global competition in the computing power industry, from individual GPU performance to interconnection efficiency of 10,000-card clusters. Optical modules have become a core strategic link in AI infrastructure. After two decades of development, three Chinese manufacturers hold over 60% of the global high-end optical module market, building unique competitive advantage in an area untouched by U.S. export controls.

2. New industry challenges: Full industrial chain self-reliance still has clear gaps: the localization rate of high-end digital DSP chips is less than 5%, leaving the sector heavily reliant on overseas suppliers and exposed to potential regulatory risk. The current LPO technology path can only cover medium- and short-distance scenarios, and cannot fully replace DSP-equipped products. Full industrial chain self-sufficiency still requires long-term investment.

3. Research insights: Latecomer economies do not necessarily need to break through technology blockades head-on. Continuous capability building in blind spots of existing blockades can also build hard-to-replicate strategic advantages. Long-term accumulation of manufacturing processes can build deep competitive moats, offering a new path for industrial upgrading.