Summary of Key Points
As AI models evolve towards trillion-parameter sizes, ultra-large-scale computing clusters have become a necessity. Traditional electrical interconnections are gradually falling behind due to bandwidth and power consumption limitations, making optical interconnections (which use light to transmit data) a new focus in the competition for AI infrastructure. Major companies such as Tencent and Huawei are actively investing in this area. NPO (Near-Package Optics), as a practical approach to transitioning towards more advanced CPO (Co-Packaged Optics) technology, has gained widespread industry recognition. Meanwhile, there is a surge of capital interest in the optical interconnection sector, with companies from chip manufacturers to equipment providers all vying for a place in this market. However, scaling up still faces challenges such as inconsistent standards and complex packaging processes. The integration of domestic computing power with optical interconnections has become a key strategic direction for these companies.
Why Have Optical Interconnections Become the “New Infrastructure” for AI Computing Power?
AI models are incredibly data-intensive; training GPT-4, for example, requires tens of thousands of GPUs working together. To efficiently transfer data between these GPUs, traditional electrical connections (copper cables) are slow and generate significant heat, leading to increased power consumption. Optical interconnections, on the other hand, use fiber optic cables, which significantly improve speed (by more than ten times) and reduce power usage, allowing for the connection of a much larger number of GPUs (ranging from several thousand to tens of thousands). Therefore, the competition in AI infrastructure is no longer about having the strongest individual GPU but about how to efficiently connect thousands of them—optical interconnections are the core technology that solves this problem.
NPO: The Most Practical Optical Interconnection Solution Currently
There are several technical approaches to optical interconnections, with CPO being the most advanced (combining optical modules and chips in the same package for maximum efficiency). However, CPO is difficult to implement due to its complex packaging requirements, making it challenging for domestic companies. NPO represents a transitional solution: it places the optical module near the chip package, reducing data transmission distances without altering the existing chip packaging process. For example, NPO eliminates the DSP (Signal Processing Chip) found in traditional optical modules, which increases the demand for higher-quality optical engines but compensates for this with reduced signal interference and lower costs and power consumption. Tencent plans to deploy NPO-based super nodes by the fourth quarter of 2026, and other companies like ByteDance and Alibaba have also adopted this technology, indicating a consensus in the industry to start with NPO before gradually moving towards CPO.
The Optical Interconnection Sector: A Hotspot for Industry and Capital
This year’s WAIC (World Artificial Intelligence Conference) featured a dedicated section on AI optical computing power. Companies such as Zhongji Xuchuang (optical modules), Yuanjie Technology (optical chips), and Xizhi Technology (optical interconnection solutions) showcased the entire supply chain from materials to equipment. The capital market has also shown significant interest:
- Zhongji Xuchuang is preparing for a listing on the Hong Kong Stock Exchange to raise funds for expansion.
- American chip giant Marvell acquired Photonic Interconnection company Celestial AI for $3.25 billion.
- Xizhi Technology went public on the Hong Kong Stock Exchange, and startups like Guangbenbit Technology received new rounds of funding.
TrendForce predicts that the CPO/NPO market will grow from approximately $100 million in 2025 to over $39 billion by 2030—this sector is expected to expand by 390 times over the next decade, explaining why everyone is competing for a share of it.
Challenges to Large-Scale Adoption
Despite the promising prospects, there are several obstacles to the widespread adoption of optical interconnections:
1. Inconsistent Standards: Different companies use different NPO solutions, requiring customizations from upstream manufacturers, which increases costs and delays.
2. Technical Difficulties: Complex packaging processes and high performance requirements for optical components (e.g., precise optical engines in NPO).
3. Ecosystem Collaboration: Optical interconnections require coordination among GPU chips, software, and equipment; for example, GPU interfaces need to be redesigned to ensure a seamless end-to-end system.
4. Cost Issues: Customized NPO solutions are currently expensive, making them unaffordable for most companies, which highlights the need for standardization to reduce costs.
The Combination of Domestic Computing Power and Optical Interconnections: A Double Advantage for Major Companies
In addition to deploying NPO, Tencent plans to increase the use of domestic computing power next year and the year after to lower the cost of AI inference (domestic GPUs are cheaper than imported ones). Why combine domestic computing power with optical interconnections? Imported GPUs do not offer their advanced interconnection technologies, and domestic GPU manufacturers lack mature solutions. Optical interconnections can fill this gap, enabling companies to reduce reliance on imports and meet the needs of large-scale clusters. However, Shen Yichen (founder of Xizhi Technology) notes that what domestic computing power lacks most is an open ecosystem—cooperation among computing, storage, and software providers is essential to fully leverage the benefits of both technologies.
Conclusion
Optical interconnections are becoming a necessity in the era of large-scale AI models. NPO is the most practical transitional solution, and the industry and capital are rapidly investing in this area. However, issues such as standards, technology, and ecosystem integration still need to be addressed. Those who can successfully scale up NPO will gain an advantage in the competition for AI infrastructure. The combination of domestic computing power and optical interconnections could be the key to China’s breakthrough in the AI industry.