Summary of Key Points
The competition in AI computing power has shifted from "chip performance" to "data transfer efficiency." CPO (Co-Packaged Optics), a technology that addresses the data transmission bottlenecks in large-scale GPU clusters, is set to see widespread adoption in 2026. The surge in orders from upstream supply chains and the shortage of production capacity confirm that this demand has moved from a theoretical concept to a commercial reality. Chinese optical chip companies have an opportunity to transition from being mere assemblers to players in the chip manufacturing sector, but they also face challenges in terms of high-end technology and market penetration.
Detailed Analysis
The "Pipeline Crisis" in AI Computing Power: Strong Chips, Yet Slow Data Transfer
Training large AI models requires the coordinated operation of thousands of GPUs, but data transfer between these chips has become a major bottleneck. Traditional solutions use copper cables and pluggable optical modules:
- Copper Cables: They can reach speeds of up to 224Gbps, but the transmission distance is limited to just 2 meters—far from crossing an entire rack.
- Traditional Optical Modules: Additional connections result in significant energy waste due to signal conversion, pushing both bandwidth density and power consumption to their limits. In essence, while chips are the "engines," the data transfer "pipelines" determine the overall system's speed; even the strongest engines cannot perform efficiently if the pipelines are too narrow or have high losses.
CPO: The Revolution That Turns Data Transfer into a Core Component of Computing Power
CPO solves this problem by integrating the optical engine (responsible for signal processing) and the switching chip (responsible for data routing) on the same substrate using advanced packaging techniques. The key principle is to maximize the efficiency of electrical signals (which travel short distances within the chip) while using fiber optics for long-distance communication, which offers higher speeds and lower losses.
- Performance Improvements: Power consumption is reduced by more than 40%, bandwidth is increased by three times, and latency is decreased by 50%. This means that the same amount of power can support larger GPU clusters, making optical interconnects a critical component of computing infrastructure.
Proof from the Upstream Supply Chain: The Real Demand for CPO
Financial reports from three leading manufacturers confirm the practical significance of CPO:
- Lumentum (a leader in optical components): Its CPO-related wafer production capacity is sold out, with revenue expected to reach $50 million in Q4 2026 and several hundred million dollars in orders for 2027.
- Coherent (an optical module manufacturer): It has secured large orders from top AI clients, with orders extending into 2028 and a four-fold increase in shipments.
- Tower (a wafer foundry): Its silicon photonics wafer production capacity is booked through 2028, with an additional $920 million invested in expansion. These figures indicate that CPO demand has moved from the laboratory to the manufacturing stage, with all links in the supply chain competing for production capacity.
The Opportunity for Chinese Optical Chip Companies
In the era of traditional optical modules, Chinese companies excelled in assembly and packaging (holding a significant global share), but they relied on imported core components such as high-speed lasers. CPO changes the game:
- New Requirements: CPOs demand smaller, more power-efficient, and higher-quality optical chips, which directly affect the effectiveness of the entire solution.
- Progress by Chinese Companies: Domestic companies like Yuanjie Technology (in developing domestic high-speed lasers) and Shijia Photonics (in passive components) are making progress. Leading optical module manufacturers such as Zhongji Xuchuang and NeoPhotonics are also investing in CPO technology.
- Remaining Challenges: There is still a gap with international leaders (Lumentum, Coherent) in high-end components, and improvements are needed in areas like thermal management and packaging integration.
The Road Ahead for CPO Adoption
Despite the benefits of CPO, several issues need to be addressed before widespread adoption:
- Technical Barriers: Fiber optic coupling requires precision at the micron level (manual labor is often required), temperatures can exceed the tolerance limits of lasers (85°C vs. 60°C), and there is a lack of industry-wide standards for compatibility.
- Cost Concerns: CPOs are more complex, and replacing a single component may require replacing the entire switch, increasing maintenance costs compared to traditional optical modules.
- Alternative Solutions: 1.6T optical modules are already in mass production, with 3.2T expected by 2028, as well as transitional solutions like LPO that reduce power consumption by two-thirds and are compatible with existing equipment. These alternatives are slowing the adoption of CPO.
Industry experts estimate that full adoption of CPO will take at least 3-5 years. In the short term, CPOs will first be used to connect racks within data centers, while the long-term growth will come from expanding their use between chips (currently, all interconnections rely on electrical connections).
Conclusion
The rise of CPO marks a shift in the AI computing power competition towards "connection efficiency." Optical interconnects are becoming a central component of computing systems, and the value chain will be restructured with increased influence for wafer manufacturers and optical chip companies. For Chinese companies, 2026-2028 is a critical period. By seizing this opportunity, they can transform from a major assembler to a leading chip manufacturer; otherwise, they may fall behind in the technological race.