The explosive expansion of AI computing power is reshaping the optical interconnect industry, yet the outcome of this technological race is not the dominance of a single path.
A recent industry report from Zheshang Securities highlights that the optical interconnect sector currently does not feature a single technology comprehensively replacing all others. Instead, it exhibits distinct characteristics where application scenarios dictate technology selection, multiple paths operate collaboratively, and they are set to coexist over the long term.
As data center electrical interconnects hit bottlenecks in bandwidth, latency, and power consumption, multiple differentiated technological paths—including silicon photonics, LPO, LRO, NPO, CPO, and TFLN—are evolving concurrently.
Betting on a single technological path in the optical module race carries significant risk. Meanwhile, value within the industrial chain is concentrating in high-barrier upstream segments like optical chips, advanced packaging, and specialized optoelectronic materials. These core components will become the key variables determining the development ceiling for each technological path.
Three Bottlenecks Drive Parallel Development of Multiple Paths
Traditional electrical interconnect solutions are facing systemic limitations. As single-node computing power surpasses hundreds of exaflops per second, copper-based electrical interconnects encounter the triple challenges of a 'bandwidth wall,' 'latency wall,' and 'power wall.' Single-channel rates struggle to exceed 400Gbps, transmission latency reaches several microseconds, and the power consumption for interconnects within a single rack can exceed 40% of the total.
Optical interconnect technology has thus become the inevitable solution, with the trend of 'optics replacing electronics' being irreversible. However, the issue is that optical interconnect itself is not a monolithic technology. According to China Mobile's "Optical Interconnect Technology White Paper for Large-Scale AI Computing Cluster Scenarios (2025)," optical interconnect technologies can be divided into two main categories: equipment-level and chip-level. The former primarily includes pluggable optical modules, while the latter encompasses near-packaged and co-packaged optical solutions like NPO and CPO.
Different technological paths prioritize different dimensions, such as power control, transmission latency, port bandwidth density, and equipment maintainability. This is the fundamental reason for the formation of a multi-path parallel landscape. Zheshang Securities analysts Deng Hefang and Zhou Yixuan point out in their report that to clarify this landscape, a comprehensive analysis is needed, combining the differentiated requirements of various scenarios regarding equipment maintainability, hardware standardization, and the maturity of the industrial chain ecosystem.
Silicon Photonics: A Platform-Level Foundation with Rising Penetration
Silicon photonics is not a specific product but a platform-level foundational technology for the entire optical interconnect field. Its core advantage lies in its high compatibility with CMOS processes, enabling ultra-large-scale mass production using mature wafer fabs like TSMC, Intel, and GlobalFoundries. It also offers ultra-high integration density and powerful optoelectronic integration capabilities.
Market data supports this view. According to a LightCounting report from May 2026, 2026 is projected to be the milestone year when transceiver sales using silicon photonic modulators first exceed 50% of the total market of over $40 billion. Yole Group predicts the silicon photonics market will grow from $278 million in 2024 to approximately $2.7 billion by 2030, representing a compound annual growth rate of 46%.
Over a longer cycle, the optical chip market is expected to grow from $4 billion in 2025 to about $15 billion by 2031. Within this, the share of silicon photonic chips is projected to rise from the current one-third to 42%, corresponding to a scale of roughly $6.3 billion. Notably, the penetration path for silicon photonics will continue to extend with the evolution of optical interconnect architectures—progressing from current scale-out networks to scale-up networks and eventually to scale-in networks within the package.
Three Divisions Within the Pluggable Camp
Within the realm of pluggable optical modules, the industry is undergoing technological differentiation by adjusting DSP configurations, forming three parallel paths: FRO (full DSP), LRO (half retiming), and LPO (full linear).
LPO was introduced in 2022 by Macom in collaboration with NVIDIA. Its core logic is to completely eliminate the DSP chip, using a pure analog linear-drive architecture to achieve significant reductions in power consumption and latency. According to Macom data, the power consumption of an 800G multimode optical module can drop from over 13W to below 4W, with an overall cost reduction of about 8%. However, the trade-offs for LPO are also significant: weak noise resistance, application scenarios limited to short-distance interconnects under 500 meters, a current lack of unified interoperability standards, and high demands on system-side SerDes performance.
LRO represents a more pragmatic compromise. It retains only one DSP on the transmitter side to ensure signal quality meets IEEE 802.3 standards, while the receiver side employs a linear analog architecture to reduce power consumption. A March 2026 technical report from the IEEE Electronic Packaging Society points out that when single-channel rates rise to 200G/lane and total module rates reach 1.6Tbps, the power consumption of a full DSP solution is expected to exceed 30W, whereas LRO can control it below 20W. This threshold means air cooling can be retained instead of requiring liquid cooling, significantly reducing deployment complexity. The report also reveals that almost all companies demonstrating 1.6T LPO solutions at OFC 2025 also showcased LRO solutions simultaneously. The industry broadly believes LRO is more viable for implementation in the 1.6T era than LPO.
NPO: The Mainstream Choice for Current Large-Scale Deployment
NPO (Near-Packaged Optics) positions itself as a pragmatic transitional solution between traditional pluggable modules and CPO. Its core design involves mounting the optical engine on the switch motherboard near the ASIC chip, shortening the electrical signal path to the centimeter level. This significantly reduces insertion loss while maintaining the replaceability of the optical engine.
The competitiveness of NPO lies in balancing performance with industrial reality. Technical experts from Alibaba and Tencent believe that while CPO is the optimal solution in terms of performance, its lack of an open ecosystem is a major concern. In contrast, NPO can rely on the mature pluggable optical module ecosystem while offering significant improvements in bandwidth density and power consumption. Chen Qin, an optical network architect at Alibaba Cloud, notes that at rates ≤224G/lane, NPO has sufficient performance headroom and can fully leverage the existing industrial chain, making large-scale deployment easier. NPO is also currently the primary technological path chosen by domestic GPU chip manufacturers.
Market size data supports this assessment. According to DataIntelo, the global near-packaged optics market was valued at $3.8 billion in 2025 and is expected to grow at a compound annual growth rate of 19.3% from 2026 to 2034, reaching $18.6 billion by 2034. North America leads with a 36.2% market share, while the Asia-Pacific region is expected to catch up with the fastest regional CAGR of 21.4%.
CPO: The Ultimate Direction, But Commercialization Challenges Should Not Be Underestimated
CPO (Co-Packaged Optics) is widely recognized within the industry as the 'ultimate solution.' By integrating the optical engine and switching ASIC on the same substrate using 2.5D/3D advanced packaging technology, the electrical signal transmission path is compressed from over 100 millimeters in traditional solutions to the millimeter level. This reduces power consumption by 30% to 50% compared to traditional solutions while achieving nanosecond-level ultra-low latency and single-channel bandwidth density of 3.2T+.
NVIDIA and Broadcom are the most aggressive proponents of CPO. NVIDIA announced its Quantum-X and Spectrum-X silicon photonic co-packaged chips at the 2025 GTC conference, with plans to deliver InfiniBand CPO systems in the first half of 2026. Broadcom delivered the industry's first 51.2Tbps CPO Ethernet switch, Bailly, in March 2024. Its CPO production line is expected to enter a critical mass production phase in the second half of 2026, with monthly output potentially jumping to the ten-thousand-unit level by the first quarter of 2027. Regarding market forecasts, LightCounting predicts the CPO market could reach $10 billion by 2030, while Coherent revised this estimate upward to $15 billion at the OFC conference.
However, the commercialization challenges for CPO are equally significant and cannot be ignored. Technically, CPO involves deep integration across multiple fields including chip design, photonic integration, advanced packaging, and thermal management. A standardized system across the entire industrial chain has not yet been formed, and the cost of a single optical engine is as high as $35,000 to $40,000. From an operational perspective, CPO adopts a 'non-pluggable' architecture where the optical engine is permanently bound to the expensive ASIC. A failure necessitates replacing the entire composite module, completely overturning the existing data center operational ecosystem. Furthermore, there is a lack of consensus on interoperability between NVIDIA's COUPE solution and Broadcom's FOWLP solution, and the absence of industry standards also slows the pace of adoption.
TFLN: A New Variable in High-End Segments
Thin-Film Lithium Niobate (TFLN), as a new-generation optoelectronic material technology, is carving out its own niche in the high-end, high-speed optical module sector. Lithium niobate crystal is referred to as the 'optical silicon' within the industry. Its natural low half-wave voltage characteristic allows modulators to be directly driven by the DSP's native low-swing electrical signal, eliminating the need for external high-power driver amplifier circuits.
Commercial breakthroughs have already emerged. A 1.6T-DR8 optical transceiver based on TFLN technology has a total operating power of only 20 watts, representing a 20% reduction compared to traditional solutions of the same specification. It also employs a single continuous-wave laser driver scheme, greatly simplifying the optical path structure and operational difficulty. The head of HyperLight points out that TFLN is a core supporting technology for future 400Gbps-per-lane optical communication systems. It has already demonstrated strong energy-saving capabilities within the current mainstream 200Gbps technology generation.
The Zheshang Securities report concludes that TFLN is not emerging as a disruptive alternative to existing mainstream technologies but rather as a critical complementary technology that fills the gaps of traditional materials in high-performance electro-optic modulation. The future industry will form a landscape where silicon photonics, indium phosphide, and TFLN coexist in parallel, selected based on need. TFLN will firmly occupy niche markets such as high-end, high-speed optical modules and radio-frequency photonic devices. Currently, TFLN has entered a phase of small-scale commercial deployment. As manufacturing processes are optimized and production yields improve, it will gradually penetrate from high-end scenarios to more general-purpose applications.
Comments