New Paradigm of Optical Interconnection Under the Computing Power Revolution

The explosive growth of AI large models and general computing power is driving the rapid upgrade of data center interconnection bandwidth from 800G to 1.6T, 3.2T and even 6.4T. Traditional pluggable optical modules are approaching their physical limits in three core dimensions: power consumption control, signal integrity and port bandwidth density. The transmission loss of electrical signals at a single-channel rate of 200Gbps and above increases sharply on PCB copper cables, requiring equalization compensation by high-power SerDes chips. As a result, the power consumption of optical interconnection accounts for more than 30% of the total power consumption in data centers.

Against this backdrop, three mainstream technical routes—CPO (Co-packaged Optics), NPO (Near-packaged Optics), and XPO (eXtra-high density Pluggable Optics)—have successively advanced to the forefront of industrialization. Adopting respective technical paths of deep optoelectronic co-packaging, short-distance board-level integration, and high-density pluggable innovation, they provide differentiated solutions for next-generation ultra-high-speed and low-power optical interconnection. The three technologies are not in a simple substitution relationship; instead, they complement each other in application scenarios, cost and technological maturity, jointly supporting the continuous evolution of computing power networks toward higher bandwidth and higher energy efficiency.

I. Three Core Technologies

1.1 CPO (Co-Packaged Optics)

CPO is recognized globally as the ultimate form of optical interconnection technology. Its core logic adopts advanced 2.5D/3D packaging technology to integrate optical engines with switch ASIC/GPU/XPU chips on the same substrate or interposer. This fundamentally restructures the optical-electrical signal transmission link, shortening the electrical signal transmission path from over 100 millimeters in traditional solutions to the millimeter level, and eliminating loss and latency caused by long PCB traces at the source.

Its typical architecture adopts a layered heterogeneous stacking design: the top layer is the computing/switch chip layer, responsible for data operation and high-speed signal processing; the interposer adopts 2.5D CoWoS/RDL technology to realize short-range chip-to-chip interconnection between chips and optical engines; the optical engine layer integrates silicon photonic chips, modulators and other components to complete optical-electrical conversion and eliminate the need for high-power DSP chips; the bottom layer consists of optical fiber arrays and optical interface layers, enabling seamless connection of internal and external links.

CPO’s core advantages lie in energy efficiency, bandwidth and reliability: compared with traditional 800G DSP optical modules, the power consumption per 800G bandwidth is only 4-5W, with a maximum energy saving rate of 73% and a 30%-50% reduction in system power consumption; the shoreline density reaches 10Tbps/mm, 25 times that of traditional SerDes, supporting a single-channel bandwidth above 3.2Tbps; Meta’s actual tests show its MTBF (Mean Time Between Failures) reaches 2.6 million hours, more than 3 times that of traditional pluggable modules.

Its large-scale deployment still faces challenges: integrating silicon photonic devices under the 3nm process relies on advanced packaging platforms such as TSMC’s COUPE; heterogeneous integration of multiple chips leads to low yield and high costs; the fixed integration of optical engines and main chips requires replacement of the entire board card in case of faults, raising operation and maintenance costs; fragmented industry standards result in poor cross-vendor compatibility and a closed ecological landscape.

1.2 NPO (Near Package Optics)

NPO serves as a core transitional solution for CPO evolution, with the core concept of "proximity without tight integration". The optical engine is decoupled from the ASIC chip and placed on the same PCB motherboard close to the SerDes port, reducing the electrical signal transmission path to the centimeter level. It improves performance while maximally retaining industrial chain compatibility and operation & maintenance flexibility, acting as the optimal smooth evolution path from traditional pluggable solutions to CPO.

Essentially a board-level integrated optical engine solution, NPO retains the pluggable feature of optical engines for independent disassembly and replacement. Electrical signals are transmitted through short-range direct drive of SerDes on the PCB, mostly adopting a linear direct-drive architecture without external DSP chips, which greatly reduces link power consumption and latency. Unlike CPO that focuses on the switch side, NPO is more adapted to intra-cabinet and inter-board interconnection scenarios of GPU accelerator cards and AI super nodes, making it a core transitional solution for Scale-up scenarios.

Its core advantage is achieving a triangular balance of "performance, cost and operation & maintenance": compared with traditional pluggable solutions, power consumption is reduced by 40%-50% and latency drops to the nanosecond level, close to CPO performance; the decoupling of optical engines and ASIC chips supports independent replacement, adapts to the existing operation and maintenance system, reuses the mature industrial chain, and requires no disruptive production process transformation; it covers a wide range of scenarios, supporting both horizontal expansion on the switch side and short-range interconnection within GPU clusters and HPC system cabinets.The core limitation of NPO lies in its positioning as a transitional solution: in ultra-high-speed scenarios, the electrical signal transmission path remains relatively long, posing challenges for signal loss compensation; unified industry packaging and interface standards have not yet been established, leading to insufficient compatibility of vendor solutions and hindering large-scale commercial application.

1.3 XPO (eXtra-dense Pluggable Optics)

XPO is an emerging technical route that rose to prominence at the 2026 OFC Conference. Led by Arista, 45 industrial chain enterprises jointly released a white paper and established the XPO MSA Alliance. It is strategically positioned as "the ultimate form of pluggable architecture". While retaining the operation and maintenance advantages of pluggable designs, it achieves an exponential leap in bandwidth density through architectural and heat dissipation innovation, serving as the optimal deployment path for Scale-out scenarios before the maturation of the CPO ecosystem.

Its core innovations are concentrated in three aspects:

First, ultra-high-density channel integration. A single module embeds 64 lanes of 200Gbps PAM4 channels, delivering a total bandwidth of 12.8Tbps — 8 times that of traditional 1.6T OSFP modules — with future potential to exceed 25.6Tbps.

Second, an "Oreo-style" sandwich liquid cooling architecture. It adopts a buckled dual-PCB paddle board design with an integrated liquid cooling plate sandwiched in the middle; core heat-generating components fit closely to the cooling plate, delivering a heat dissipation capacity of 400W and lowering the temperature by 20-25℃ compared with air cooling.

Third, optimized mechanical structure and power supply. A wrench-style release tab facilitates plugging and unplugging; 48-50V high-voltage DC input reduces current and line loss.

XPO boasts remarkable core advantages: fully compatible with the existing operation and maintenance system, it supports on-site hot swapping without modifying switch architecture or computer room processes, resulting in low deployment barriers; it complies with the full range of optical interconnection standards, covering all scenarios including inter-cabinet and inter-data center connections, with far higher adaptability than NPO and CPO; it can reuse existing industrial chain resources such as silicon photonic chips and DSPs without relying on advanced packaging platforms, enabling controllable mass production costs and yield, with large-scale commercial use expected in 2027.

Its core challenges are as follows: the optical engine is still located at the edge of the PCB, far from the ASIC chip, requiring high-power SerDes and ultra-low-loss PCB to compensate for signal loss, leading to lower energy efficiency than NPO and CPO; the native liquid cooling design imposes mandatory requirements on data center liquid cooling infrastructure, increasing deployment costs and system complexity.

II. Multi-dimensional Comparison of NPO, CPO and XPO Technologies

The three major technical routes adopt different architectural designs and evolution logics, showing significant differences in core performance, deployment cost, ecological maturity and other dimensions. The detailed comparison is as follows:

III. Industrial Implementation Status and Scenario Adaptation Logic

At present, the three major technical routes have formed a clear division of labor and adaptation pattern, coexisting in a complementary manner, and achieving differentiated implementation based on diverse customer demands and deployment scenarios.

3.1 CPO: Leading Manufacturers Take the Lead in Pilots with Gradual Breakthroughs

CPO implementation is concentrated among leading enterprises with solid technical strength and robust ecological closed-loop capabilities, focusing on core nodes of ultra-large-scale AI computing power clusters.

Internationally, NVIDIA plans to deliver InfiniBand and Ethernet CPO products in 2026, and its Rubin Ultra GPU super-node has adopted the CPO solution; Broadcom’s Bailly switch chip has been delivered, with the supporting CPO solution expected to enter mass production in the second half of 2026; Meta has completed reliability verification for CPO switches.

Domestically, Ruijie Networks and H3C have showcased CPO switch solutions; Accelink Technology, InnoLight Technology and other enterprises have completed the development of CPO optical engine samples, gradually keeping pace with global progress.

3.2 NPO: Accelerated Verification of Transitional Solutions with Highlighted Scenario Adaptability

Boasting a balanced advantage in performance and cost, NPO has become the core solution for intra-cabinet interconnection of AI servers and GPU accelerator cards, and has entered the testing and verification phase among leading cloud vendors.

Google’s TPU v9 cluster has adopted NPO technology; domestically, Accelink Technology has launched the world’s first 3.2T silicon photonics NPO module and completed full-system verification with leading cloud vendors; HG Genuine, Higon Core and other enterprises have also released relevant products, driving the industrial implementation of NPO technology from laboratory research to engineering application; equipment manufacturers including Huawei and ZTE have rolled out NPO prototypes, focusing on the board-to-board interconnection scenario of AI servers.

3.3 XPO: Rapid Follow-up Across the Industrial Chain with Clear Commercial Prospects

With perfect compatibility with the existing ecosystem, XPO has gained widespread support from the entire industrial chain since its launch, emerging as a highlight in the optical interconnection industry in 2026.

Arista has collaborated with vendors including Marvell and Coherent to complete XPO standard formulation and prototype development. Domestically, as a founding member of the XPO MSA, HG Genuine has globally launched the first 12.8T XPO module. InnoLight Technology, Eoptolink, HG Genuine and other enterprises have displayed XPO prototype solutions at OFC 2026. Sample verification is expected to be finished in the second half of 2026, with large-scale commercialization set to be realized in 2027.

IV. Future Technological Evolution Trends and Industry Outlook

4.1 Long-term Coexistence of Technical Routes with Scenario-based Adaptation as the Core

In the next 3 to 5 years, the three major technical routes will form a tripartite stalemate pattern: XPO will take the lead in large-scale commercial application for horizontal expansion scenarios in ultra-large-scale data centers and become the mainstream in the pluggable market in the 1.6T/3.2T era; NPO will dominate the short-range interconnection scenarios inside AI server cabinets and serve as a core transition solution before CPO matures; CPO will be first deployed in ultra-large-scale closed AI clusters and core nodes of ultra-high-density switches. With the unification of standards and declining costs, it will gradually penetrate all scenarios and evolve into the ultimate long-term development direction.

4.2 Accelerated Standardization Process with Ecological Collaboration as the Key

Standard fragmentation is the core bottleneck restricting the three major technical routes at present. In the future, international organizations such as IEEE and OIF will accelerate the formulation of standards for packaging, interfaces and testing to promote cross-vendor compatibility. Among them, XPO advances the fastest driven by the MSA alliance and will take the lead in realizing collaborative mass production across the entire industrial chain; CPO will gradually shift from closed ecosystems of leading manufacturers to openness and standardization, lowering the threshold for large-scale implementation.

4.3 Deepened Technology Integration with Optoelectronic Collaborative Design as the Innovation Direction

Future innovation of the three major technical routes will further deepen toward optoelectronic collaborative design and full-link optimization: CPO and NPO will drive the integration of devices such as lasers and modulators to realize the collaborative design of photonic and electronic chips; XPO will optimize liquid cooling and high-speed signal design to narrow the energy efficiency gap with CPO. All three routes will achieve in-depth integration with silicon photonics, advanced packaging and liquid cooling technologies, enabling optical interconnection to evolve from a communication supporting facility into the core infrastructure of AI computing power clusters.

4.4 Opportunities for the Domestic Industrial Chain to Achieve Overtaking in a New Track

Domestic manufacturers already hold a major global market share in the traditional pluggable optical module market. The technological transformation of NPO, CPO and XPO creates opportunities for domestic players to achieve leapfrog development on a new track. Domestic enterprises have made technological breakthroughs in core links such as optical engines, silicon photonic chips and high-speed connectors, and are gradually integrating into the global mainstream supply chain. With the unification of industry standards and market scaling up in the future, domestic manufacturers are expected to occupy a more pivotal position in the next-generation optical interconnection market relying on mature manufacturing capabilities and complete industrial chain supporting advantages, and support the independent and controllable development of China’s AI computing power infrastructure.

Conclusion

The sustained demand for AI computing power drives optical interconnection technology to evolve from traditional pluggable modules into three new technical routes: NPO, CPO and XPO, accomplishing a paradigm revolution from board-level interconnection to chip-level interconnection. There is no absolute superiority or inferiority among the three routes, which only differ in scenario adaptation: XPO addresses the current demand for large-scale commercial deployment, NPO offers a smooth evolution path, and CPO defines the ultimate future form. In the future, with technological iteration and ecological maturity, the three routes will jointly advance data center optical interconnection toward higher density, lower power consumption and better reliability, laying a solid network foundation for breakthroughs in large AI models and general computing power.