In recent years, gigabit optical networks have developed rapidly, becoming a solid all-optical foundation supporting the digital and intelligent transformation of society. According to the economic operation report of the telecommunications industry for the first half of 2026 released by the Ministry of Industry and Information Technology, as of the end of June 2026, the scale of fixed broadband access users in China reached 701 million, with 258 million gigabit users. The number of 10G PON (passive optical network) ports capable of providing gigabit network services reached 32.86 million, demonstrating significant achievements in gigabit network construction. Looking ahead to the "15th Five-Year Plan" period, the construction of 10-gigabit networks is accelerating. It is projected that by the end of 2030, China will build 1 million 50G PON ports, gradually expanding the coverage of 10-gigabit access networks.
Currently, AI (artificial intelligence) technology is comprehensively empowering various industries. Content production and business decision-making increasingly rely on the support of large AI models, and AI applications are showing a trend of diversified rapid development. In August 2025, the State Council issued the "Opinions on Deepening the Implementation of the 'AI+' Action," proposing that by 2030, artificial intelligence in China will fully empower high-quality development, with the penetration rate of new-generation smart terminals, agents, and other applications exceeding 90%. In smart homes, AI relies on home automation systems to achieve appliance linkage and voice control, while also handling daily tasks such as health monitoring, learning assistance, and security alerts, thereby enhancing user comfort and safety.
In the government and enterprise service sector, AI is widely applied in scenarios such as smart office, intelligent quality inspection, automatic document organization, and market analysis and prediction, significantly improving production efficiency and management levels. The rapid development of AI businesses has placed new requirements on optical networks. While optimizing transmission performance such as bandwidth and latency, they must also support new capabilities like intrinsic intelligence and computing power coordination.
As the first-hop entry point of operator networks, the optical access network serves as a critical infrastructure for business-to-computing integration and home/enterprise AI services. This article focuses on the current status of the AI optical access network industry, addresses the trends of smart applications and terminal intelligence, and elaborates on the evolution path and key technologies of optical access networks towards an AI-intrinsic integrated computing-network carrying base. It also analyzes the application practices of intelligent optical access networks in typical scenarios.
Evolution Trends of AI-Intrinsic Optical Access Networks
The evolution of AI-intrinsic optical access networks is not merely a single bandwidth upgrade but a systematic paradigm reconstruction centered on three aspects: enhancement of communication connectivity capabilities, upgrading of intrinsic intelligence capabilities, and coordination of intelligent computing processing capabilities. The technological evolution trends of AI-intrinsic all-optical access networks are shown in Figure 1.
Communication Connectivity: Symmetric Large Bandwidth and Deterministic Enhancement
According to statistics from the National Data Administration, the current daily average Token consumption in China reaches 140 trillion. AI businesses will drive substantial growth in optical network traffic, with AI inference traffic expected to account for 25% by 2035. Additionally, AI businesses will change the traffic patterns of optical access networks. Traditional home broadband services are primarily downstream-oriented, with upstream traffic accounting for only 10%~15%, whereas the upstream traffic proportion for AI businesses is expected to reach 40%~50%. Beyond continuous bandwidth improvement, optical access networks in the AI era are upgrading from the traditional "best-effort" network service model to a "deterministic experience" carrying system that offers commitments and guarantees.
This requires not only reducing end-to-end latency but also imposing rigid requirements on latency stability (low jitter) and reliability assurance. Businesses such as AI industrial control and autonomous driving typically require end-to-end latency to be controlled at the millisecond level and jitter reduced to the microsecond level to ensure real-time response.
Intrinsic Intelligence: Intelligent Evolution of Network Elements
Against the backdrop of expanding traffic scale, increasing operational complexity, and rising device power consumption, optical access networks will deploy AI agents to "agentify" network elements and the entire network domain, enabling real-time dynamic sensing of key network metrics. Specifically, on the network element side, agents will evolve from single analysis modules into intelligent operating entities intrinsic to OLT (optical line terminal), ONU (optical network unit)/FTTR (Fiber to the Room) devices. Based on multi-dimensional real-time data and using device mechanisms and operational knowledge as reasoning bases, these agents perform local AI inference and decision-making, equipping optical networks with capabilities such as autonomous prediction, root cause localization, and intelligent optimization. On the network side, AI-driven businesses will transform the traditional static pre-configured management system into a dynamic intelligent management system driven by both "tasks" and "intentions." This enables unified intelligent scheduling of distributed agents, building a distributed multi-agent system to meet the carrying requirements of AI-type businesses.
Intelligent Computing Processing: Collaborative Evolution of End-Edge-Cloud
The end-side ONU/FTTR serves as a crucial "neural ending" for Token operations. Combining the extreme intelligent experience requirements and ultra-low cost characteristics of the end side, the edge-side OLT becomes the preferred scheduling anchor for intelligent computing processing and the optimal carrying node for sinking cloud-side computing power and models. Computing power coordination centered on the OLT naturally possesses advantages of low latency and local data security. Relying on the nearly hundred-Teraflops computing power deployed on the OLT side, it can support the efficient carrying of ultra-large parameter models under end-edge collaboration. Meanwhile, based on edge-cloud collaboration, the OLT proactively initiates demands for model optimization and knowledge updates to the cloud side, achieving on-demand model updates pushed to OLT nodes, thus forming a closed loop of end-edge-cloud collaborative intelligent computing business processing.
Key Technologies of AI-Intrinsic Optical Access Networks
In the AI era, optical access networks will continue to evolve into comprehensive all-optical foundations featuring intrinsic intelligence and communication-computing fusion. Based on key technologies such as symmetric 50G PON large bandwidth, AI intrinsic intelligence, and end-edge collaborative computing power buses, new types of AI-intrinsic optical access networks will be constructed.
Enhancement of Basic Communication Capabilities
To meet the demands for symmetric large bandwidth and deterministic low latency for AI businesses, next-generation 10-gigabit optical access networks should adopt symmetric 50G PON. Compared to 10G PON, the 50G PON system increases line speed by five times but introduces challenges related to fiber dispersion and nonlinear effects, making compatibility with the existing C+ grade 32 dB power budget of ODN (optical distribution network) difficult. Therefore, it is necessary to significantly improve the performance of optoelectronic devices on the transmitting and receiving sides. Introducing optical devices such as SOA (semiconductor optical amplifier) and new-structure 50G APD (avalanche photodiode) enhances transmission performance. Simultaneously, introducing soft-decision LDPC (low-density parity-check code) FEC (forward error correction) improves bit error tolerance. Furthermore, dynamic adaptive equalization oDSP (optical digital signal processor) solutions are required to achieve differentiated dispersion compensation for25G and 50G rate signals within the same port in the upstream direction. In the downstream transmitter, superimposing low-frequency envelope amplitude modulation onto the DFB (distributed feedback laser) drive current effectively suppresses SBS (stimulated Brillouin scattering) nonlinear damage caused by increased fiber input power.
To address the deterministic requirements of AI businesses, 50G PON should support single-frame multi-burst low-latency functions. By scheduling ONU burst signals multiple times within the 125 μs frame cycle in the upstream direction, business waiting time is reduced. Additionally, regarding the issue where opening registration windows during ONU online status causes latency and jitter for online ONUs, a third wavelength can be introduced for online registration based on Combo terminals. This eliminates registration window opening, reducing the maximum latency of the PON system from the millisecond level to within 200 μs.
Looking to the future,200G PON has become the clear direction for the evolution of 50G PON. However, the generational transition to 200G PON faces numerous technical challenges, the most fundamental being the unification of technical routes, which is currently a key research focus for academia and industry. ITU-T initiated technical pre-research for 200G PON in 2022. After multiple rounds of discussion, three technical directions have been basically identified: direct detection with direct modulation, standard coherent, and hybrid coherent. The direct detection with direct modulation scheme has significant cost advantages but faces risks of failure in 200G PON systems due to current industrial chain limitations. Standard coherent offers the best performance selection, providing ample performance margins; however, how to apply "subtraction" to this technology to balance cost and performance is key to pushing coherent technology down to PON. Hybrid coherent technology combines the advantages of both, implementing 200G PON systems via coherent detection of intensity-modulated formats, making it a promising candidate technical solution.
Key Technologies for AI Intrinsic Intelligence
Optical access networks reconstruct device software and hardware architectures to enable the deployment of intrinsic agents within OLT and ONU/FTTR devices. Considering device cost and power consumption constraints, ONU/FTTR sides typically cannot deploy large amounts of computing power or hardware resources such as memory and video memory. Therefore, computing power within 1 Teraflop is selected for deployment on the end side to support lightweight models, achieving functions such as Wi-Fi (wireless fidelity) tuning and IoT (Internet of Things) device control on end-side devices. On the OLT side, CPU (central processing unit), NPU (neural processing unit), and storage capabilities must be enhanced to support the deployment of hundred-Teraflop computing power and ten-billion-parameter large models, enabling end-side agent invocation and intra-OLT domain optimization.
Access network agent collaboration is built upon the layered computing architecture of end-edge-cloud collaboration, with different network elements assuming differentiated roles. As the edge core agent, the OLT carries hundred-Teraflop computing power and assumes the function of cross-element core scheduling and collaborative decision-making. As terminal-side lightweight agents, ONU/FTTR devices complete tasks such as business sensing, traffic prediction, and low-latency inference locally, reporting information requiring global coordination to the OLT. The cloud platform, acting as the global agent, is responsible for cross-domain strategy orchestration, model training and distribution, and decision-making for complex scenarios. Currently, access network agent collaboration mainly relies on two types of new protocols: the A2A (Agent to Agent) protocol provides a standardized method for direct communication and collaboration between agents, allowing multi-agents to negotiate control strategies through structured dialogue while decomposing complex intentions; the MCP (model context protocol) provides standardized interfaces for agents to access device tools, data, and capabilities. Computing and network resources are abstracted through natural language combined with structured descriptions, adapting to A2A and MCP new communication protocols to build an access network agent collaboration system suitable for AI businesses.
China Mobile completed the demonstration of dynamic energy saving and intelligent operation and maintenance based on AI intrinsic intelligence in live networks in 2026. By learning historical traffic features at the PON port level and relying on AI inference algorithms to predict future network traffic trends, measures such as deep sleep of idle networks, dynamic frequency and voltage adjustment of devices, dynamic speed adjustment of high-speed SerDes, and working mode switching are taken timely based on predictions. This achieves the minimization of comprehensive energy consumption for optical access network devices driven by AI intrinsic intelligence. Through high-frequency collection, real-time reporting of data including ONU, ODN, MAC (medium access control) logs, and OLT is realized. Based on professional operation and maintenance knowledge bases and large O&M models, autonomous sensing, analysis, and judgment of network faults are achieved, outputting standardized repair suggestions, thereby effectively improving the automation level of broadband network O&M and fault handling efficiency.
End-Edge Collaborative Computing Power Bus
To achieve efficient scheduling of computing resources, a PON end-edge computing power bus can be constructed based on extensions of OMCI (ONT management and control interface) and eOMCI (enhanced ONT management and control interface). This fully reuses existing PON device management and control mechanisms, introducing computing resource management and scheduling capabilities without changing the business carrying system, thereby reducing costs associated with new control protocols and network modifications. Meanwhile, through joint sensing of computing power and network status, computing power scheduling no longer relies solely on CPU/NPU load indicators but comprehensively considers factors such as user location, network path, link status, resource load, and business latency. This achieves collaborative decision-making between "computing power selection" and "network selection," balancing determinism, reliability, and system cost.
This technology is mainly used to implement the following functions. First, in the resource scheduling dimension, it achieves unified abstraction of heterogeneous resources such as CPU, NPU, GPU, and storage, solving issues of resource fragmentation and invisibility; it establishes a network-aware computing power scheduling mechanism to achieve proximity computing, dynamic load balancing, and fault rerouting, improving resource utilization efficiency. Second, in the business processing dimension, through computing power and network collaboration, it transforms the traditional "access first, then go to cloud" business model into "data proximity, computing power proximity, service collaboration," reducing data movement and end-to-end inference latency to meet data security and low-latency computing service needs on the edge/end side. Third, in the network evolution dimension, leveraging mature PON device, management, O&M, and security systems allows intrinsic computing power to be incorporated into existing access networks with lower incremental costs, evolving computing power from single-device capability to network-level infrastructure capability.
The computing power bus is not simply adding a set of computing power scheduling software but establishing a unified resource abstraction layer, collaborative control layer, and service orchestration layer for intrinsic computing power in access networks. This evolves "network access" further into "network access + computing power access," extending to the cloud to open up end-edge-cloud collaborative computing-network fusion service capabilities. It meets the high requirements for computing resources and network latency posed by the large-scale application of future agents, supporting operators in transitioning from providing "connection" services to providing comprehensive "connection + computing power + intelligence" services. The architecture of the end-edge-cloud collaborative optical access network computing power bus is shown in Figure 2.
China Mobile completed the live network verification of end-edge-cloud collaborative computing power bus technology in 2026. This enabled AI-FTTR to quickly invoke built-in intelligent computing models in the OLT, completing complex user intent analysis and accurate capture within sub-second timescales. It efficiently scheduled and controlled IoT devices, achieving deep integration and linkage between sensing information and IoT control.
Summary and Outlook
10-gigabit optical networks will continue to evolve into optical access networks integrating communication, sensing, and intelligent computing. By introducing AI intrinsic intelligence and network capabilities connecting "people, things, and computing power," they will support the scaled implementation of Token services in diverse scenarios. In the future, they will continue to evolve towards multi-dimensional integration of "reliable connection + computing power coordination + intelligent services": In terms of basic communication capabilities, symmetric 50G PON will provide large-bandwidth, low-latency transmission channels, evolving towards 200G PON technology; in terms of multi-device agent collaboration, an efficiently interconnected intelligent ecosystem will be built, upgrading the network from a single transmission pipe to an intelligent infrastructure with sensing, decision-making, and collaboration capabilities; in terms of computing power invocation, breaking barriers to end-edge-cloud collaboration through edge computing poolization and computing power bus standardization will build native intelligent capabilities for optical access networks, comprehensively empowering the digital and intelligent transformation of thousands of industries.
This article was published in Communications World, Issue 10, 2026; Original Title: Technical Evolution of Next-Generation AI-Intrinsic 10-Gigabit All-Optical Access Networks.
Authors: Wang Ning, Jia Xiaoshuo, Hu Leya, Li Junwei, Zhang Dechao, Research Institute of China Mobile Communications Corporation Limited
Editor/Layout: Zhu Wenfeng
Reviewer: Mei Yaxin
Supervisor: Liu Qicheng