Currently, AI has entered a phase of large-scale application. New technologies and businesses such as large models, agents, scientific computing, and intelligent manufacturing are iterating rapidly, driving the accelerated expansion of computing infrastructure. As of the end of June 2026, China's intelligent computing power scale reached 2185 EFLOPS (FP16), a year-on-year increase of 177%; 52 intelligent computing facilities with over 10,000 cards have been built. Computing infrastructure is accelerating its evolution towards a three-tier collaborative layout system of "hub-region-edge".
The demand for computing power in large model training and inference continues to grow. Computing resources within a single data center can no longer meet business development needs. Cross-data center, cross-campus, and even cross-regional computing resource scheduling and coordination have gradually become an important direction for industry development. Computing resources are evolving from mutually independent computing nodes into a networked resource system that is interconnected, schedulable, and collaborative.
The large-scale application of AI drives the rapid expansion of computing infrastructure and cross-regional computing coordination. Optical networks are evolving from traditional transmission pipes to AI computing infrastructure. Their connectivity, scheduling, and assurance capabilities directly impact computing efficiency. Meanwhile, optical networks will accelerate the comprehensive integration of communication, sensing, and computing elements, laying a solid foundation for the large-scale application of AI.
This article focuses on the three major characteristics of all-optical networks: "New Media," "New Systems," and "New Control." It analyzes the development trends of key technologies such as hollow-core/multi-core fibers, ultra-high-speed optical transmission, WSON (Wavelength Switched Optical Network), OCS (Optical Circuit Switching), digital twins, and computing-network synergy. Combined with China Telecom's practical experience with its All-Optical Network 3.0 strategy, it proposes a phased evolution path for the next decade.
01Changes in Business Forms in the AI Era Drive the Accelerated Evolution of Next-Generation Optical Networks
Business forms are a significant driver of network evolution: the Internet era promoted the widespread adoption of fiber optic and mobile networks; the cloud computing era facilitated cloud-network convergence; and the AI era is driving network infrastructure further towards the comprehensive integration of computing, communication, storage, sensing, and control elements.
The new requirements AI imposes on networks exhibit distinct systemic characteristics.
First, connected objects are continuously evolving. Currently, network connections are expanding from traditional person-to-person and person-to-information interactions to diverse entities such as cloud centers, computing clusters, edge nodes, smart terminals, machines, sensors, and agents.
Second, business space is constantly expanding. Computing resources and business nodes are no longer limited to data centers but extend continuously into metropolitan, regional, and backbone networks, and further into multi-dimensional spaces including air, sky, ground, and sea.
Third, performance requirements are steadily improving. Businesses such as AI training and inference, and multimodal interaction have significantly higher demands for bandwidth, latency, and stability. Especially in large-scale distributed training scenarios, massive parameter synchronization and data exchange between nodes require networks to possess high throughput, low jitter, and low packet loss capabilities; in mission-critical scenarios, networks are also required to provide fast fault recovery and differentiated SLA (Service Level Agreement) assurance.
Fourth, functional elements are increasing. The core function of traditional optical networks was information transmission. In the AI era, networks need to further integrate sensing, computing, intelligence, control, and storage capabilities. Networks must not only transmit data but also sense business status, understand business intent, and dynamically adjust resource supply based on business needs.
Against this backdrop, optical networks, as a crucial component and key carrier base of new infrastructure, are entering a new stage of coordinated technological and industrial development. Facing the intelligent era, China Telecom proposed the "All-Optical Network 3.0" development strategy centered on "All-Optical Intelligent Connectivity," clarifying three visions: "Integration of Optics, Cloud, and Intelligence," "Integration of Optics, Sensing, and Business," and "Integration of Sky, Ground, and Sea." This essentially aims to upgrade optical networks from independent basic carrier infrastructure to a comprehensive capability system deeply integrated with cloud, computing, and business.
02All-Optical Network 3.0: From "Connection Network" to "Carrier for AI Computing Power"
All-Optical Network 1.0 was characterized by "fiber replacing copper," focusing on solving fiber coverage and access issues. All-Optical Network 2.0 centered on "all-optical switching," using technologies like ROADM (Reconfigurable Optical Add/Drop Multiplexer) to drive networks towards high capacity, flattening, and full opticalization. Entering the AI era, All-Optical Network 3.0 centers on "All-Optical Intelligent Connectivity," with further expansion of network connection objects, carrying scenarios, and functional boundaries.
From a technical perspective, the capability evolution of All-Optical Network 3.0 for the AI era can be summarized into three aspects: "New Media," "New Systems," and "New Control," as shown in Figure 1.
"New Media" primarily addresses capacity, distance, density, and latency issues in optical networks. Through innovations in hollow-core fibers, multi-core fibers, high-fiber-count cables, and iterative evolution of new optical network architectures, it continuously expands the physical boundaries of optical connectivity.
"New Systems" primarily address how optical networks carry large-scale AI services. By fully utilizing ultra-high-speed, ultra-large-capacity transmission, ROADM/WSON, OCS, and opto-electronic collaborative computing technologies, it drives optical networks to evolve towards ultra-large bandwidth, ultra-high reliability, ultra-large scale, and ultra-low latency.
"New Control" primarily addresses how network resources are efficiently utilized. Through technologies such as digital twins, AI, agents, and open control management, it achieves intelligent lifecycle management of "Planning—Construction—Maintenance—Optimization—Operation," pushing optical networks from manual operation and passive response to proactive sensing, intelligent decision-making, and closed-loop execution.
These three aspects jointly constitute the technical system of next-generation optical networks. "New Media" expands boundaries, "New Systems" release value, and "New Control" enhances efficiency. Ultimately, this upgrades optical networks from traditional "information transmission pipes" to comprehensive infrastructure bases carrying AI computing power.
From an international perspective, major global cloud providers and telecom operators are accelerating their optical network layouts for AI services. Google Cloud launched the Virgo Network for large-scale AI training, exploring cross-data center, high-bandwidth, low-latency network architectures. NTT (Nippon Telegraph and Telephone Corporation) relies on all-photonics networks to advance data center interconnection and distributed computing. Overall, cloud providers focus on collaborative design of computing and networking, while operators emphasize strengthening cross-regional optical connectivity and scheduling capabilities. Both parties jointly drive the deep coupling of optical networks and AI services.
New Media: Evolving from Expanding Fiber Capacity to Breaking Optical Connection Boundaries
The large-scale deployment of AI services drives rapid growth in optical network capacity and connection density. Traditional single-mode fibers are gradually revealing performance bottlenecks in terms of capacity, latency, and duct space resources, necessitating coordinated evolution at three levels: fiber medium, cable density, and network architecture.
In terms of fiber medium, hollow-core fibers reduce light propagation delay in the medium, providing a new technical path for latency-sensitive applications such as financial trading and data center interconnection. Multi-core fibers utilize spatial dimensions to enhance single-fiber capacity, particularly suitable for scenarios with limited spatial resources such as inside data centers and submarine cables. High-fiber-count optical cables alleviate duct resource constraints by increasing the number of fiber cores per cable unit and enhancing transmission capacity.
From the perspective of international development and industrial ecosystem, new optical media are evolving from breakthroughs in fiber body performance to system adaptation and large-scale engineering applications. Microsoft has deployed hollow-core fibers in Azure networks for metropolitan data center interconnection. NTT is advancing research on splicing, branching, and testing technologies for four-core multi-core fibers, exploring their application in data center interconnection and submarine cables. However, hollow-core fibers still face challenges such as gas absorption loss, heterogeneous structure fusion, and testing/operation maintenance. Multi-core fibers need to resolve issues regarding inter-core crosstalk, supporting devices, and engineering splicing.
In terms of standardization, ITU-T SG15 (International Telecommunication Union Telecommunication Standardization Sector Study Group 15) has initiated the GSTR.hcf study report on hollow-core fiber technology, advanced the formulation of standards for weakly coupled multi-core fibers such as G.smmcf and G.csmcf, and continued to improve the single-mode fiber standard system including G.652, G.654, and G.657. Future efforts should strengthen collaboration among industrial segments such as optical fibers/cables, optical devices, transmission equipment, and test instruments. Robust performance testing, system adaptation, and engineering application specifications should be established to drive new optical media from performance verification to large-scale application.
Regarding new fiber applications, China Telecom actively conducts live network practices. For hollow-core fibers, in 2025, China Telecom deployed the world's longest cross-border hollow-core fiber and ultra-low latency transmission system on the Dongguan-Hong Kong 110 km route, reducing mutual access latency between Guangdong and Hong Kong exchanges to within 1 ms. Collaborating with industry partners, it completed a demonstration of ultra-long span unamplified WDM transmission technology on the live network, achieving 51.3 Tbit/s transmission capacity over a 206.5 km link without repeaters, setting a new record for capacity-distance product in similar live network scenarios. For multi-core fibers, China Telecom completed the first-ever live network validation of cross-city multi-intelligent-computing-center distributed training based on multi-core fibers in Guangzhou-Shenzhen, with an interconnection distance of 409.61 km and business performance reaching over 97% of centralized training.
Facing the construction of the "Eastern Data Western Computing" project and the national integrated computing network, China Telecom continues to promote the construction of the "New Eight Vertical and Eight Horizontal" backbone optical cable network. By densifying networks between the eight major hubs, enriching node exits, optimizing routing, promoting the integration of primary and secondary trunk lines, and deploying G.654.E fibers, it constructs a new backbone optical cable infrastructure featuring high reliability, low latency, and high-density interconnection for computing nodes. Meanwhile, China Telecom accelerates engineering technology breakthroughs and live network validation for operations and maintenance/splicing of high-fiber-count optical cables, promoting their application in backbone networks.
New Systems: Evolving from High-Speed Transmission to Integrated Connectivity Inside and Outside Computing Power
First, the high-speed, large-capacity transmission system continues to upgrade. 400G has become a mature commercial technology in backbone networks; 800G is accelerating pilot and commercial deployment on live networks; and 1.6 Tbit/s is the direction for forward-looking layout in the next phase. Additionally, technologies such as multi-band WDM (Wavelength Division Multiplexing), higher-order modulation, DSP (Digital Signal Processor), and FEC (Forward Error Correction) are driving single-fiber capacity towards over 100 Tbit/s.
Second, all-optical switching capabilities are penetrating deeper into computing scenarios. ROADM provides wavelength-level flexible connectivity; WSON further offers fast recovery capabilities at the optical layer; OCS extends optical switching from traditional transmission networks to inside data centers and even metropolitan-level computing clusters, providing high-reliability, low-latency, and dynamic optical connectivity for AI services.
From an international and industrial perspective, AWS (Amazon Web Services) expands long-distance transmission capabilities through self-developed DWDM (Dense Wavelength Division Multiplexing) systems. NTT advances on-demand configuration of Tbit/s-class optical channels relying on IOWN (Innovative Optical and Wireless Network). Verizon conducts verification of single-wavelength 1.6 Tbit/s metropolitan optical transmission.
In the supply chain, optical chips are evolving towards higher rates and higher integration, but high-speed modulators and detectors still face constraints in bandwidth, yield, and cost. Optical modules are accelerating upgrades to 800G and 1.6T, constrained by coherent DSP power consumption, advanced packaging, and heat dissipation capabilities. The contradiction between high rates and transmission distance/system energy efficiency is becoming increasingly prominent. Transmission equipment is developing towards multi-band, ultra-large capacity, and flexible optical switching, but broadband optical amplifiers, WSS (Wavelength Selective Switch), and optical layer power management constrain system capacity expansion. OCS still faces challenges in switching scale, control coordination, and deployment costs.
In terms of standardization, IEEE P802.3dj is advancing physical layer specifications for Ethernet from 200G to 1.6T. OIF (Optical Internetworking Forum) has released implementation agreements for 800ZR and 1600ZR. ITU-T SG15 is conducting studies on TR.ION-aiDC, Beyond 1T, and high-speed FlexO interfaces. Future efforts should strengthen collaborative design of optical chips, modules, and transmission equipment, refine system interfaces and multi-vendor interoperability standards, and drive optical systems from single-point technical breakthroughs to end-to-end collaboration.
Centering on high-reliability all-optical connectivity, China Telecom collaborated with industry partners to optimize routing calculation, control protocols, WSS, and OTU (Optical Transponder Unit) links, constructing multi-node MESH WSON with 50 ms deterministic recovery capability. Laboratory validation with 8 nodes and live network validation with 4 nodes were completed, providing a high-reliability all-optical connectivity solution for distributed computing interconnection. At the same time, China Telecom explores mapping AI communication topologies to the metropolitan optical layer. By constructing dynamic logical topologies via fiber ring networks, ROADM, and OCS, multiple AI data centers form metropolitan-level intelligent computing clusters: parallel wavelength channels adapt to AllReduce aggregate communication; ROADM optical pass-through and dynamic optical connectivity construct low-latency Mesh topologies for many-to-many communications such as All-to-All and MoE (Mixture-of-Experts), gradually making optical networks part of the computing cluster topology.
In the future, optical computing will promote "transmission-computing fusion." Through opto-electronic collaboration, optical signal processing, and photonic AI acceleration, it explores undertaking highly parallel, low-power computing tasks in networks or at the edge, expanding network functions from "Transmission—Switching" to "Transmission—Switching—Computing."
New Control: Evolving from Network Automation to Computing-Network Synergistic Intelligence
Facing the large-scale operational needs of high-speed optical networks, New Control relies on digital twins, AI, and agents to achieve full lifecycle management moving from passive O&M to proactive sensing, intelligent decision-making, and closed-loop execution.
In the planning and construction phase, digital twins map physical networks, simulating and optimizing routing, expansion, and relay sites. China Telecom has applied new technologies to 400G ROADM expansion planning to improve construction efficiency.
In the network maintenance phase, combining data-and-rule-driven fault diagnosis and dynamic benchmark adaptive inspection agents enables precise fault localization and performance inspection. Related capabilities have been deployed on China Telecom's live network.
In the network optimization phase, leveraging fiber sensing and AI capabilities enables early warning of external cable damage, same-route hazards, and submarine cable anomalies. Through gray-box digital twin transmission quality assessment and AI cascade models, network performance improved by approximately 30%. Related solutions have completed live network validation.
In the business operation phase, AI accurately identifies business characteristics and predicts bandwidth demands, scheduling resources and ensuring SLAs according to business priority, driving the transformation from "leased line operation" to intelligent services with deterministic business experiences.
From an international and industrial perspective, optical network control is evolving from single-domain automation to digital twin, agent-driven, and cross-domain collaboration. NTT, together with industry partners like Orange and Telefonica, conducts automatic configuration of Tbit/s-class optical channels and multi-vendor interoperability validation based on open interfaces and digital twins. In terms of standardization, OIF has initiated studies on optical network digital twins and enhanced network operations. ITU-T SG15 has completed ION-2030 and GSTR.ENO technical reports and is advancing standardization work on ENO architectures. However, issues such as inconsistent data models for heterogeneous devices, insufficient reliability of AI decisions, and inadequate cross-layer/cross-domain interface coordination still constrain the large-scale application of intelligent control capabilities. Future efforts should strengthen the construction of standard systems for unified resource models, open control interfaces, and intelligent capability evaluation, promoting multi-vendor collaboration and full-lifecycle intelligent control.
Facing AI computing services, the evolution focus of New Control will expand from the intelligence of optical networks themselves to computing-network synergistic intelligence. By establishing a collaborative mechanism between computing service intent, network resource status, and optical layer execution capabilities, it translates the requirements of model training and inference tasks for bandwidth, latency, reliability, and connection topology into executable network scheduling strategies. Dynamic optimization is achieved through business quality feedback, making optical connection resources a perceivable, callable, and orchestratable foundational capability for computing platforms.
Evolution Path of All-Optical Computing Infrastructure Base
"New Media, New Systems, and New Control" jointly constitute the core of the all-optical computing infrastructure base. Based on technological development and application needs, they present a phased evolution trend, as detailed in Table 1.
Table 1 Phased Evolution Path of All-Optical Computing Infrastructure Base
| Evolution Stage | By 2030 | By 2035 |
| New Media | Small-scale application of new fibers such as hollow-core and multi-core fibers;
Accelerate the popularization of G.654.E; Backbone optical cable networks basically complete the construction of efficient direct-connect optical cable networks |
On-demand large-scale deployment of hollow-core fibers in hotspots with low-latency requirements; Pilot high-fiber-count optical cables based on single-mode fiber with over 500 cores;
Complete the construction of efficient direct-connect (low-latency) and strategic base (wide-coverage) trunk optical cable networks |
| New Systems | Backbone Network: Introduce 800 Gbit/s channels in hotspot areas; Achieve multi-rate converged ROADM all-optical switching network;
Metro Network: Introduce 800 Gbit/s or 1.6 Tbit/s capabilities between DCs; Introduce 50 ms telecom-grade WSON recovery capability |
Backbone Network: Achieve wide coverage of 400 Gbit/s; Gradually deploy 1.6 Tbit/s; Simultaneously promote C+L+S band innovation and gradual commercialization;
Metro Network: Complete flattened convergence of core/aggregation layers in large and extra-large cities; Grid networking; Access layer evolves towards grid networks |
| New Control | Fully support management of optical cable quality, transmission performance, and business SLA;
Achieve millisecond-level sensing, minute-level fault tracing, and minute-level service provisioning |
Optical networks fully possess L5-level autonomous optical network capabilities; Achieve end-to-end autonomy across the "Plan-Build-Maintain-Optimize-Operate" lifecycle;
Possess cross-domain twin and self-evolution capabilities; Dynamically adapt to unknown business scenarios and external environmental changes |
03Summary, Initiatives, and Outlook
In summary, from "Fiber Replacing Copper" to "All-Optical Switching," and then to "All-Optical Intelligent Connectivity," optical networks facing the AI era possess three basic characteristics.
First, optical networks are evolving from transmission pipes to the core infrastructure base for AI computing power. Network boundaries inside and outside computing centers are weakening. Optical networks have deeply covered internal scenarios of intelligent computing centers and cross-center interconnections, and have begun participating in computing resource scheduling. Optical networks are no longer just an external carrier base for computing power but have become an indispensable foundational capability of the computing system.
Second, the value of optical networks is extending from "single pipe" to "more deterministic, more stable differentiated connectivity capabilities." 800G, 1.6T, and multi-band transmission enhance computing interconnection capacity; WSON and tiered SLA assurance guarantee computing business continuity; OCS enhances computing topology elasticity; and computing-network synergy improves resource scheduling capabilities.
Third, optical networks will evolve from "single connection" to the integration of communication, sensing, intelligence, computing, control, and storage capabilities. Competitiveness is no longer determined solely by transmission speed but depends on the coordinated development of New Media, New Systems, and New Control.
The large-scale implementation of the all-optical computing infrastructure base still faces multiple practical challenges. Technologically, key technologies such as engineering O&M of new fibers like hollow-core, ultra-high-speed coherent optical modules above 1.6T, and optical switching/computing-network synergistic control are not yet fully mature. Capabilities such as cross-domain computing-network twins and business intent perception still require continuous iteration. Industrially, technical interfaces across supply chain segments are inconsistent; costs for new fibers and high-speed devices remain high; equipment adaptation to computing services is insufficient; and engineering practice experience for large-scale live network deployment is relatively lacking. Policy-wise, the coordinated planning mechanism for computing networks and optical networks needs improvement. Guarantees for infrastructure resources such as optical cable ducts and hub sites for intelligent computing interconnection, along with supporting policy supplies, need strengthening.
In view of this, this article issues three action initiatives to all parties in the industry.
Initiative One: Accelerate the Construction of the Standard System for All-Optical Computing Infrastructure Bases. Centering on the three major technical directions, comprehensively promote the formulation of standards in key areas such as new fibers, high-speed optical transmission, all-optical switching, computing-network synergistic control, and business SLAs. Unify technical interfaces, O&M specifications, and performance evaluation systems. Leverage the advantages of domestic operators' live network applications, collaborate with the supply chain for joint verification, promote the conversion of autonomous technological achievements into international standards, and enhance China's influence in international standard-setting and industrial collaboration.
Initiative Two: Promote Collaborative Breakthroughs in Key Core Technologies through "Industry-Academia-Research-Application" Synergy. Gather forces from operators, research institutes, equipment vendors, and intelligent computing construction entities. Focus on "bottleneck" and forward-looking technologies such as engineering of hollow-core fibers, ultra-high-speed optical devices, dynamic all-optical switching, computing-network digital twins, and optical computing. Build joint experimental verification platforms. Strengthen full-chain breakthroughs from basic research and prototype development to live network pilots, promoting autonomous controllability and mature implementation of core technologies.
Initiative Three: Construct an Open and Win-Win Industrial Ecosystem. Break down barriers between the computing industry and the optical communication industry. Promote deep collaboration between computing, networking, and AI services. Encourage cross-industry innovative cooperation models. Cultivate new services integrating computing and networking. Seize the opportunity of the "Six Networks" construction, guide all parties in the supply chain to participate collaboratively in building all-optical computing infrastructure bases, achieving co-research of technology, sharing of results, and co-creation of value.
This article was published in *Communications World*, Issue 10, 2026; Original Title: "Evolution and Practice of All-Optical Computing Infrastructure Bases in the AI Era."
Authors: Li Junjie, Liu Yuyang, Gong Yadong, Research Institute of China Telecom
Editor/Layout: Zhu Wenfeng
Proofreader: Mei Yaxin
Supervisor: Liu Qicheng