In recent years, demand for computing power has surged. Survey data indicates that the annual growth rate of AI (Artificial Intelligence) training computing power in China reached 35%, while the annual growth rate of inference computing power was as high as 42%. Business scenarios are gradually diverging into lightweight scenarios represented by assisted Q&A and core production scenarios such as risk control scheduling and engineering design. Core production scenarios directly involve enterprise intellectual property, imposing stringent requirements on computing elasticity, data security, and network latency.
Behind the explosion in computing demand, a profound transformation is occurring on the industrial supply side. China’s "Eastern Data, Western Computing" project is progressing steadily, establishing a three-tier computing network deployment system comprising "hub-level—regional-level—edge-level." Nationwide, 147 hub-level intelligent computing DCs (Data Centers), 230 regional-level intelligent computing DCs, and 1,916 edge-level intelligent computing nodes have been built, forming a spatial layout of eight major computing hubs and ten major computing clusters. While computing resources have been preliminarily deployed nationwide, computing resources do not equate to usable computing-network services. Where is the computing power? How can it interconnect? How can it be used securely and efficiently? These questions constitute the three core challenges currently constraining industry digital transformation.
Under traditional models, enterprises primarily acquire computing power through two paths: self-built infrastructure or public cloud leasing. Self-built computing is an asset-heavy model requiring massive upfront investment; facility upgrades take 3–6 months, and tidal fluctuations in computing demand easily lead to idle hardware waste. Public cloud leasing offers smooth investment and elastic scaling but faces widespread concerns that "uploading data to the cloud equals surrendering core competitiveness." As the main force in communication infrastructure construction, telecom operators possess multi-network resources including basic communication networks, industry-specific private networks, and computing networks. They have nationwide coverage of AIDCs (AI Data Centers) and OTN (Optical Transport Networks). Operators have the capability to connect the entire chain of computing supply, network connectivity, and secure usage, building a Computing Power Intelligent Network tailored for the Token era.
Analysis of Contradictions Between Computing Supply Landscape and Industry Usage Realities
Positioning and Industrial Practice of the Three-Tier Computing Network
The national computing network has formed a three-tier hierarchical system. Nodes at different levels serve differentiated functional roles and cater to distinct user groups.
Hub-level computing relies on the eight national computing hubs, serving as the core layer for national-level scheduling and hosting. Typical hub computing centers achieve IT room power capacities up to 1 GW with an 85% green electricity ratio. They serve top-tier internet companies and large telecom operators by handling multimodal large model training tasks. Leveraging abundant western green energy resources, they facilitate the green implementation of high-energy-consumption large model operations. Regional-level computing corresponds to the ten computing clusters and provincial computing centers, primarily serving provincial governments, state-owned enterprises, and mid-to-large internet firms. Some provincial intelligent computing centers utilize local hydropower resources to provide 1,000 PFLOPS of intelligent computing capability, offering AI inference services to research institutions and enterprises. This fully leverages local green energy advantages, achieving coordinated development of computing and electricity. Edge-level computing is deployed in prefecture-level cities and industrial parks, serving local governments, local SOEs, and small-to-medium internet enterprises. By sinking business processing to the source location, it reduces access latency.
Although the physical resource layout for the three-tier system is complete, network fragmentation persists between nodes at different levels. Cross-layer and cross-region computing scheduling still relies on traditional public networks. Training and inference for large models generate ultra-high bandwidth traffic, which ordinary networks struggle to support given the demands for large granularity, low latency, and high reliability.
Key Industry Regulatory Requirements: Data Residency as a Rigid Constraint
Laws and regulations across various industries draw clear red lines for compliant data usage. Ensuring data does not leave the factory, data remains within its domain, and raw data does not flow across domains has become a mandatory prerequisite for AI businesses in all sectors.
In the industrial sector, compliance with the Implementation Plan for Enhancing Data Security Capabilities in the Industrial Sector (2024–2026) requires strict adherence to keeping data within the factory premises. Production data must remain within the enterprise's physical boundaries, with core computational analysis performed locally. Only desensitized results are transmitted externally to prevent leakage of critical data regarding production processes and equipment operation.
The financial industry strictly follows the Data Security Management Measures for Banking and Insurance Institutions, implementing requirements that data must not leave its designated domain. Logical security domains are defined based on security levels. Sensitive data such as risk control records and transaction flows are deployed in dedicated protected facilities, with strict limitations on cross-domain data flow.
The healthcare industry adheres to health data security guidelines, strictly enforcing that raw data must not leave the domain and that data must be usable but invisible. Highly sensitive data such as medical records and genetic information only output computed results and model parameters externally. Raw data remains undisclosed throughout the process, and all data products must undergo privacy ethics reviews.
Regulatory requirements and business needs converge on one goal: computing power can be accessed remotely, but raw sensitive data cannot flow freely. Traditional public cloud models require uploading raw data to remote facilities, conflicting with regulatory mandates. Conversely, purely local self-built computing cannot share the capabilities of external large-scale intelligent computing pools. This contradiction is the most prominent challenge facing current large model deployments and the core problem the Computing Power Intelligent Network aims to solve.
Comparison of Four Enterprise Computing Usage Models
Currently, four mainstream solutions exist for enterprises acquiring AI computing power: purchasing integrated machines, building self-operated DCs, renting cloud computing, and distributed collaborative inference, as shown in Figure 1. Each model has shortcomings and struggles to simultaneously balance cost, elasticity, and security.
Figure 1: Comparison of Enterprise Computing Usage Models
First, purchasing integrated machines. Local deployment of intelligent computing integrated machines ensures data residency and suits SMEs experimenting with AI. However, these machines have hardware limits; performance significantly degrades when running large models exceeding 70B parameters. Connecting to external super-computing pools also requires expensive dedicated line bandwidth investments.
Second, building self-operated DCs. The advantage is that data remains entirely within the enterprise, ensuring security and controllability. However, disadvantages are stark: enterprises bear full costs for facility construction, hardware procurement, and O&M personnel. AI hardware iterates rapidly, leading to quick obsolescence of compute cards. Tidal business demands cause significant asset idleness, placing extreme demands on IT teams. Only a few large enterprises meet the conditions for viable implementation.
Third, directly renting remote public cloud computing. Enterprises avoid hardware investment and benefit from strong elastic scaling. The fatal flaw is that raw data must be uploaded to the public cloud, violating industry data security regulations. Additionally, long-distance transmission incurs high bandwidth costs, and protocol restrictions cause significant loss in computing efficiency, impacting the performance of large model inference and fine-tuning.
Fourth, distributed collaborative inference. Multiple computing nodes collaborate to complete training and inference tasks, suitable for large enterprises performing model fine-tuning. However, the number of adaptable models is currently limited, and cross-node collaboration demands extremely high dedicated line bandwidth, resulting in persistently high overall construction costs.
New business scenarios call for innovative models: deploying integrated machines locally to ensure data residency while remotely accessing operator-managed large-scale intelligent computing pools. This approach keeps computing efficiency loss within 5% over a300 km transmission distance, balancing security, elasticity, and cost. Such a business model relies fundamentally on the Computing Power Intelligent Network.
Overall Strategy and Technical Support for Building the Computing Power Intelligent Network
The core essence of the Computing Power Intelligent Network in the Token era is not merely deploying more servers, but achieving integrated fusion of computing resources, network transmission, data security, and business orchestration. With the business goal of "Two Connections, Two Uses," we promote the convergence of the basic communication network, industry-specific networks, and the computing network. This involves connecting AIDC computing nodes with national backbone OTNs to achieve millisecond-level interconnection of computing centers, millisecond-level access to computing resources, and sub-10ms reachability for computing applications in metropolitan areas. The goal is to build new computing infrastructure characterized by "data stays in the domain, computing can be accessed remotely."
Total Business Goal: Two Connections, Two Uses, Three-Network Convergence
The "Two Connections, Two Uses" development target established for operator computing-network services serves as a general reference for building the Computing Power Intelligent Network. "Two Connections" refers to linking AIDCs with the national backbone and converging the three networks. On one hand, intra-provincial computing resources connect to the national backbone network via intra-provincial DCI (Data Center Interconnect) and national backbone OTN. Core facilities act as new egress points for government and enterprise business, interfacing with the national computing network to export local computing services to regional clusters while importing external resources for provincial users. On the other hand, we push for the convergence of the basic communication network, industry-specific networks, and the computing network, bridging the intelligent network and the computing network to break "network silos" and enable efficient cross-network computing scheduling.
"Two Uses" refers to secure computing usage and efficient data utilization. Secure computing usage entails meeting industry regulatory requirements, ensuring enterprise data residency, and isolating raw sensitive data securely. Efficient data utilization spans the entire data lifecycle, constructing a complete chain from "data annotation forming datasets stored -> dataset training yielding industry models -> models outputting inference Token services." This connects the full "Storage-Training-Inference" business chain, providing comprehensive Token service capabilities to enterprises. The target network form for computing-network business development is shown in Figure 2.
Figure 2: Target Network Form for Computing-Network Business Development
Regarding latency metrics, three tiers of millisecond-level targets are established: interconnection latency between computing centers < 1 ms; access latency for key locations like parks and factories to computing resources <1 ms; end-to-end latency for terminal devices accessing computing applications < 10 ms. This achieves a millisecond-level computing experience within metropolitan areas, meeting the needs of latency-sensitive scenarios such as industrial control, smart healthcare, and government AI.
Underlying Network Technology Support: IPv6+ All-Optical DCI Network
Existing network architectures are centered around proprietary services. Transitioning to a computing-centric network requires evolving the OTN architecture to center on computing power. Simultaneously, intelligent network scheduling capabilities must evolve to support rapid, flexible business scheduling, one-hop cloud access, and computing-network coordination.
The network foundation of the Computing Power Intelligent Network is built on the "IPv6+" (Internet Protocol Version 6 Plus) technology system, OTN all-optical DCI, and SPN (Slicing Packet Network). It encompasses technical capabilities such as SRv6 (Segment Routing IPv6) for minimalist connectivity, FlexE (Flexible Ethernet) for deterministic bandwidth, iFIT (In-situ Flow Information Telemetry) for real-time detection, and APN6 (Application-Aware IPv6 Networking) for intelligent business assurance. These technologies support the upgrade from the Internet of Everything to the Intelligent Internet of Everything, shifting from consumer internet to industrial internet, and providing end-to-end technical assurance for the Computing Power Intelligent Network.
All-optical DCI interconnection is the base of the Computing Power Intelligent Network. Internally, DCs adopt a FullMesh architecture. Optical cross-connect devices are sunk to edge nodes to enable direct wavelength routing, with a flattened network architecture guaranteeing low latency. We create large-granularity DCI dedicated lines and OTN government-enterprise dedicated line products, enabling activation within 7 working days. ASON (Automatically Switched Optical Network) ensures "five nines" (99.999%) business reliability, supporting end-to-end automated creation of computing-network services. SPN provides logically isolated transmission channels for industry users, ensuring secure business transmission. The all-optical interconnection network for computing-business is shown in Figure 3.
The overall network architecture is layered: The business access layer provides unified access for government/enterprise, home, and wireless AI services. The OTN aggregation and core layers handle intelligent scheduling of computing traffic. DCI enables all-optical interconnection among AIDCs within the province. National backbone OTN supports cross-provincial computing interconnection. Furthermore, it interfaces with group intelligent computing PE dedicated line egresses, opening access paths to national computing pools.
Hardware Product Carrier: Tiered Intelligent Computing Integrated Machines for Diverse Scenarios
The computing-network integrated machine is the hardware carrier for the Computing Power Intelligent Network on the enterprise side, enabling efficient coordination between local deployment and cloud computing. It features five characteristics: "Small, Complete, Silent, Elastic, Secure." "Small" means the hardware size approaches a standard PC tower, allowing flexible deployment. "Complete" signifies integrated computing, networking, and storage. "Silent" indicates low noise (< 55 dB), suitable for office environments. "Elastic" denotes the ability to elastically invoke cloud computing. "Secure" represents full-process security management ensuring data residency.
Integrated machines come in high, medium, and low tiers, supporting local deployment of 72B, 32B, and 14B large models respectively. Through edge-cloud collaboration, they can also access computing power for 744B ultra-large models. Hardware uses built-in water cooling, with a maximum power consumption of 2,000 W. It supports RDMA lossless wide-area networks and deeply adapts to the operator's Computing Power Intelligent Network infrastructure. Deploying integrated machines on the enterprise side retains raw business data without external transmission. When massive computing power is needed, the system remotely accesses operator intelligent computing pools via the Computing Power Intelligent Network, transmitting only model parameters and Token inference requests. Raw data remains local, precisely matching data supervision requirements in industries like manufacturing, finance, and healthcare, effectively resolving the conflict between data security and computing scale.
Building Computing Power Intelligent Network Solutions for Industry Users
For consumers, operators build wireless bearer networks, charging for traffic while users do not pay for base station equipment. For the home broadband market, operators build wired metropolitan bearer networks, charging for bandwidth while users do not purchase OLT (Optical Line Terminal) access devices. Following this mature commercial logic, in the Token era, operators can build a Computing Power Intelligent Network (broadly defined as a DC) for industry users. This constitutes the core business model of the Computing Power Intelligent Network. The construction mode is shown in Figure 4.
Figure 4: Construction Mode of the Computing Power Intelligent Network
Core Layer Architecture of the Computing Power Intelligent Network
Intelligent Computing Connectivity: 1 pair of Digital Link PEs + Parameter Plane Switch;
General Computing Connectivity: 1 pair of General Computing PEs (reusable);
Digital Link Access: 1 pair of Digital Link PEs;
Functional Platform: Digital Link / Computing Scheduling / Computing Operations.
Aggregation and Access Layer Architecture of the Computing Power Intelligent Network
All-Optical DCI Interconnection: The base of the Computing Power Intelligent Network. Internal DCs use a FullMesh architecture. Optical cross-connect devices are sunk to edge nodes for direct wavelength routing. Large-granularity DCI dedicated lines and OTN government-enterprise dedicated lines are created, supporting end-to-end automated creation of computing-network services.
SPN Private Network Extension: Reuses existing coverage resources to bridge user-side and computing-resource-side connections. Hard slicing achieves security isolation, while network resource reuse enables bandwidth overselling.
Data Highland Aggregation: Concentrated deployment in third-party DCs and industrial parks, facilitating grid connection of edge storage and computing power.
Industry Entry Point Access: Dedicated network entry points for government/enterprise, healthcare, education, etc.
User-Side Architecture of the Computing Power Intelligent Network
Integrated Machine Access: Built-in computing boards and transmission CPEs allow these machines to serve as edge access nodes for the Computing Power Intelligent Network.
Dedicated Line Access: Compatible with mainstream forms such as conventional high-bandwidth dedicated lines and digital-intelligent interconnection dedicated lines.
Innovative Billing Logic for the Computing Power Intelligent Network: For industry users, charges are primarily levied on data and Token traffic usage, eliminating high initial installation fees for dedicated lines. Network construction costs are borne by operators, alleviating early-stage industry investment pressure. Operator revenue is realized through two methods: first, back-end revenue sharing with computing pool operators; second, front-end value-added service fees covering security hardening, business elasticity assurance, high-reliability SLAs (Service Level Agreements), managed O&M, and intranet application services, creating new revenue growth points.
Typical Application Scenarios for the Computing Power Intelligent Network
Industrial Manufacturing Scenario
Production processes and equipment operation data in manufacturing enterprises are core commercial secrets, with regulations mandating data residency within the factory. Enterprises deploy low-tier intelligent computing integrated machines on-site to perform local inference tasks like quality inspection and production line monitoring, keeping raw production data within the plant. When fine-tuning large models using massive historical production data is required, the system remotely accesses operator hub-level intelligent computing pools via the Computing Power Intelligent Network's SPN. The transmission process uploads no raw production data, sending only desensitized features and model gradient parameters. Upon completion, updated models are returned to the local factory. This meets the massive computing demands of large model training while adhering to data residency regulations, with end-to-end latency satisfying millisecond-level industrial requirements.
Smart Government and Judicial Scenario
In government and judicial scenarios, vast amounts of case files and materials constitute sensitive administrative data. Prefecture-level government units deploy intelligent computing integrated machines to perform local file parsing and case-assisted reasoning. For large-scale batch fine-tuning of case files, they access provincial operator regional intelligent computing centers. Secure isolated transmission is achieved via Computing Power Intelligent Network OTN dedicated lines. Original case materials never leave the government intranet domain; only Tokens and model parameters are exchanged, supporting AI-assisted case handling and intelligent file archiving. Provincial government platforms can also schedule edge computing resources across prefectures via the Computing Power Intelligent Network, achieving coordinated dispatch of provincial government AI business.
Healthcare Industry Scenario
Hospital medical records and genomic sequencing data have extremely high privacy protection requirements, necessitating strict data residency. Medical institutions deploy intelligent computing integrated machines to perform local inference tasks like image recognition and record organization. When scientific research model training requires large-scale computing support, the system remotely accesses intelligent computing pools via the Computing Power Intelligent Network, transmitting only feature parameters. Raw medical records and genetic data remain entirely within the hospital, achieving "data usable but invisible," balancing medical research innovation with patient privacy security.
General AI Business for SMEs
Many SMEs have large model usage needs but cannot afford the high costs of fully self-built intelligent computing facilities. Deploying small-to-medium intelligent computing integrated machines allows local access to enterprise knowledge bases, ensuring documents and business data remain within the domain. Daily Q&A and customer service inference run locally. When complex business fine-tuning is needed, the system elastically invokes intelligent computing resources from the operator's Computing Power Intelligent Network, billing by Token volume. This achieves lightweight investment and on-demand expansion, avoiding asset sunk risks caused by rapid computing hardware iteration.
Challenges and Outlook
Currently, the construction of the Computing Power Intelligent Network faces several practical challenges. First, standards for cross-layer and cross-operator computing-network scheduling are not yet fully unified; business orchestration, metering/billing, and SLA assurance systems across different computing pools need further standardization. Second, the Computing Power Intelligent Network represents an innovative business model; transitioning from traditional dedicated line revenue to "Token Traffic + Value-Added Services" requires adjusting business models and accounting systems. Third, the industrial ecosystem for edge-cloud collaborative integrated machines and RDMA lossless wide-area networks needs maturation to reduce overall hardware costs. Fourth, data security norms vary widely across industries, requiring the Computing Power Intelligent Network's security management capabilities to continuously align with specific industry regulatory details.
Opportunities are equally evident: National integrated computing network policies continue to land and advance, with the scale of three-tier computing infrastructure expanding steadily. Large models are penetrating deeply into all industries, with Token-based metering becoming the mainstream form of computing consumption. Operators possess comprehensive advantages in networks, facilities, and government/enterprise user bases, naturally fitting the role of primary builders for the Computing Power Intelligent Network.
In the future, the Computing Power Intelligent Network will connect the entire chain of "Computing Deployment—Computing Interconnection—Computing Usage," freeing enterprises from the dilemma of choosing between "computing elasticity" and "data security." Leveraging this innovative business model, operators will upgrade from traditional bandwidth and dedicated line service providers to integrated computing-network service providers in the Token era, helping industries across all sectors achieve secure and controllable scaled deployment of large models.
This article was published in Communications World, Issue 10, 2026; Original Title: "Planning and Practice of the Computing Power Intelligent Network Based on DCI All-Optical Interconnection."
Author: Wang Lin, China Mobile Design Institute