Computing Power Networks (CPNs)

• Computing Power Networks (CPNs) are a paradigm that treats distributed cloud, edge, and terminal resources as networked computing power, enabling integrated service orchestration.
• Their multi-layer architecture abstracts and pools heterogeneous resources while using centralized or hierarchical controllers to optimize computing and connectivity jointly.
• Key techniques include optimization for service mapping, federated learning adaptations, and energy-aware scheduling to enhance resource utilization and quality of service.

Computing Power Networks (CPNs) are a network paradigm in which computing power is treated as a first-class networked resource: distributed cloud, edge, and terminal resources are interconnected, abstracted, perceived, allocated, and scheduled jointly with storage and networking, so that services request “computing power + connectivity” rather than bandwidth alone (Sun et al., 2022). In this view, a CPN interconnects heterogeneous computing, storage, and networking resources distributed across cloud, edge, and end devices; abstracts computing power as a networked service; and orchestrates and schedules computing resources on-demand according to service requirements, aiming at global optimization of resource utilization and service QoS (Sun et al., 28 Sep 2025). Recent work further extends the concept from a mostly ground, static CPN to integrated cloud–edge–air and space-centered forms in which mobility, sensing, and cyber–physical reconfiguration become part of the computing fabric itself (Sun et al., 24 Nov 2025).

1. Concept and distinguishing characteristics

A CPN emerged from the observation that cloud computing, edge computing, and smart devices have produced ubiquitous deployment of computing resources, yet traditional network architectures cannot efficiently leverage these distributed resources because of the computing power island effect (Sun et al., 2022). A survey characterizes CPN as a proposed post-MEC, post-cloud networking paradigm in which computing power itself becomes a networked resource that is sensed, routed, and scheduled across heterogeneous end–edge–cloud infrastructures (Sun et al., 2022).

The defining distinction is that a conventional communication network focuses on data transport, while cloud and edge computing focus on where computation happens but largely treat the network as a transparent pipe. A CPN “unifies both”: the network is aware of computing resources and actively discovers, selects, and orchestrates compute nodes for each service (Sun et al., 24 Nov 2025). In the same direction, recent surveys describe CPN as integrating cloud + edge + end resources, using the network to model and virtualize computing power, route tasks based on both communication and computation conditions, and perform global, service-aware orchestration (Sun et al., 28 Sep 2025).

This operational model is commonly framed as Computing-as-a-Service (CaaS), with computing power treated analogously to bandwidth. Three design principles recur across the literature: service-oriented requests for “computing power + connectivity,” global coordination through centralized or logically centralized controllers, and programmability through SDN/NFV-like control over function placement, routing, and adaptation (Sun et al., 24 Nov 2025).

2. Architectural model and control structure

The survey literature describes a multi-layer CPN stack. One representative architecture consists of a CPN Infrastructure Layer, CPN Resource Pooling Layer, CPN Resource Information Announcement Layer, CPN Computing Scheduling Optimization Layer, CPN Service Layer, and CPN Orchestration and Management Layer (Sun et al., 2022). The infrastructure layer contains end devices, edge nodes, cloud resources, and heterogeneous network access; the resource pooling layer abstracts compute, network, and service resources; the information announcement layer disseminates state; the scheduling layer performs task offloading and resource allocation; the service layer hosts AI and application microservices; and the orchestration and management layer handles computing power modeling, OAM, orchestration, and security (Sun et al., 2022).

A baseline architectural decomposition used in more recent work includes five components: Computing Nodes, Network Infrastructure, Control / Management Plane, Orchestration & Virtualization, and Interfaces to Services (Sun et al., 24 Nov 2025). In this formulation, the substrate includes cloud data centers, regional clouds, MEC or edge servers, and possibly capable terminals; the forwarding plane may be SDN-controlled; the control plane is a centralized or hierarchical CPN orchestrator with a global view of computing capacities, loads, and link quality; orchestration includes VNFs and Service Function Chains (SFCs); and service interfaces expose QoS constraints such as latency, bandwidth, reliability, CPU/GPU cycles, memory, and model type (Sun et al., 24 Nov 2025).

The same architectural logic is often expressed as a cloud–edge–end fabric, and recent air–ground work extends it to a cloud–edge–air framework in which UAVs and low-altitude platforms become mobile computing and networking nodes under the same orchestrator (Sun et al., 28 Sep 2025). This suggests that the architectural core of CPN is not a fixed placement hierarchy but a programmable control system over heterogeneous, geographically distributed resources.

3. Core technical functions and formal abstractions

The literature repeatedly identifies a stable functional core. A recent survey lists eight capabilities: computation offloading, computing power matching, computing power virtualization, computing–network integration (CNI), computing-aware routing (CAR), computing-aware service orchestration, programmable networking, and ubiquitous computing power (Sun et al., 28 Sep 2025). The 2022 survey complements this with three enabling control processes: computing power modeling, information awareness and announcement, and resource allocation with network forwarding (Sun et al., 2022).

Information awareness is multi-dimensional. CPN controllers may collect user mobility and task requirements, node-side compute/storage/network/service states, and link-level metrics such as bandwidth, delay, jitter, and loss (Sun et al., 2022). Information dissemination may be centralized, distributed through extended IGP/BGP mechanisms, or hybrid, depending on the desired trade-off among convergence speed, view granularity, and deployment cost (Sun et al., 2022).

A standard analytical abstraction is joint communication–computation optimization. One representative formulation minimizes a weighted sum of delay and energy:

$$\begin{aligned} \min_{\boldsymbol{x}, \boldsymbol{f}, \boldsymbol{p}} \quad & \sum_{i \in \mathcal{U}} \Big( \alpha_i T_i(\boldsymbol{x}, \boldsymbol{f}, \boldsymbol{p}) + \beta_i E_i(\boldsymbol{x}, \boldsymbol{f}, \boldsymbol{p}) \Big) \ \text{s.t.} \quad & \sum_{i} x_{i,k} C_i \leq F_k^{\max}, \quad \forall k \in \mathcal{N}, \ & \sum_{i} x_{i,k} R_i \leq B_k^{\max}, \quad \forall k \in \mathcal{N}, \ & x_{i,k} \in \{0,1\},\ \sum_{k} x_{i,k} = 1,\quad \forall i . \end{aligned}$$

Here, tasks or users are assigned to computing nodes while respecting CPU and bandwidth capacities (Sun et al., 28 Sep 2025).

For cloud-native service placement, a CPN substrate is often modeled as an undirected graph $G^s = (N^s, L^s)$ with computing nodes and network links, while a service request is modeled as $G^v_i = (N^v_i, L^v_i)$ with service functions and logical links (Xie et al., 10 Jul 2025). Binary variables $x^{u^v}_{m^s} \in \{0,1\}$ encode service-function placement, and $f^{u^v w^v}_{p,k} \in \{0,1\}$ encodes logical-link embedding over precomputed tunnels (Xie et al., 10 Jul 2025). A distinctive feature of the CPN formulation is that multiple service functions of the same service entity may co-locate on one computing node, which changes the combinatorics relative to conventional virtual network embedding (Xie et al., 10 Jul 2025).

4. Optimization, learning, and orchestration frameworks

A major line of work formalizes service mapping as the core optimization problem of a cloud-native CPN. “A Fragmentation-Aware Adaptive Bilevel Search Framework for Service Mapping in Computing Power Networks” defines the problem, proves that even the single-request version is NP-hard through reductions connected to graph bisection and k-way partitioning, and introduces Adaptive Bilevel Search (ABS), combining graph partitioning-based reformulation, a bilevel optimization architecture, and fragmentation-aware evaluation via NRED, CBUG, and PNVL (Xie et al., 10 Jul 2025). In complex scenarios, ABS achieves up to 73.2% higher computing resource utilization and a 60.2% higher service acceptance ratio compared to the best-performing baseline (Xie et al., 10 Jul 2025).

Federated learning has become a second major CPN workload model. “FedAPTA” treats a CPN as a distributed infrastructure with a central server and heterogeneous devices running different tasks, and addresses multi-task deployment through adaptive layer-wise pruning, heterogeneous model recovery, and task-aware aggregation with HDBSCAN clustering (Yuan et al., 4 Aug 2025). In the reported CPN setting of 50 heterogeneous devices and 5 distinct classification tasks, the framework improves average accuracy over FedLPS from 88.80% to 89.54% on ResNet18 under i.i.d. partitioning and from 83.37% to 85.87% under non-i.i.d. partitioning; on ShuffleNetV2, it improves from 77.19% to 79.96% under i.i.d. and from 70.98% to 75.30% under non-i.i.d. partitioning, with up to 4.23% gain over some baselines on specific tasks (Yuan et al., 4 Aug 2025). The same study reports parameter reduction from 11.01M to 2.20M for ResNet18 and from 1.21M to 0.24M for ShuffleNetV2 at pruning ratio $\rho=0.8$ (Yuan et al., 4 Aug 2025).

Energy-aware orchestration has likewise been cast as a CPN-native control problem. “A Novel Hierarchical Co-Optimization Framework for Coordinated Task Scheduling and Power Dispatch in Computing Power Networks” models a grid-coupled CPN in which a day-ahead stochastic unit commitment stage is coupled to real-time economic dispatch and DRL-based task scheduling (Luo et al., 6 Aug 2025). In simulations on an IEEE 30-bus system with 5 CPN nodes, TSCO reduces curtailment by over 60% compared to CO-Opt and DC-Frame, while maintaining a job success rate of approximately 98.5% and average job tardiness of approximately 12.3 s; relative to CO-Opt, emissions are 41.5% lower and total cost is 17.5% lower once carbon cost is included, and relative to DC-Frame, emissions are 16.2% lower and total cost is 12.7% lower (Luo et al., 6 Aug 2025).

Security and transaction mechanisms have also been instantiated explicitly. “BECS” defines a 6G CPN as a three-layer user–edge–cloud computing fabric coordinated by 6G base stations, formulates computing sharing as a six-objective MOOP, and combines NSGA-III-KDR with a permissioned blockchain and a zero-knowledge pseudonym scheme (Yan et al., 2024). In the reported setting with 300 user devices, 200 edge devices, and 100 cloud devices, the proposed pseudonym scheme yields a total computational overhead of 8.664 ms, lower than the compared schemes, while NSGA-III-KDR improves resource utilization and privacy entropy relative to standard NSGA-III, MOEA/D, and NSGA-II-SDR (Yan et al., 2024).

5. Deterministic, mobile, aerial, and space extensions

Deterministic CPN research adds hard timing semantics to the basic paradigm. “Deterministic Computing Power Networking” defines Det-CPN as a CPN integrated with deterministic network technology so as to provide determinacy for latency, jitter, path, and computing (Jia et al., 2024). Its architecture introduces an Infrastructure Layer, a Computing and Network Convergence Layer split into a Computing and Network Decision Plane (CNDP) and a Computing and Network Execution Plane (CNEP), and an Application Service Layer (Jia et al., 2024). On the networking side it relies on TSN, “5G+TSN,” DetNet, SR, and lossless data center networking with RDMA; on the computing side it uses task prioritization, preemption, resource reservation, and resource pre-adjustment (Jia et al., 2024). The motivating scenarios are explicitly time-bounded, including Cloud VR with motion-to-photon latency usually below 20 ms and industrial PLC communication with typical maximum end-to-end message delivery time in the range $100\,\mu s$ to $50\,ms$ (Jia et al., 2024).

A second major extension integrates CPN with the Low-Altitude Economy (LAE). Recent work characterizes conventional CPNs as static, ground-based systems with fixed locations and limited environmental awareness, then proposes a cloud–edge–air collaborative CPN in which both ground CPN nodes and LAE nodes are cooperative entities under a unified orchestrator (Sun et al., 24 Nov 2025). The control model adopts an “agentification” paradigm: edge and flying agents support perception, planning, action, and reflection, while Duplex MCP allows agents to enter an Elicitation state and trigger re-planning when link failure, overload, or new hotspots violate plan preconditions (Sun et al., 24 Nov 2025). In the reported case study over a square area with 20 CPN edge nodes and 20 LAE nodes, the integrated CPN–LAE environment achieves higher overall task completion than either pure CPN or pure LAE, and under dynamic hotspot conditions obtains a substantially higher task success rate than either baseline (Sun et al., 24 Nov 2025).

UAV-enabled CPNs make this mobility explicit at the physical-layer and performance-analysis level. One analytical model defines a request zone of radius $R_u$, a service zone populated by computing nodes according to a homogeneous PPP $\Phi_C$ with density $G^s = (N^s, L^s)$0, and a UAV relay at altitude $G^s = (N^s, L^s)$1, then introduces task completion probability as the probability that at least one computing node can satisfy the end-to-end latency budget (Deng et al., 17 Dec 2025). The per-user success probability is written as

$G^s = (N^s, L^s)$2

which directly couples communication and computation (Deng et al., 17 Dec 2025). In the reported numerical results, fixing UAV altitude at 300 m, expanding CN distribution radius from 200 m to 1000 m raises task completion probability from approximately 46.65% to approximately 99.14%, a 2.13× improvement (Deng et al., 17 Dec 2025). A subsequent energy-constrained extension jointly optimizes UAV altitude and transmit power under a hybrid fuel-cell and battery architecture, and reports that average performance drop compared to unconstrained operation can be about 50% (Deng et al., 14 Jan 2026).

Space-CPN extends the same logic to LEO constellations. “Brain-Inspired Decentralized Satellite Learning in Space Computing Power Networks” treats satellites as distributed computing nodes connected by intra-plane and inter-plane ISLs, uses ring all-reduce within each orbit plane and RelaySum across a minimum-diameter inter-plane aggregation tree, and couples this to spiking neural networks on neuromorphic hardware (Yang et al., 27 Jan 2025). The convergence analysis shows explicit dependence on network diameter, and experiments compare a chain tree of diameter 7 with an optimized tree of diameter 3, with the optimized tree converging faster while reaching similar final accuracy (Yang et al., 27 Jan 2025). The same study reports that SNNs consume roughly 10× less energy per layer than corresponding ANNs at similar accuracy, and that after 60 inter-plane communication rounds the RelaySum-based scheme remains more than 10% absolute test accuracy ahead of gossip and all-reduce baselines in the reported setting (Yang et al., 27 Jan 2025).

6. Open challenges and research directions

The research agenda remains broad. The 2022 survey identifies computing power modeling, information synchronization and scalability, mobility enhancement, communication bottlenecks, energy consumption, deterministic ultra-low latency, security and trust, and pricing and incentives as open issues (Sun et al., 2022). Even the basic question of a unified computing-power metric remains unsettled, because practical systems expose heterogeneous CPU, GPU, NPU, FPGA, memory, storage, network, and service-level characteristics (Sun et al., 2022).

Recent surveys sharpen these challenges in integrated cloud–edge–air systems. They emphasize digital-twin-enhanced LAE–CPN integration, security and privacy, energy-aware integration, heterogeneous LAE nodes and multi-agent coordination, and extension toward space–air–ground integrated systems (Sun et al., 28 Sep 2025). The integrated air–ground study adds mobility management, resource heterogeneity, control complexity at scale, and standardization and interfaces as immediate obstacles to deployment (Sun et al., 24 Nov 2025).

Deterministic variants expose another set of unresolved problems: integrated modeling and control across operator and cloud-provider boundaries, integration of forwarding, computing and caching, and transmission control redesign for deterministic environments rather than best-effort transport (Jia et al., 2024). This suggests that future CPN work will depend not only on better optimization algorithms but also on common abstractions, cross-domain APIs, and interoperable control planes.

Across the literature, the long-term trajectory is clear: CPN evolves from a static, ground-centric resource orchestration system toward a mobile, context-aware, AI-native, cyber–physical computing fabric in which routing, service placement, sensing, learning, energy, and even physical node mobility are jointly optimized (Sun et al., 24 Nov 2025).