From Empowerment to Vulnerability: The Computation–Energy Paradox of AI-Enabled Power-Transport Systems

Abstract

The transition towards smart megacities has deeply integrated Artificial Intelligence (AI) with power–transport networks. While AI empowers complex operations like multi-network coordinated dispatch and emergency rescue, current algorithm-centric perspectives largely ignore its massive physical energy costs. Accordingly, this Perspective examines the dual role of AI, considering it not only as an intelligent decision-support tool but also as a potential source of additional stress on physical infrastructure. First, through a structured synthesis of the representative literature, we deconstruct the functional dependencies between algorithms and physical infrastructures, identifying how AI reshapes the operational paradigms of power, ground transport, and aerial networks under routine and emergency scenarios. We then introduce the concept of the “Computation–Energy Paradox.” Integrating conceptual analysis with a quantitative case study of a typical community, we illustrate a plausible failure mechanism: during extreme disasters, intensified AI invocation for emergency management generates surging computational loads, which paradoxically exacerbate power shortages and reduce the operating margin of already weakened systems. In addition, we analyze core engineering bottlenecks, including spatiotemporal computation–energy mismatches and physical constraints in extreme edge environments. To address these challenges, we outline a prospective roadmap encompassing lightweight emergency AI and computation–power-coordinated offloading mechanisms. Finally, the sustainable development of such systems suggests a paradigm shift: AI must evolve from a purely virtual algorithm into a physical component of an integrated compute–power–transport system.

1. Introduction

The global transition towards smart megacities has driven the deep integration of Artificial Intelligence (AI), particularly large foundation models, with physical urban infrastructures [1,2]. This integration is transforming traditional, isolated power and transport networks into a highly coupled “compute–power–transport” symbiotic nexus [3,4]. In normal operational states, AI serves as a powerful enabler. For instance, the widespread deployment of Electric Vehicles (EVs) and Unmanned Aerial Vehicles (UAVs) relies heavily on AI for complex tasks such as vehicle-to-grid (V2G) scheduling, autonomous grid inspections, and traffic optimization [5,6,7,8]. However, this pervasive intelligence comes at a staggering physical cost. Recent data indicates that a single generative AI query consumes nearly ten times the electricity of a standard search engine request [9], and global data center power demand is projected to exceed 1000 TWh by 2026 [10]. While current literature and review studies extensively document how AI algorithms optimize system efficiency and resilience, this narrative predominantly focuses on a one-way “empowerment” perspective [11,12,13]. It treats AI as an omnipotent, resource-free cognitive layer, largely ignoring the physical “bottom line” of energy consumption: that massive computing power is itself an extremely energy-intensive and rigid load.

This algorithm-centric view masks a critical vulnerability during extreme disaster scenarios [14,15], such as severe hurricanes or widespread blackouts (e.g., the 2021 Texas power crisis or recent extreme weather events) [16,17]. Under such conditions, the physical power–transport system experiences severe structural damage and energy imbalances [18,19]. Consequently, the system heavily invokes AI for emergency response, such as UAV swarm-based 3D disaster reconstruction and EV fleet evacuation routing [20,21]. Here, a critical blind spot in existing research emerges: the failure to recognize the physical dependence of AI on the underlying grid. This creates a complex, two-fold crisis. First, the physical disruption of the power grid can directly incapacitate data centers and communication nodes, rendering the envisioned AI rescue strategies infeasible [22,23]. Second, even if edge computing is forcefully deployed, the massive concurrent data throughput and high-frequency AI inference requests instantly transform into a potentially concentrated, high-power demand [10,24,25]. Instead of saving the system, this sudden surge in computational load exacerbates the existing power deficit, creating adverse feedback on the fragile infrastructure.

The research gap addressed in this Perspective is therefore not the general observation that AI and data centers consume electricity. It is the limited treatment of emergency computing as a coupled physical load within damaged power–transport infrastructures. Existing studies commonly examine computing efficiency, resilient microgrid operation, or AI-enabled emergency coordination separately. The proposed Computation–Energy Paradox connects these strands by focusing on the operational feedback that may arise when safety-relevant computing demand increases while local generation, cooling, communication, and network flexibility are simultaneously constrained.

To fill this gap, this paper develops a structured Perspective framework, tracing the evolutionary trajectory of AI in megacity coupled systems from “empowerment” to “vulnerability.” We no longer treat AI merely as an optimizer, but critically examine its dual role as both a “cognitive brain” and a “potential system burden.” Based on this, this Perspective aims to broaden the current algorithm-centric perspective, urging future research to explicitly account for the physical energy costs of AI within urban infrastructures. The main contributions of this paper are as follows:

• A structured synthesis of AI applications within the power–transport coupled system is conducted. Moving beyond isolated use cases, this synthesis organizes AI deployments across distinct physical domains—namely, the power grid, ground transport, and aerial networks. By mapping these applications across normal operations and emergency response scenarios, this paper deconstructs the functional dependencies between algorithms and physical infrastructures, establishing a cross-domain baseline of how AI currently empowers the urban nexus.
• The “Computation–Energy Paradox” is conceptually formulated for extreme disaster response across coupled power, ground-transport, and aerial networks. Through a community-scale case study, we identify a plausible feedback mechanism: intensified emergency AI invocation creates a concentrated electrical demand that can further reduce the operating margin of an already damaged islanded system. The focused 12-case sensitivity comparison shows that this additional grid burden persists across the tested load magnitudes and connection buses, while its severity varies systematically with demand size and electrical location under fixed resource-control and protection settings. This conditional finding is central to the contribution: the paradox identifies a coupled physical mechanism and a planning variable, rather than asserting that AI demand will cause universal grid collapse.
• The core engineering bottlenecks underlying this paradox are examined, including spatiotemporal computation–energy mismatches and physical constraints in extreme edge environments. To address these challenges, we outline a prospective roadmap including lightweight emergency AI and compute–power coordinated offloading mechanisms. This provides a structured conceptual basis and a technological outlook for resolving the fundamental conflict between high-intensity intelligent rescue efforts and fragile physical infrastructures.

2. AI Applications in Power–Transport Systems

This section synthesizes representative studies on AI applications across power systems, ground transport, aerial networks, and cross-network coordination. The discussion covers both routine operation and emergency response, with particular attention to the roles of edge computing, computational demand, and the physical dependence of AI services on energy infrastructure.

2.1. Normal Operations and Optimization

With the large-scale integration of renewable energy resources, distributed energy resources (DERs), and electrified loads, modern power systems are gradually evolving from conventional centralized architectures toward highly dynamic and bidirectionally coupled smart grids. Compared with traditional power systems, modern smart grids are characterized by stronger load fluctuations, increased uncertainty in renewable generation, and increasingly complex multi-timescale interactions among generation units, energy storage systems, demand-side resources, and distributed terminals [26]. Under such conditions, conventional rule-based and static optimization approaches are increasingly unable to satisfy the requirements of real-time prediction, adaptive coordination, and resilient operation in highly distributed energy systems. Consequently, AI technologies have been extensively introduced into modern smart grid infrastructures to enhance operational awareness, forecasting capability, and autonomous decision-making [27]. Existing studies have demonstrated that AI can effectively support a wide range of smart grid applications, including short-term load forecasting, renewable energy prediction, demand-side management, distributed energy coordination, and intelligent stability control [28,29]. In particular, recent advances in machine learning (ML) and deep learning (DL) have significantly improved the capability of capturing nonlinear temporal dependencies and spatial correlations in large-scale energy datasets, thereby enabling the transition of power systems from passive energy delivery networks toward intelligent, adaptive, and self-coordinated cyber–physical energy systems [30]. Furthermore, the emergence of advanced paradigms such as explainable artificial intelligence (XAI), federated learning (FL), graph neural networks (GNNs), and multi-agent reinforcement learning (MARL) indicates that AI is gradually evolving from a standalone prediction tool into a system-level orchestration engine for future energy infrastructures [31]. As illustrated in Figure 1, AI applications in normal operations are organized into four complementary categories from a system perspective. The first three categories correspond to representative physical infrastructure domains, namely power systems, ground transport networks, and aerial networks, where AI is directly deployed for perception, prediction, optimization, and autonomous operation. These domains are selected because they collectively determine urban energy flow, mobility, and situational awareness, forming the fundamental cyber–physical foundation of AI-enabled smart megacities. Building upon these individual domains, the final category focuses on cross-network coordination, where AI enables information exchange, resource sharing, and collaborative decision-making across heterogeneous infrastructures.

2.1.1. Power System Operations

Accurate spatiotemporal forecasting has become one of the fundamental requirements for smart grid operation due to the increasing penetration of renewable energy resources, distributed energy systems, and electrified loads. Unlike traditional power systems with relatively stable and predictable load profiles, modern smart grids exhibit strong temporal variability and complex spatial correlations caused by renewable intermittency, dynamic electricity consumption behaviors, and geographically DERs [32]. Conventional statistical forecasting approaches, such as autoregressive integrated moving average (ARIMA) and Kalman filtering methods, often struggle to capture the nonlinear and multi-scale dynamics of modern energy systems [33]. To address these limitations, ML and DL techniques have been increasingly adopted for short-term load forecasting (STLF), renewable generation prediction, and demand-side energy estimation. Early studies mainly relied on support vector machines (SVMs), random forests (RFs), and shallow neural networks [34,35], while recent research has shifted toward deep learning architectures such as recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and convolutional neural networks (CNNs) for capturing long-range temporal dependencies in electricity consumption patterns [36,37]. More recently, GNNs, Transformer-based architectures, and spatiotemporal graph neural networks (STGNNs) have demonstrated strong capability in modeling both temporal dependencies and topological spatial correlations in highly distributed power systems [38,39,40,41]. These approaches enable power systems to transition from passive energy delivery networks toward predictive and adaptive cyber–physical energy infrastructures with enhanced operational awareness and real-time decision support capabilities.

Beyond predictive intelligence, AI has also been increasingly utilized for source–grid–load–storage coordination in modern smart grids. As renewable energy resources are distributed, battery energy storage systems (BESSs), flexible loads, and prosumer-side energy interactions continue to expand, conventional centralized optimization approaches face growing challenges in handling the high-dimensional, nonlinear, and dynamic characteristics of modern energy systems [42,43]. In this context, AI-driven coordination strategies, particularly deep reinforcement learning (DRL) and MARL, have emerged as promising solutions for adaptive energy management and real-time distributed dispatch [44]. Existing studies demonstrate that DRL-based energy management frameworks can effectively optimize power flow, storage scheduling, demand response, and renewable energy utilization under uncertain operating conditions [45]. Compared with traditional optimization methods that rely heavily on accurate mathematical models and predefined operating constraints, DRL enables energy systems to learn adaptive operational policies directly from historical interactions and real-time environmental feedback [46]. More recently, MARL-based coordination strategies have attracted increasing attention for large-scale distributed energy systems, where multiple agents collaboratively manage interconnected generation units, storage devices, and flexible loads in decentralized smart grid environments [46,47]. Such developments indicate that AI is gradually transforming modern power systems from conventional rule-driven infrastructures toward self-adaptive and autonomously coordinated energy networks with enhanced operational flexibility and distributed intelligence.

2.1.2. Ground Transport Operations

The rapid growth of EV penetration has significantly transformed the interaction between ground transport networks and urban power grids. Rather than being treated merely as isolated mobile loads, EVs under AI orchestration are structurally integrated into the broader ground transport network infrastructure, presenting unique spatial-temporal dynamics [48,49]. Large-scale, uncoordinated vehicle charging across the urban roadway layout introduces severe operational challenges to localized distribution networks, including peak load amplification, voltage fluctuations, transformer overloading, and increased operational uncertainty [50]. Consequently, AI-enabled coordinated charging and routing strategies have attracted extensive research attention for improving ground transport network flexibility and maintaining grid stability under high penetration scenarios. Existing studies indicate that the orchestration of the ground transport network has gradually evolved from uncontrolled charging toward intelligent coordinated charging and V2G frameworks [51]. Early charging coordination approaches mainly relied on deterministic scheduling and price-based demand response mechanisms to shift charging demand away from peak periods [52]. However, such approaches often exhibit limited adaptability under highly dynamic transportation behaviors and uncertain electricity market conditions. To address these limitations, reinforcement learning (RL) and DRL methods have increasingly been adopted for adaptive road network traffic alignment and real-time energy management [53,54]. Recent studies demonstrate that DRL-based charging frameworks can effectively optimize charging cost, renewable energy utilization, and grid operational stability while simultaneously satisfying user mobility requirements and battery constraints [55,56].

More recently, research attention has shifted toward large-scale V2G aggregation and multi-agent EV coordination across complex urban road network nodes. Since EV fleets collectively possess considerable distributed storage capacity, V2G technologies enable the ground transport network to provide ancillary services such as frequency regulation, peak shaving, load balancing, and renewable energy support [57]. In this context, MARL and hierarchical coordination frameworks have emerged as promising solutions for large-scale EV aggregation under uncertain mobility patterns and distributed operating environments [58]. Existing studies further indicate that AI-enabled ground transport network coordination can significantly reduce charging costs, mitigate load fluctuations, and improve renewable energy accommodation capability in highly electrified urban energy systems [59,60]. Such developments demonstrate that the transport network is gradually evolving into an intelligent mobile energy network capable of dynamically interacting with future cyber–physical energy infrastructures.

2.1.3. UAV Infrastructure Operations

The rapid development of UAVs has significantly enhanced the intelligence and operational flexibility of modern power infrastructure inspection systems. Compared with traditional manual inspection approaches, UAV-assisted inspection systems provide substantial advantages in terms of operational efficiency, accessibility, safety, and real-time monitoring capability, particularly for geographically distributed transmission networks and complex terrain environments [61,62]. With the increasing deployment of renewable energy infrastructures and large-scale transmission corridors, conventional inspection methods often face limitations in handling high-frequency monitoring requirements and rapidly evolving operational conditions [63]. Consequently, AI-enabled UAV systems have gradually become an important component of intelligent power infrastructure management within modern smart grids. Existing studies indicate that UAV-based inspection frameworks are increasingly integrated with computer vision (CV) and deep learning techniques for automated defect detection and infrastructure monitoring [64]. Early UAV inspection systems mainly relied on image acquisition and manual interpretation of transmission line conditions [65]. However, recent advances in deep learning have significantly improved the capability of automatic feature extraction and defect recognition in UAV-captured images. CNNs, object detection frameworks, and semantic segmentation models have been widely adopted for identifying insulator damage, conductor faults, corrosion, vegetation intrusion, and structural abnormalities in transmission infrastructure [66,67]. Existing research demonstrates that AI-assisted defect recognition frameworks can significantly improve inspection accuracy while reducing manual labor intensity and operational costs [68].

More recently, research attention has shifted from static image recognition toward intelligent aerial sensing and autonomous UAV coordination. Since power transmission systems naturally exhibit large-scale spatial distribution characteristics, UAV swarms combined with edge intelligence and real-time communication frameworks have attracted increasing research interest for large-area infrastructure monitoring and adaptive inspection planning [69]. In particular, AI-enabled UAV systems integrated with simultaneous localization and mapping (SLAM), real-time path planning, and edge-based inference mechanisms have demonstrated strong capability in dynamic environment perception and autonomous navigation [70]. Such developments indicate that UAV systems are gradually evolving from simple aerial inspection platforms toward intelligent aerial sensing infrastructures capable of supporting adaptive monitoring, distributed coordination, and real-time cyber–physical awareness in future smart energy systems.

2.1.4. Cross-Network Coordination

The first three subsections review AI applications within representative physical infrastructure domains, whereas the final subsection focuses on AI-enabled coupling and coordination across these heterogeneous domains. With the increasing integration of electrified transportation systems, distributed renewable energy resources, and intelligent aerial infrastructures, modern urban energy systems are gradually evolving toward highly interconnected multi-network cyber–physical systems. In such environments, power grids, ground transport networks, UAV systems, and communication infrastructures are no longer operated independently, but instead exhibit strong coupling relationships through energy exchange, mobility interaction, and real-time information coordination [71]. Power systems provide electricity supply and distributed flexibility for EV charging infrastructures, UAV docking stations, and edge computing facilities, while ground transport networks contribute mobile storage capacity and flexible charging demand through EV fleets. Meanwhile, UAV systems extend computational capability and infrastructure awareness by supporting aerial inspection and flying edge computing services for latency-sensitive intelligent applications. These heterogeneous resources are therefore interconnected through energy exchange, computation offloading, information sharing, and mobility support, requiring coordinated optimization rather than isolated operation. Consequently, conventional single-domain optimization approaches often struggle to handle the large-scale uncertainty, dynamic resource allocation, and multi-agent interaction characteristics of future smart city infrastructures [72]. In this context, AI-enabled multi-network integrated coordination has emerged as an important research direction for enhancing operational flexibility and system-level intelligence in urban energy ecosystems.

Existing studies indicate that game-theoretic optimization and multi-agent coordination frameworks have become increasingly important for managing the interactions among ground transport fleets, UAV systems, and smart grid infrastructures. Since EV charging behaviors, UAV mission scheduling, distributed renewable dispatch, and edge computing resource allocation often involve multiple stakeholders with partially conflicting objectives, game-theoretic models have been widely adopted to achieve distributed decision-making and economic coordination under decentralized operating conditions [73]. In particular, Stackelberg games, non-cooperative games, and cooperative game frameworks have demonstrated strong capability in modeling charging competition, energy trading, UAV resource allocation, and distributed resource coordination in integrated smart energy systems [74].

The routine applications summarized in

Table 1. Synthesized Mapping of AI Core Architectures for Routine Multi-Network Optimization.

Network Dimension	Representative System Components	Core AI/ML Algorithms	Operational Metrics & Optimization Targets
Power Grid	Distributed Generation, Battery Energy Storage Systems, Substation Transformers	Long Short-Term Memory, Transformers, Spatiotemporal Graph Neural Network, Deep Reinforcement Learning/Multi-Agent Deep Reinforcement Learning	Maximize renewable hosting capacity; minimize short-term load forecasting error; optimize power flow distribution.
Ground transport network	Electric Vehicle Charging Stations, Public Roads, Fleet Charging Piles	Deep Reinforcement Learning, Hierarchical Reinforcement Learning, Multi-Agent Reinforcement Learning	Minimize peak-to-valley load ratios; reduce charging waiting times; mitigate spatial voltage deviations on distribution feeders.
Aerial Network	Unmanned Aerial Vehicles Swarms, Transmission Corridors, Inspection Sensors	Convolutional neural networks, YOLO-based Detection, Semantic Segmentation Networks, simultaneous localization and mapping	Maximize defect feature extraction accuracy; optimize autonomous 3D flight path trajectories; minimize manual labor costs.
Cross-Domain Coordination	Cyber–Physical Coordination Hubs, Vehicle-to-Grid Aggregators, Edge Computing Platforms	Stackelberg Games, Cooperative Multi-Agent Reinforcement Learning, Federated Learning	Maximize global operational efficiency; minimize cross-domain coordination latency; enable collaborative decision-making across heterogeneous infrastructures.

also differ in their computational and deployment characteristics. Forecasting and planning methods may involve substantial offline training but comparatively moderate online inference demand, whereas real-time UAV vision, multi-agent dispatch, and cross-network coordination require more continuous and latency-sensitive computation. Cloud-based approaches generally provide higher computing capacity and scalability through centralized resources but depend strongly on communication availability, whereas edge and onboard approaches improve responsiveness and autonomy while facing tighter limitations in hardware capacity, energy availability, and large-scale deployment. Because computational and energy metrics are reported inconsistently across the literature, the comparison focuses on relative deployment characteristics rather than direct numerical benchmarking.

In terms of deployment maturity, applications such as load forecasting, renewable-energy prediction, and image-based inspection are generally closer to practical deployment because they have been studied for longer and supported by more established datasets and implementation experience. In contrast, large-scale multi-agent coordination, cross-network AI dispatch, and fully autonomous emergency intelligence still face barriers related to communication reliability, computing resources, and system-level validation.

More recently, RL, DRL, and MARL techniques have been increasingly introduced into integrated multi-network systems for adaptive coordination and real-time economic dispatch [75]. Existing studies demonstrate that MARL-based coordination frameworks can jointly optimize EV charging schedules, UAV task allocation, renewable energy utilization, distributed storage management, and flying edge computing resources under highly dynamic operating conditions, enabling collaborative decision-making across coupled power–ground–aerial systems rather than independent optimization of individual infrastructures [76]. Furthermore, edge-cloud collaborative architectures and federated learning frameworks have recently attracted growing attention for supporting distributed intelligence and scalable coordination in large-scale urban cyber–physical systems [77]. Furthermore, the core algorithmic deployments, target infrastructures, and primary performance metrics of these routine operations are structurally cross-compared in Table 1. Such developments indicate that future smart energy systems are gradually transitioning from isolated infrastructure management toward deeply integrated “ground–aerial–power” collaborative ecosystems, where power systems provide energy supply and distributed flexibility, ground transport networks offer mobile storage and charging resources, aerial networks extend computational capability through intelligent inspection and flying edge computing, and AI serves as the cognitive coordination layer that synchronizes information exchange, resource allocation, and joint economic dispatch across the three interconnected networks.

2.2. Emergency Response and Disaster Recovery

The increasing frequency of extreme weather events, natural disasters, and large-scale infrastructure disruptions has significantly intensified the resilience challenges faced by modern urban energy systems. Under disaster conditions, conventional centralized operation frameworks often become vulnerable to cascading failures caused by power outages, communication interruptions, transportation constraints, and infrastructure damage [78,79]. Thus, research attention has gradually shifted from conventional economic optimization toward AI-enabled resilient operation and emergency response coordination in coupled power–transport systems [80]. Existing studies demonstrate that AI technologies can effectively support fault diagnosis, grid self-healing, emergency energy dispatch, autonomous disaster assessment, and cross-network resource coordination under highly uncertain and degraded operating conditions [81,82]. Existing evidence demonstrates that AI is gradually evolving from an operational optimization tool toward a critical resilience-enabling component for future cyber–physical energy infrastructures. Consistent with the taxonomy adopted for routine operations, emergency AI deployments are first reviewed within three representative physical infrastructure domains, namely power systems, ground transport networks, and aerial systems. Building upon these individual resilience functions, the final subsection focuses on cross-domain emergency coordination, where AI integrates heterogeneous infrastructures into a unified disaster-response ecosystem. To depict the tactical deployment of AI during crisis scenarios, Figure 2 outlines the cross-network emergency response framework, detailing the collaborative recovery flow through Section 2.2.1, Section 2.2.2 and Section 2.2.3 under structural failure conditions.

2.2.1. Power System Resilience

Under extreme weather events and large-scale infrastructure disruptions, maintaining the resilience and survivability of modern power systems has become a critical challenge for future smart grids. Unlike normal operating conditions that primarily focus on economic optimization and operational efficiency, disaster scenarios often involve cascading failures, communication interruptions, infrastructure damage, and severe supply–demand imbalance [83]. Consequently, AI-enabled resilient operation strategies have attracted increasing attention for enhancing the adaptive recovery capability of modern power systems under highly uncertain and degraded operating environments. Existing studies indicate that AI technologies have been widely adopted for fault diagnosis, self-healing control, and resilient microgrid operation during post-disaster restoration processes. Early intelligent fault management frameworks mainly relied on expert systems, rule-based decision-making, and optimization-based restoration methods for fault localization and service recovery [84]. However, such approaches often exhibit limited adaptability when facing dynamically evolving failure conditions and large-scale distributed outages. To address these limitations, ML and DL techniques have increasingly been integrated into fault detection and self-healing frameworks for improving fault classification accuracy, outage identification speed, and autonomous restoration capability [85,86]. Existing studies demonstrate that AI-assisted self-healing systems can effectively support adaptive feeder reconfiguration, fault isolation, and distributed restoration planning under uncertain post-disaster conditions [87].

More recently, research attention has shifted toward resilient microgrid islanding and AI-assisted black-start coordination. Since distributed renewable energy systems, BESSs, and local distributed generators can provide temporary autonomous operation capability during large-scale outages, microgrid islanding strategies have become an important solution for enhancing power system resilience [88]. In this context, DRL and MARL methods have demonstrated strong capability in adaptive island partitioning, distributed energy scheduling, and post-disaster restoration coordination under highly dynamic operating conditions [89]. Furthermore, AI-enabled black-start frameworks integrated with distributed decision-making and real-time situational awareness mechanisms have recently attracted growing research interest for accelerating power system recovery and improving restoration flexibility after extreme events [90]. These findings highlight that AI is gradually transforming smart grids from passive fault recovery infrastructures toward adaptive and self-healing resilient energy systems capable of autonomous emergency response and distributed restoration coordination.

2.2.2. Ground Transport Emergency Mobility

Under extreme weather events and large-scale disasters, the urban road transport network undergoes severe structural degradation, shifting the focus from routine dispatch to critical emergency survival. Within this stressed paradigm, EVs serve as vital assets for resilient road transport network operations, acting as both swift evacuation nodes and mobile emergency power suppliers under highly uncertain road blockages and grid interruptions [91,92]. Managing emergency mobility across the compromised road network layout requires AI-enabled coordination to simultaneously optimize cross-regional traffic evacuation, safeguard critical survival loads, and execute mobile energy delivery. Existing studies indicate that emergency ground transport frameworks increasingly integrate evacuation routing, charging coordination, and resilience-oriented dispatch strategies. During disaster evacuations, large-scale ground transport charging demand may introduce severe spatiotemporal stress on both traffic and distribution networks, potentially resulting in charging congestion, traffic bottlenecks, and cascading grid overloads [93]. To address these challenges, RL, DRL, and data-driven optimization frameworks have been increasingly adopted for adaptive evacuation routing and emergency charging coordination [94]. Existing research demonstrates that AI-enabled emergency charging frameworks can effectively optimize charging allocation, evacuation efficiency, and distribution system stability while considering traffic congestion, battery state-of-charge (SoC), and dynamically changing disaster conditions [95].

More recently, research attention has shifted toward utilizing ground transport vehicles as mobile energy storage and emergency power support resources during post-disaster restoration processes. Since ground transport vehicles collectively possess substantial distributed storage capacity and high mobility flexibility, V2G technologies enable these vehicles to provide temporary emergency power support for shelters, hospitals, communication infrastructures, and isolated microgrids during prolonged outages [96,97]. In this context, AI-driven mobile energy routing across the damaged road network and cross-regional ground transport coordination frameworks have emerged as promising solutions for improving resilience recovery capability under highly damaged infrastructure conditions [98]. Existing studies further indicate that coordinated road transport network mobile energy dispatch can significantly enhance restoration speed, reduce critical load loss, and improve post-disaster recovery flexibility in urban distribution systems [99]. Such developments indicate that ground transport networks are gradually transitioning from flexible transportation loads toward intelligent mobile energy carriers capable of supporting distributed emergency response and resilient energy recovery in future cyber–physical infrastructure systems.

2.2.3. UAV Emergency Response

Under post-disaster conditions, UAVs have gradually evolved from conventional aerial inspection platforms toward critical emergency response infrastructures capable of supporting disaster assessment, situational awareness, and communication recovery in damaged urban environments. Compared with ground-based emergency systems, UAVs possess significant advantages in terms of rapid deployment, high mobility, flexible sensing coverage, and accessibility in heavily damaged or inaccessible regions [100,101]. In large-scale disasters where transportation networks, communication infrastructures, and power systems may simultaneously experience severe disruptions, AI-enabled UAV systems have attracted increasing attention for improving emergency response efficiency and resilient situational awareness.

Existing studies indicate that AI-assisted UAV frameworks have been widely adopted for rapid disaster identification, post-disaster mapping, and autonomous environment reconstruction. Early UAV emergency response systems mainly relied on manual image interpretation and remote monitoring for disaster assessment tasks [102]. However, recent advances in CV, DL, and semantic segmentation technologies have significantly improved the capability of automated damage recognition and real-time disaster analysis using UAV-captured aerial imagery [103,104]. Existing studies demonstrate that CNNs, object detection models, and transformer-based visual frameworks can effectively identify collapsed buildings, damaged transmission infrastructures, flooded regions, wildfire propagation, and blocked transportation corridors under highly dynamic disaster conditions [105]. Furthermore, SLAM techniques integrated with UAV swarms have demonstrated strong capability in post-disaster three-dimensional (3D) reconstruction and autonomous environmental perception under partially destroyed infrastructure conditions [106].

More recently, research attention has shifted toward AI-enabled UAV communication relay and distributed aerial coordination in communication-constrained environments. Since post-disaster regions often suffer from severe communication outages and degraded network connectivity, UAV-assisted aerial communication systems have emerged as promising solutions for establishing temporary wireless coverage and emergency information transmission [107]. In this context, RL, DRL, and multi-agent UAV coordination frameworks have increasingly been adopted for adaptive flight path planning, distributed sensing, and resilient communication relay optimization under uncertain disaster environments [108,109]. Existing studies further indicate that edge intelligence and onboard AI inference mechanisms can significantly improve real-time disaster perception and autonomous coordination capability in UAV emergency response systems [110]. These studies indicate that UAV systems are gradually transitioning from isolated sensing platforms toward intelligent aerial emergency infrastructures capable of supporting distributed situational awareness, autonomous communication recovery, and real-time cyber–physical coordination in future resilient megacities.

2.2.4. Cross-Network Emergency Coordination

Under large-scale disasters and extreme operating conditions, the resilience of future urban infrastructures increasingly depends on the coordinated interaction among power systems, ground transport networks, communication infrastructures, and aerial sensing platforms. In such environments, isolated emergency management strategies often exhibit limited capability in handling cascading failures, multi-resource competition, and dynamically evolving disaster conditions [111]. Consequently, AI-enabled cross-network emergency coordination has emerged as a critical research direction for enhancing system-wide resilience and adaptive recovery capability in coupled cyber–physical infrastructures. Existing studies indicate that integrated air–ground collaborative rescue and distributed emergency resource allocation frameworks have attracted increasing research attention in disaster response systems. Since post-disaster environments often involve damaged transportation infrastructures, interrupted communication networks, and unevenly DERs, coordinated scheduling among ground transport vehicles, UAVs, and distributed energy systems has become increasingly important for improving emergency response efficiency [112,113]. In particular, UAV systems can rapidly provide aerial situational awareness and temporary communication coverage, while ground transport fleets equipped with V2G capability can support mobile energy delivery and emergency power restoration in affected regions [114]. Existing studies demonstrate that collaborative air–ground coordination frameworks can significantly improve rescue efficiency, reduce response delay, and enhance critical infrastructure survivability under highly uncertain disaster environments [115].

More recently, RL, DRL, and MARL techniques have been increasingly introduced into disaster-oriented multi-network coordination systems for adaptive emergency dispatch and distributed decision-making [116]. Existing studies show that MARL-based coordination frameworks can effectively optimize UAV swarm deployment, ground transport emergency routing, mobile energy allocation, and distributed communication relay under dynamically changing disaster conditions [117]. Furthermore, edge intelligence and federated coordination architectures have recently attracted growing attention for supporting decentralized emergency response and resilient situational awareness in communication-constrained environments [118].

In summary,

Table 2. Synthesized Mapping of AI Core Architectures for Emergency Multi-Network Coordination.

Emergency Subsection	Crisis-State Functional Role	Specialized AI Methodology	Targeted Resilience Goal & Post-Disaster Objective
Power System Resilience	Self-Healing & Black-Start Agent	Deep Reinforcement Learning, Expert Systems, Graph Neural Networks	Rapid fault localization, automated feeder reconfiguration, and islanded microgrid stability control under partial blackout.
Ground Transport Emergency Mobility	Mobile Energy Bank & Evacuation Router	Data-Driven Optimization, Online Reinforcement Learning, Spatial Heuristics	Alleviate emergency traffic gridlock; maximize evacuation efficiency; dispatch mobile energy storage to critical survival loads.
UAV Emergency Response	High-Mobility Sensor & Communication Relay	Deep Computer Vision, YOLO-based Detection, Multi-Agent Deep Reinforcement Learning, Simultaneous Localization and Mapping	Automated structural defect classification from aerial streams; real-time 3D disaster zoning; adaptive communication relay positioning.
Cross-Network Emergency Coordination	Coordinated Air–Ground Rescue Engine	Multi-Agent Reinforcement Learning, Edge Inference	Optimize global survival utility; manage multi-resource spatial competition; sustain ad-hoc network integrity under physical breakdowns.

compiles and evaluates the specialized AI methodologies, disaster-driven roles, and specific resilience objectives essential for cross-network survival, Such developments indicate that future resilient urban infrastructures are gradually transitioning from isolated emergency management frameworks toward deeply integrated “air–ground–power” collaborative ecosystems, where distributed edge intelligence and AI coordination become indispensable for emergency perception, resource allocation, and adaptive recovery, while simultaneously introducing unprecedented computing dependence on fragile physical energy infrastructures.

From a resilience perspective, the emergency applications in Table 2 can likewise be distinguished by task criticality, latency tolerance, computational intensity, and deployment maturity. Fault diagnosis and regional damage assessment may permit short scheduling delays, while obstacle avoidance, communication relay control, and immediate evacuation guidance have lower latency tolerance. Cloud offloading can reduce local computing demand when communication survives, whereas onboard or edge execution provides autonomy at the cost of tighter local power and thermal limits.

The preceding section synthesizes observations from the representative literature. The following section moves from that literature-derived foundation to the authors’ formulation of the Computation–Energy Paradox and then examines the proposed mechanism through a numerical case study.

3. The Overlooked Hidden Danger: The Conflict Between Computing Surges and Power Security Under Disasters

As AI evolves from isolated prediction tools toward a cross-network cognitive coordination layer, its dependence on distributed computing infrastructures becomes increasingly indispensable. While the preceding review highlights the immense utility of AI in crisis management, it reveals a pervasive assumption in the current literature: the presumption that computing power is an abstract, infinitely available resource. This assumption is fatally flawed. The integration of high-performance AI into disaster response creates a profound, physically grounded conflict. This section conceptually traces how the invocation of advanced algorithms can generate large and concentrated electrical loads that consume the remaining voltage and frequency operating margins of fragile post-disaster distribution networks.

3.1. Physical Interpretation: From AI Computation to Energy Demand

The fundamental hazard emerges from the intersection of computational physics, facility thermodynamics, and power systems engineering. Disaster management fundamentally transforms the cyber–physical workload from predictable, distributed background processes into concentrated, highly concurrent inference spikes. The imperative to save lives dictates that algorithms process massive volumes of complex data with near-zero latency. Consider the deployment of UAV swarms for post-earthquake search and rescue. Effective operations require high-resolution (1080p or 4K) real-time aerial video analytics to identify victims and structural obstacles. Advanced instance segmentation models, such as YOLOv8-seg or RTUAV-YOLO variants, are required to process these continuous video feeds at 30 frames per second (FPS) or higher [119,120,121]. In a coordinated response, continuous transmission from just 50 drones can easily saturate a 500 Mbps cellular uplink, necessitating heavy reliance on localized edge data centers. Simultaneously, evacuating a megacity involves tracking and routing tens to thousands of EVs. Dynamic routing models, such as ALD-EVRP, must continuously execute graph-based optimizations (e.g., dynamic Dijkstra) alongside traffic flow forecasting [122,123]. The combination of parallel video stream decoding and continuous combinatorial optimization results in a massive, synchronized call to the edge computing hardware.

The following analysis considers a high-intensity emergency scenario whose facility demand depends on workload composition, communication requirements, computing hardware, utilization, batching strategy, cooling conditions, and local energy availability.

The translation of this algorithmic complexity into physical electrical demand is highly non-linear. AI workloads, unlike general-purpose IT services, are characterized by ultra-low inertia, sharp power surges, and extreme computational density [124,125]. The dynamic energy consumption per inference request can be modeled through the relationship between active parameters (P_active) and batch size (B) [126]. For GPU-accelerated workloads, the energy function f_E can be generalized as [127]:

$$f_E(P_{\mathrm{active}},B)=\alpha e^{\beta B}{P}_{\mathrm{active}}+\gamma$$

(1)

where α, β, and γ represent hardware-specific regression coefficients. In standard cloud deployments, operators maximize energy efficiency by utilizing large batch sizes (e.g., B = 64 or B = 256), which amortizes the high fixed power costs of memory bandwidth and GPU activation [128,129,130]. However, in emergency disaster response, latency tolerance can be very limited for safety-critical tasks. Waiting to batch UAV video frames or EV routing updates would result in unacceptable delays [124,131]. Consequently, emergency edge servers are forced to operate at small batch sizes (e.g., B = 4), which is used here as a representative latency-constrained edge-inference configuration balancing real-time latency requirements and GPU utilization. At such small batch sizes, arithmetic intensity remains significantly lower than in conventional cloud inference, while GPU memory bandwidth and static power consumption remain largely unchanged, resulting in substantially higher energy consumption per inference. Modern accelerators, such as the NVIDIA A100 and H100, exhibit maximum power draws of 300 W to over 700 W per chip [132,133]. A single high-density AI rack can demand between 30 kW and 100 kW of continuous electrical power. When multiple racks are saturated by incoming disaster data, the server-level power draw (P_IT) instantly scales into the hundreds of kilowatts.

Furthermore, the raw IT power demand represents only a fraction of the actual load imposed on the electrical grid [134,135,136]. The total facility power (P_DC) is dictated by the Power Usage Effectiveness (PUE) metric:

$$P_{\mathrm{DC}}=P_{\mathrm{IT}}\times PUE=P_{\mathrm{IT}}+P_{\mathrm{cooling}}+P_{\mathrm{distribution}}$$

(2)

Under ideal conditions, hyperscale data centers achieve highly optimized PUE values around 1.10. In an extreme disaster scenario, however, physical damage to infrastructure often knocks out primary cooling towers and chilled water systems. To prevent the densely packed AI servers from catastrophic thermal throttling, the edge facility must rely on backup, highly inefficient localized cooling mechanisms, such as maximum-RPM forced-air fans [124,137,138]. For the emergency scenario, the PUE of the localized edge center is assumed to fall within 1.40–1.60 under degraded cooling conditions. The 1.67 MW IT load is an aggregated multi-rack engineering assumption rather than the measured demand of a single model, GPU, or server. Applying the assumed PUE of 1.50 yields an illustrative facility-level demand of approximately 2.50 MW. The evidentiary status and modeling role of the principal assumptions are summarized in

Table 3. Principal assumptions and evidentiary status of the emergency EDC.

Parameter	Illustrative Value	Basis or Evidentiary Status	Role in the Illustrative Scenario
Concurrent UAV video streams	50 streams	Scenario-specific engineering assumption	Represents a high-intensity emergency vision workload
Aggregate uplink demand	Approximately 500 Mbps	Scenario-level estimate associated with concurrent video transmission	Motivates reliance on localized edge processing
Inference batch size	B = 4	Representative latency-constrained inference assumption	Illustrates the efficiency–latency trade-off of small-batch inference
Accelerator power	Approximately 300–700 W per accelerator	Representative hardware range discussed in Section 3.1	Supports the rack-level IT-demand estimate
Aggregated IT load	1.67 MW	Engineering assumption for multiple high-density computing racks	Input to the facility-load calculation
Power Usage Effectiveness	PUE = 1.50	Degraded emergency-facility assumption	Accounts for cooling and other facility overhead
Total EDC demand	Approximately 2.50 MW	Calculated from IT load × PUE	Facility-level load imposed on the islanded microgrid
EDC connection point	Bus 29 in Area B	Selected case-study configuration	Defines the electrical location of the load step
Admission limits	Voltage: 0.90 p.u.; frequency: 59.30 Hz; |ROCOF|: 2.0 Hz/s; evaluation delay: 0.06 s; maximum provisional connection: 0.16 s	Selected short-timescale protection settings for the illustrative case	Determine acceptance or rejection of the attempted EDC load

.

The next physical link in this causal chain occurs at the distribution-network level. When a disaster strikes, line outages and degraded operating conditions can force the surviving feeder into a weakened radial configuration. The voltage profile of such a network can be characterized by the DistFlow branch relation, where the voltage drop between adjacent nodes i and i + 1 is approximated by (3). When the PCC is subsequently lost, the surviving feeder operates as an islanded microgrid with limited equivalent inertia. At the short electromechanical timescale considered here, the network and frequency dynamics are coupled: the radial power-flow solution determines the electrical power demand, while the coherent island-frequency state evolves according to the system-equivalent swing equation in (4) [139]. Consequently, a rapid EDC load increase raises the electrical power requirement before finite governor and storage responses can fully compensate, producing transient voltage, frequency, and ROCOF deviations.

$$V_{i+1}^2\approx V_i^2-2(R_i{P}_i+X_i{Q}_i)$$

(3)

$${\displaystyle \frac{2H_{\mathrm{sys}}}{f_0}}{\displaystyle \frac{d\Delta f}{dt}}={\displaystyle \frac{P_m-P_e}{S_{base}}}-D_{sys}{\displaystyle \frac{\Delta f}{f_0}}$$

(4)

In (3), R_i and X_i denote the resistance and reactance of branch i, while P_i and Q_i are the corresponding active- and reactive-power flows. In (4), H_sys is the equivalent inertia constant, f₀ is the nominal frequency, Δf is the coherent island-frequency deviation, P_m and P_e are the equivalent mechanical/grid-forming input and power-flow-determined electrical output, respectively, S_base is the system power base, and D_sys represents aggregate frequency damping. The electrical power calculated from the network solution is used directly in the electromechanical frequency dynamics.

The effective elasticity of emergency AI demand remains task-dependent. For low-flexibility, safety-critical services, abrupt curtailment may degrade UAV perception or delay routing updates, whereas lower-priority tasks can retain moderate flexibility through task prioritization, reduced resolution or update frequency, compression, delayed execution, or offloading. From the power-system perspective, the disturbance is not unique to AI hardware: a non-AI load with the same magnitude, connection point, power factor, and admission rate could produce a similar electrical response. The particular concern is that emergency computing demand may rise rapidly when the grid is already weakened, while curtailment may simultaneously interrupt safety-related services. According to (3), a large and rapidly admitted active-power surge, combined with increased effective feeder impedance, deepens the voltage drop; according to (4), the associated electrical-power imbalance simultaneously drives frequency and ROCOF excursions. This coupled response illustrates the Computation–Energy Paradox: the computational intelligence deployed to mitigate physical damage can itself consume scarce operating margin in the surviving islanded grid and thereby contribute to further service loss.

3.2. Simulation and Result Analysis

To translate the conceptual argument into a quantitative illustration, a dynamic case-study architecture is considered. We utilize a post-disaster IEEE 33-bus distribution system operating as a single islanded microgrid partitioned into three functional areas (Areas A, B, and C), coupled with an electrified transportation network, as illustrated in Figure 3 [140]. The scenario begins when an extreme flood severs the Point of Common Coupling (PCC), after which the surviving distribution feeder remains electrically connected internally but operates without upstream-grid support. Area A is primarily supported by local renewable generation and BESS, and Area B supplies critical loads through BESS and gas-turbine resources, while Area C relies on distributed renewable generation and local storage. Meanwhile, floods inundate vital transportation routes and interrupt power supply at several charging infrastructures, resulting in large-scale emergency mobility demand. A stationary, facility-scale emergency EDC, connected to the local distribution network in Area B, is immediately activated to coordinate AI-assisted emergency services through high-density GPU computing resources. In this case study, the EDC is modeled as a fixed computing facility rather than a mobile or vehicle-mounted unit. The principal computational, facility, and protection assumptions used in the case study are summarized in Table 3. With the physical configuration established, the scenario proceeds through three continuous 30-min stages: Pre-Disaster (0–30 min), Early Post-Disaster (30–60 min), and AI Saturation (60–90 min). Figure 4 traces the corresponding growth in emergency computing activity and facility-level EDC demand. The coupling is implemented at the scenario level rather than as one monolithic physical model. The transportation component defines emergency mobility demand, EV service conditions, AI-assisted routing requirements, and user-utility consequences, while the power-system component evaluates the response to EDC and charging-related loads. The two components interact through charging-station locations, electricity-dependent mobility services, the selected EDC connection point, and the loss of AI-assisted guidance after EDC rejection. UAV-vision, EV-routing, and sensor-processing tasks are aggregated into IT demand, after which PUE is applied to obtain facility-level demand.

To assess the grid impact of the 2.50-MW EDC demand identified in Figure 4, a short-timescale dynamic simulation was implemented in MATLAB R2024b. At each 0.01-s time step, a backward–forward-sweep power flow is solved for the weakened radial IEEE 33-bus network using the current EDC load and generation/storage outputs. The power-flow solution provides the bus voltages, network losses, and the electrical power supplied by the equivalent grid-forming source. This electrical power is then used in the system-equivalent swing equation in (4) to update the island frequency and ROCOF. The gas turbine in Area B provides governor-based active-power support, while the BESS units provide P–f active-power and Q–V reactive-power support within their capacity limits. The EDC is represented as a ramped constant-power load with a fixed power factor. The admission logic monitors the voltage at the connection bus, the island frequency, and ROCOF, and rejects the EDC load when a specified protection limit is exceeded or the maximum provisional connection time is reached. The model captures the short-timescale interaction between the distribution-network power flow and the main electromechanical response. The resulting baseline dynamic response is shown in Figure 5.

Using the selected baseline assumptions, emergency AI services are activated within Area B for UAV coordination, emergency routing, and cross-network situational awareness. The EDC reaches the assumed high-utilization condition under latency-constrained inference and a degraded PUE of 1.50. At t = 2.0 s, the resulting 2.50-MW facility-level demand attempts to connect at Bus 29. As shown in Figure 5, the attempted admission causes a clear short-timescale disturbance: the Bus 29 voltage reaches a minimum of 0.7470 p.u., the island frequency decreases to 59.812 Hz, and the maximum absolute ROCOF reaches 4.9259 Hz/s. The voltage falls below the 0.90-p.u. admission limit and the ROCOF exceeds the 2.0-Hz/s limit, while the frequency remains above the 59.30-Hz minimum-frequency setting.

At t = 2.06 s, the admission criteria reject the attempted emergency AI/GPU load, corresponding to a provisional residence time of 0.06 s. After clearing, the Bus 29 voltage returns toward its pre-event level, while the island frequency and ROCOF recover through the governor and BESS responses. The baseline case therefore shows that a rapidly activated emergency computing load can consume part of the limited operating margin of an already weakened islanded grid, even when admission protection prevents sustained connection.

Sensitivity to load magnitude and connection location. The baseline establishes the mechanism at one operating point; a focused comparison is therefore used to test whether the observed stress is confined to Bus 29 and 2.50 MW. Twelve single-point admission cases combine four Area B buses (26, 28, 29, and 31) with three facility-load levels (1.50, 2.00, and 2.50 MW). All cases use the same network model, electromechanical parameters, resource-control settings, and admission-protection logic as the baseline. Simultaneous multi-bus admission is not introduced, allowing the effects of load magnitude and electrical location to be interpreted separately. As summarized in

Table 4. Sensitivity of the EDC admission test to load magnitude and connection bus.

Connection Bus	EDC Load (MW)	Voltage Nadir at Connection Bus (p.u.)	Island Frequency Nadir (Hz)	Maximum |ROCOF| (Hz/s)	Rejection Time (s)
26	1.50	0.9129	59.902	2.4842	2.06
26	2.00	0.8936	59.867	3.3790	2.06
26	2.50	0.8734	59.831	4.3159	2.06
28	1.50	0.8718	59.899	2.5777	2.06
28	2.00	0.8376	59.861	3.5666	2.06
28	2.50	0.7995	59.820	4.6552	2.06
29	1.50	0.8460	59.897	2.6360	2.06
29	2.00	0.8006	59.857	3.6972	2.06
29	2.50	0.7470	59.812	4.9259	2.06
31	1.50	0.8105	59.895	2.7092	2.06
31	2.00	0.7403	59.851	3.9188	2.06
31	2.50	0.6407	59.794	5.5804	2.06

, increasing the EDC demand at Bus 26 from 1.50 to 2.50 MW decreases the connection-bus voltage nadir from 0.9129 to 0.8734 p.u., decreases the island-frequency nadir from 59.902 to 59.831 Hz, and increases the maximum absolute ROCOF from 2.4842 to 4.3159 Hz/s. For the same 2.50-MW demand, moving the connection from Bus 26 toward Buses 28, 29, and 31 progressively reduces the voltage nadir to 0.7995, 0.7470, and 0.6407 p.u., respectively.

The 12-case results show clear sensitivity to both load magnitude and electrical location. At each fixed connection bus, increasing the EDC load consistently deepens the voltage depression, lowers the frequency nadir, and increases ROCOF stress. Electrical location produces an additional effect. For the same 2.50-MW facility demand, moving the connection from Bus 26 to Bus 31 decreases the connection-bus voltage nadir from 0.8734 to 0.6407 p.u., decreases the island-frequency nadir from 59.831 to 59.794 Hz, and increases the maximum absolute ROCOF from 4.3159 to 5.5804 Hz/s. The much stronger voltage sensitivity toward the downstream buses indicates that electrical connection strength is particularly important for localized EDC admission.

All 12 cases are rejected at t = 2.06 s because the protection logic applies a fixed 0.06-s evaluation delay and the ROCOF criterion has already been exceeded by the first permitted protection check in every tested case. The identical rejection time therefore does not imply identical grid stress; instead, the voltage, frequency, and ROCOF trajectories quantify substantially different disturbance severities before the same admission action is issued. Because PUE enters the grid through ${\mathrm{P}}_{EDC}={\mathrm{P}}_{IT}\times \mathrm{PUE}$, alternative IT-load and PUE combinations are represented here through their resulting facility-level demand. Although these results remain conditional on the selected network and operating parameters, they demonstrate that the disturbance is not confined to a single EDC size or connection point. EDC admission should therefore account for both required facility demand and the electrical strength of the prospective connection location.

Beyond grid-side operating stress, the computation–energy paradox also propagates to end-user decision quality through AI service interruption. As shown in Figure 6, this section evaluates the decision-making utility of 300 representative EVs sampled from the broader traffic flow. In this context, the AI system assists users in complex decisions, such as route planning and charging station selection, to optimize time efficiency and meet energy demands. Consequently, a higher utility value reflects a greater satisfaction of comprehensive user needs, including time, energy, and financial costs [18]. The simulation reveals that AI significantly enhances performance: under normal conditions, AI optimization raises the average utility from −0.047 (without AI) to 0.354. Even during extreme scenarios, AI-supported guidance maintains a resilient average utility of 0.222. However, the principal user-level observation is the pronounced impact of AI-guidance failure. When trapped EVs lose all navigation and decision support in extreme scenarios, the average utility plummets to −0.163. This severe degradation highlights that without AI assistance, users fail to make effective evacuation decisions, leading to a surge in stranded vehicles and fundamentally damaging overall user utility. This indicates that algorithm-centric resilience strategies can be incomplete if they do not account for physical power constraints. The case results suggest that emergency AI workloads should be represented as dynamic physical loads rather than abstract computational resources. Effective islanded operation and admission-protection mechanisms are therefore essential for maintaining the resilience of AI-enabled power–transport systems under extreme disasters.

4. Key Challenges and Technological Outlook

The conceptual formulation of the “Computation–Energy Paradox” exposes a critical vulnerability inherent in modern urban infrastructure. However, translating this theoretical paradigm into actionable engineering solutions requires a profound deconstruction of the operational bottlenecks constraining system resilience. This section examines the fundamental reasons why these challenges remain intractable in current engineering practice and outlines forward-looking technological pathways. By adapting mature technologies from adjacent fields, this section aims to help address the fundamental conflict between high-intensity intelligent rescue operations and fragile physical infrastructures.

4.1. Analyzing the Core Challenges

A primary engineering bottleneck stems from the spatio-temporal energy–computation mismatch, which precipitates an intractable “energy island” dilemma during extreme events. The geographical epicenters of disaster zones, which demand the highest density of computational power for real-time disaster analysis and dynamic route planning [141,142], are precisely the regions where power infrastructure suffers the most catastrophic damage. Although recent literature has extensively considered post-disaster road blockages when exploring resilient routing strategies for mobile energy storage systems (e.g., heavy-duty power generation trucks) [5,143,144,145,146,147], a fundamental friction persists: the dynamic dispatch of physical energy resources inherently lags behind the instantaneously surging computational demands of emergency response. Consequently, this spatio-temporal decoupling manifests not only as a critical state of “computational demand without power supply” at the disaster epicenter, but also as “stranded energy capacity without computational utilization” within safe peripheral zones or isolated microgrids. As illustrated in Figure 7, this bilateral spatial decoupling creates a severe operational void, rendering localized, high-wattage AI inference unsustainable. Although literature [148] attempts to bypass ground obstacles using UAVs for wireless power transfer to edge nodes, the energy transfer efficiency of current aerial platforms remains grossly insufficient to sustain the continuous high-power consumption required for AI inference. Therefore, this fundamental spatio-temporal decoupling intrinsically paralyzes the deployment of intelligent rescue operations.

Compounding this spatio-temporal mismatch is the heterogeneous flexibility of emergency computational workloads. In conventional power grid operations, demand-side response mechanisms effectively mitigate power shortages by shedding or deferring non-critical loads (e.g., HVAC systems or routine EV charging) [149,150,151]. However, this elasticity paradigm must be adapted to each task’s urgency and safety relevance. During disaster response, AI tasks—such as autonomous UAV obstacle avoidance, real-time structural damage assessment, and dynamic life-saving dispatch—are low-flexibility demands for which latency and interruption may be unacceptable [4,152]. For such low-flexibility tasks, abrupt curtailment may reduce operational effectiveness and introduce safety risks; demand reduction should therefore be governed by task-aware priorities. Although recent literature extensively investigates energy optimization in mobile edge computing through flexible resource management—such as trading output accuracy via approximate computing [153], adopting elastic CPU frequency scaling [154], and implementing joint task caching strategies [155]—it must be emphasized that for genuinely life-critical applications, the acceptable latency or interruption window may be extremely limited. Accordingly, compute–power coordination should reserve capacity for safety-critical functions while reducing, delaying, compressing, or offloading less urgent workloads when energy availability is constrained. Conventional load-shedding remains relevant, but it must be supplemented by task-aware computational prioritization.

Furthermore, the physical boundaries of extreme edge environments impose severe thermodynamic and hardware constraints on AI deployment. During disasters, emergency response heavily relies on surviving mobile communication base stations, UAV edge nodes, and temporary command vehicles [156,157], yet the Uninterruptible Power Supply (UPS) capacities and cooling architectures of these facilities are extremely limited. Recent advancements in edge AI advocate for deploying high-performance, large-scale models on edge devices to enhance local autonomous decision-making capabilities [158,159]. However, forcibly executing such compute-intensive workloads in disaster-stricken, resource-constrained environments inevitably triggers rapid hardware thermal throttling or even catastrophic system crashes [160]. Although extensive research has explored dynamic thermal management (DTM) strategies, such as optimizing the thermal resistance of forced-convection heat sinks [161], implementing OS-level thermal-aware job scheduling [162], or relaxing strict temperature constraints to minimize data center HVAC energy consumption [163], these approaches are predicated on conventional operating environments. In a disaster scenario driven by continuous, compute-intensive emergency AI tasks, these flexible workarounds fall short, ultimately forcing the system to rely heavily on active cooling mechanisms (such as high-speed fans or liquid cooling pumps) to prevent immediate physical meltdowns. These mechanisms inherently consume substantial power. In an emergency context, activating these cooling mechanisms directly devours the already scarce UPS reserves, thereby recreating a fatal thermodynamic paradox. This objective physical constraint—a continuous trade-off between cooling demands and battery endurance—establishes a hard ceiling for edge AI performance, making the adoption of traditional high-precision AI paradigms physically unviable during a crisis.

4.2. Potential Technological Outlook

To overcome these physical bottlenecks, transitioning toward energy-aware Green Edge AI emerges as a critical technological pathway [164,165,166]. The core objective of this approach is to decouple emergency end-user devices from energy-intensive cloud architectures, enabling them to execute critical disaster response with ultra-low power consumption. Mature model compression technologies can be strategically adapted to meet these stringent requirements. As summarized in

Table 5. Comparative Analysis of Energy-Aware Edge AI Technologies for Emergency Scenarios.

Technology	Core Mechanism	Energy Reduction Potential	Strategic Advantages in Disaster Contexts
Model Pruning	Systematically removes redundant weights and neurons from pre-trained networks.	High (up to 60–70% reduction in FLOPs)	Enables high-performance models to bypass strict UAV memory and thermal limits, sustaining real-time autonomous obstacle avoidance before battery depletion.
Knowledge Distillation	Transfers intelligence from a large “teacher” model to a lightweight “student” model.	Moderate to High	Drastically cuts computational load while retaining core inference logic, allowing for rapid deployment on low-power, generic edge nodes.
Spiking Neural Networks (SNNs)	Utilizes discrete, event-driven binary spikes, mimicking biological neural processing.	Very High (orders of magnitude on neuromorphic chips)	Excels in continuous sensor monitoring and event-triggered processing; its ultra-low standby power perfectly aligns with prolonged grid outages.

, techniques such as model pruning and knowledge distillation can significantly streamline deep neural networks [167,168,169]. By systematically stripping away redundant network weights and transferring core intelligence to lightweight structures, these models drastically reduce computational overhead [170,171]. For example, this enhancement in energy efficiency directly translates into extended operational lifespans for battery-powered rescue equipment, ensuring that critical AI inference processes remain uninterrupted under severe power constraints. Furthermore, Spiking Neural Networks (SNNs) provide a highly promising neuromorphic alternative: by processing information through discrete, event-driven binary spikes rather than continuous floating-point values [172,173], SNNs can achieve orders-of-magnitude reductions in energy consumption. By mimicking the efficiency of biological nervous systems, SNNs can be seamlessly deployed on specialized neuromorphic hardware, offering a highly robust solution for continuous monitoring and decision-making during prolonged power outages [174]. The integration of these green AI technologies directly resolves the physical constraints of the extreme edge, offering a practical methodology for sustaining intelligent operations when traditional power infrastructure is paralyzed.

Meanwhile, addressing the spatio-temporal energy–computation mismatch requires a paradigm shift toward spatiotemporal energy–computation routing and mobility scheduling. This approach deeply integrates cross-disciplinary frameworks, such as the Computing Force Network (CFN) and Vehicle to Everything (V2X) [6,175,176,177], into disaster management systems. A highly viable strategy is the implementation of compute-coordinated task offloading mechanisms. Advanced computing force routing algorithms can be tailored to dynamically distribute computational workloads across geographically dispersed nodes. By offloading non-real-time, compute-intensive tasks (e.g., global damage mapping or post-disaster reconstruction simulations) from severely impacted epicenters to idle data centers in unaffected regions, the localized energy burden is profoundly alleviated, thereby freeing up precious power for zero-latency rescue operations. Crucially, extending this paradigm to Vehicle-to-Data-Center (V2DC) interactions and mobile compute–power vehicles provides a transformative and highly scalable solution [178,179,180]. Expanding upon the traditional use of idle EVs as distributed energy storage units, a massive fleet of EVs can simultaneously serve as high-capacity mobile power banks and distributed edge computing nodes [5,177,180]. As depicted in Figure 8, during a disaster, these vehicles can be rapidly dispatched to the accessible periphery of the disaster zone. Instead of attempting to navigate impassable roads to deliver physical batteries to the epicenter, they establish an ad-hoc computational perimeter—directly providing emergency “compute–power supply” to localized edge data centers. This approach ingeniously bypasses physical road blockages by substituting the slow physical transport of energy with the instantaneous wireless transmission of data and computation. Concurrently, to guarantee the seamless execution of V2DC and compute routing, resilient ad-hoc communication architectures—such as Satellite-Ground Integrated Networks (SIGNs) [181,182,183]—must be deployed synchronously. These networks ensure the highly reliable, low-latency transmission of complex data packets and critical energy dispatch commands.

Ultimately, overcoming these engineering challenges demands a holistic integration of the aforementioned advanced technologies. By seamlessly fusing lightweight emergency AI with dynamic spatiotemporal energy–computation routing mechanisms, the inherent conflict between high-intensity intelligent rescue demands and fragile physical infrastructures can be systematically resolved. This technological evolution transcends mere algorithmic upgrades; it represents a fundamental transition wherein AI operates not merely as a virtual entity, but as a critical physical component of a resilient “compute–power–transport” ecosystem, thereby guaranteeing the sustainable operational continuity of future megacities during catastrophic events.

These pathways are prospective and differ in maturity, implementation cost, and the bottleneck addressed. Model pruning and knowledge distillation are comparatively mature software-level approaches with relatively low deployment barriers, although their achievable savings depend on acceptable accuracy and latency. SNNs may provide larger efficiency gains on suitable neuromorphic hardware but remain less mature and hardware-dependent. V2DC requires coordinated vehicle, communication, charging, and control infrastructure, while satellite–ground networking entails substantial system-level investment. None of these pathways is quantitatively evaluated in the present illustrative simulation; their practical performance, cost, and interoperability remain subjects for future implementation-oriented research.

From an economic perspective, reducing, compressing, delaying, or offloading flexible workloads is generally the lower-cost short-term response when temporary losses in accuracy, resolution, or update frequency are acceptable. Additional edge-computing capacity is more relevant when processing latency or server saturation is the principal bottleneck, whereas storage, cooling, or backup generation becomes more valuable when local energy capacity and reserve are limiting. Investment is more likely to be justified when AI services are safety-critical, disasters are recurrent or prolonged, cloud offloading is unreliable, interruption costs are high, and the infrastructure also creates value during normal operation. A rigorous decision would require location-specific capital costs, energy prices, disaster probabilities, outage durations, and avoided-risk values; such techno-economic optimization is beyond the present scope.

This Perspective uses a structured and representative literature synthesis rather than a formal systematic-review or meta-analysis protocol. The numerical results depend on the selected workload intensity, hardware assumptions, batching configuration, PUE, grid topology, generation and storage availability, EDC location, and protection settings. The grid-side simulation uses a reduced-order electromechanical model for the illustrative case study. The IEEE 33-bus-based community cannot represent every city, edge architecture, or disaster condition, and the proposed technological pathways are not operationally evaluated in the present simulation. Broader validation requires additional computing architectures, system scales, resource mixes, connection locations, protection settings, and disaster scenarios.

5. Conclusions

This Perspective has examined the dual role of AI within the megacity power–transport nexus, tracing its trajectory from a cognitive enabler to a structural vulnerability. The key findings and future directions are summarized as follows:

• Literature-based observations on AI empowerment: AI has profoundly reshaped the operational landscape of urban power and transport systems. From routine multi-network optimization, such as EV cluster scheduling and UAV grid inspections, to emergency disaster response, AI acts as a critical cognitive enabler. It significantly enhances system efficiency, situational awareness, and operational resilience, demonstrating immense potential in driving the evolution of smart megacities.
• Conceptual contribution—the “Computation–Energy Paradox”: Despite its cognitive benefits, AI introduces a hidden structural vulnerability during extreme disasters. When the physical grid is severely damaged, the intensified invocation of emergency AI (e.g., real-time UAV video streaming and dynamic path planning) generates surging computational loads whose flexibility depends on task criticality. Paradoxically, this sudden megawatt-level energy demand exacerbates the existing power deficit, reducing the operating margin of islanded systems and potentially contributing to further service loss. In this context, AI transforms from a rescuing “brain” into a “super load” that can challenge the system’s remaining physical resilience margin.
• Case-study evidence and engineering implications: In the baseline case, an attempted 2.50-MW EDC load at Bus 29 reduces the voltage to 0.7470 p.u., produces a frequency nadir of 59.812 Hz and a maximum absolute ROCOF of 4.9259 Hz/s, and is rejected at t = 2.06 s. The sensitivity analysis further shows that the grid impact increases with EDC load and varies with the electrical location of the connection, indicating that emergency computing demand should be considered explicitly in post-disaster grid operation and admission decisions.
• Prospective technological pathways: To resolve these bottlenecks, future engineering practice may consider a targeted technological roadmap: deploying energy-aware Green Edge AI (e.g., model pruning and Spiking Neural Networks) to bypass physical hardware limits, and implementing V2DC compute routing to dynamically offload tasks, effectively substituting physical energy transport with wireless data transmission. Future work should further examine data-informed computing–grid models, coordinated EDC admission, and task-prioritization strategies under different disaster and infrastructure conditions.