PaperlayerView exact revision on arXiv

Source-linked AI summary

Electricity Demand and Grid Impacts of AI Data Centers: Challenges and Prospects

Xin Chen, Xiaoyang Wang, Ana Colacelli, Matt Lee, Le Xie

arXiv:2509.07218v6eess.SY

TL;DR

AI data centers are rapidly increasing electricity demand and creating grid challenges across planning, operation, markets, and real-time stability. This review synthesizes their infrastructure and workload patterns, then discusses coordinated solutions involving grids, data centers, and end-users, including flexibility and collaborative planning.

  • Problem

    Rapid AI growth is increasing data-center electricity demand and creating multi-timescale challenges for reliable power-system operation.

  • Method

    The paper provides a comprehensive review of AI data-center infrastructure, electricity-demand patterns, grid impacts, and potential solutions.

  • Results

    Inference accounts for approximately 60% of the AI energy footprint, training 30%, and preparation and fine-tuning the remaining 10%.

  • Takeaways & Limitations

    Grid-friendly AI operation requires coordinating flexibility from storage, onsite resources, converters, cooling, computing workloads, and end-user demand.

Abstract

from arXiv · show

The rapid growth of artificial intelligence (AI) is driving an unprecedented increase in the electricity demand of AI data centers, raising emerging challenges for electric power grids. Understanding the characteristics of AI data center loads and their interactions with the grid is therefore critical for ensuring both reliable power system operation and sustainable AI development. This paper provides a comprehensive review and vision of this evolving landscape. Specifically, this paper (i) presents an overview of AI data center infrastructure and its key components, (ii) examines the key characteristics and patterns of electricity demand across the stages of model preparation, training, fine-tuning, and inference, (iii) analyzes the critical challenges that AI data center loads pose to power systems across three interrelated timescales, including long-term planning and interconnection, short-term operation and electricity markets, and real-time dynamics and stability, and (iv) discusses potential solutions from the perspectives of the grid, AI data centers, and AI end-users to address these challenges. By synthesizing current knowledge and outlining future directions, this review aims to guide research and development in support of the joint advancement of AI data centers and power systems toward reliable, efficient, and sustainable operation.

I. INTRODUCTION

AI’s rapid expansion is creating unusually large, concentrated, and volatile electricity loads that challenge grid reliability and power quality. The paper reviews these loads and proposes coordinated solutions for reliable, sustainable integration.

  • Motivation: AI data-center growth is increasing grid pressure because advanced workloads require immense electricity consumption.Training GPT-3 was estimated at 1.29 GWh, while GPT-4 consumption rose substantially further.
  • AI Data Centers: AI data centers differ from conventional facilities through dense high-performance computing hardware optimized for large-scale AI workloads.They support workloads such as LLM training and inference using specialized computing infrastructure.
  • Load Characteristics: AI workloads create high-power-density loads, with racks reaching 30–100+ kW compared with 7–10 kW in traditional server racks.A ChatGPT query is estimated at 2.9 Wh, nearly ten times a regular Google search at 0.3 Wh.
  • Load Characteristics: Training, fine-tuning, and inference can produce sharp, difficult-to-forecast fluctuations, including hundreds of megawatts within seconds.These rapid changes complicate system balancing and reliable grid operation.
  • Grid Impacts: Power-electronic grid interfaces expose systems to low inertia, fast dynamics, harmonic distortion, and potential stability and power-quality problems.These effects differ fundamentally from conventional electromechanical loads.
  • Paper Scope: The review organizes AI data-center grid challenges across long-term planning, short-term operation and markets, and real-time dynamics and stability.It also presents solutions involving grid operators, data centers, and end-users.

II. OVERVIEW OF AN AI DATA CENTER

AI data centers combine specialized computing, storage, networking, power infrastructure, and cooling systems to support large-scale AI workloads. Their scale and grid connection requirements distinguish them from traditional data centers.

  • System Architecture: A typical AI data-center architecture integrates IT hardware, electrical power infrastructure, cooling systems, and supporting facilities.The architecture shown is illustrative rather than universal because transformer voltages and UPS configurations vary across facilities.
  • AI Versus Traditional Data Centers: Compared with traditional facilities, AI data centers can reach several hundred megawatts or gigawatt scale and support more computationally intensive workloads.Traditional enterprise sites commonly range from a few megawatts to around 10 MW, while large cloud facilities may reach about 100 MW.
  • IT Hardware: AI data centers use dense CPU, GPU, and TPU clusters housed in high-density server nodes and racks.This infrastructure enables scalable deployment, power distribution, and optimized cooling for large-scale computations.
  • Networking: High-speed networking connects storage systems, compute nodes, servers, and racks through technologies including NVLink, InfiniBand, and RoCE.Fat-tree and dragonfly topologies support high-throughput, low-latency communication.
  • Grid Power Supply: AI data centers typically connect to medium- or high-voltage distribution grids, while hyperscale facilities may connect directly to transmission networks.U.S. interconnection voltages range from about 13.2 kV for small sites to 115–230 kV for hyperscale facilities.

2) Uninterruptible Power Supply (UPS):

UPS systems maintain continuous AI data-center power through operating modes selected according to grid conditions. Backup generation and cooling support resilience and safe operation, while cleaner alternatives remain constrained by cost or intermittency.

  • Uninterruptible Power Supply (UPS): UPS systems provide instantaneous stored power during grid disruptions until backup generators start and synchronize with the load.Common storage technologies include VRLA, lithium-ion, and emerging solid-state batteries.
  • Uninterruptible Power Supply (UPS): AI data centers switch among bypass, double-conversion, and battery-backup modes according to grid voltage and power conditions.Illustrative thresholds place bypass at 0.9–1.0 pu, double conversion at 0.7–0.9 pu, and battery backup below 0.7 pu.
  • Backup Generation: Diesel generators remain common backup sources because they offer mature technology, low capital cost, high power density, and fast startup.Cleaner options include natural-gas generators, hydrogen fuel cells, onsite renewables, and battery storage.
  • Backup Generation: Hydrogen fuel cells have demonstrated megawatt-scale feasibility but require costly hydrogen infrastructure, storage, safety systems, and logistics.Demonstrations included a 3 MW prototype and a 1.5 MW system paired with battery storage.
  • Cooling Facilities: Cooling is essential for dissipating heat from high-density GPU and TPU clusters, and technology selection depends on density, efficiency, cost, design, and water availability.AI growth is accelerating adoption of liquid-based and intelligent cooling solutions.

D. Other Supporting Facilities

Supporting facilities and efficiency measures complete the AI data-center infrastructure. Electricity demand is growing rapidly, with IT equipment and cooling forming major consumption components and PUE providing a common efficiency measure.

  • Supporting Facilities: Supporting facilities include physical layout, security, networking, energy storage, water treatment, emergency lighting, and maintenance spaces.These systems support reliable, secure, and efficient operations beyond core IT, power, and cooling infrastructure.
  • Electricity Consumption: Global data centers consumed about 415 TWh in 2024, and demand is projected to reach around 945 TWh by 2030 with AI as the primary growth driver.The 2024 level represented about 1.5% of global electricity consumption.
  • Electricity Consumption Structure: IT equipment typically accounts for more than 60% of AI data-center electricity use, while cooling consumes approximately 30–40%.The remaining 10–30% supports facilities such as lighting, offices, monitoring, and security, though shares vary by design and operating conditions.
  • Efficiency: Power Usage Effectiveness (PUE) measures total facility electricity consumption relative to IT-equipment electricity consumption.Lower PUE indicates greater efficiency, with 1.0 representing all facility energy delivered to IT equipment.
  • Efficiency: Modern large-scale facilities achieve PUE below 1.3, while Google reports 1.09 and Meta reports 1.08 across their data centers.These values indicate that IT equipment constitutes the dominant electricity share.

B. AI Computing Load Patterns at Different Stages

AI computing is the dominant electricity-consuming component of AI data centers, but its demand patterns differ substantially across preparation, training, fine-tuning, and inference. Training sustains the highest loads, while fine-tuning and inference introduce burstiness, variability, and lifecycle-scale energy considerations.

  • 1) Preparation Stage:: Preparation covers model and data configuration, but its flexible, dispersed workflows make electricity consumption difficult to quantify precisely.Data processing choices, deployment environments, and workflow structures substantially affect preparation-stage energy use.
  • 2) Training Stage:: Training is the most electricity-intensive stage, with GPT-3 estimated at 1.29 GWh and GPT-4 at over 50 GWh.GPT-3 involved approximately 14.8 days of continuous computation on 10,000 NVIDIA V100 GPUs; GPT-4’s estimate is about 40 times higher than GPT-3’s.
  • 2) Training Stage:: Training loads ramp up initially, remain near maximum for prolonged periods, and exhibit large swings from computation, communication, checkpointing, transfers, and pauses or resumptions.Sustained full-power operation also produces substantial heat, increasing cooling demand and total energy use.
  • 3) Fine-tuning Stage:: Fine-tuning generally uses less electricity than pre-training but has moderate, bursty demand and variable cooling loads during computation, validation, and hyperparameter exploration.Because fine-tuning occurs frequently across organizations, its cumulative lifecycle energy consumption can become non-negligible; GreenTrainer reduces workload by up to 64% without significant accuracy loss.
  • 4) Inference Stage:: Inference is least intensive per query but can exceed training’s cumulative energy use at scale, with unpredictable bursts and strong daily usage patterns.At Google, inference accounts for approximately 60% of total AI energy usage, compared with 40% for training; per-query use varies with model, prompt, and task complexity.
  • 4) Inference Stage:: Figure 4 illustrates GPU-power load patterns across training, fine-tuning, and inference using data derived from prior experimental measurements.The detailed source settings are referenced through Figures 9, 13, 15, and Table III in [11].

C. Available Datasets, Summary, and Discussion

AI computing electricity demand varies substantially across workflow stages, while available measurements and datasets remain concentrated at laboratory and device scales. The paper connects these load characteristics to planning challenges and discusses coordinated infrastructure responses.

  • Available Datasets and Load Characteristics: Inference accounts for approximately 60% of AI’s energy footprint, compared with 30% for training and 10% for model preparation and fine-tuning.
  • Available Datasets and Load Characteristics: Existing datasets primarily provide laboratory-scale experiments and GPU-level measurements rather than full-scale operational data from commercial AI data centers.
  • Available Datasets and Load Characteristics: Measured AI loads vary across training, fine-tuning, and inference, with inference characterized by short, highly variable bursts driven by real-time requests.
  • Long-Term Planning and Interconnection: Fifteen U.S. states accounted for about 80% of national data center load in 2023, intensifying regional grid stress and upgrade needs.
  • Long-Term Planning and Interconnection: Transmission and distribution upgrades may require 5-10 years, creating a timing challenge for projected AI data-center demand growth.
  • Long-Term Planning and Interconnection: Co-planning data centers, generation, and grids can address investment, congestion, reliability, and decarbonization tradeoffs under uncertain future conditions.

B. Short-Term Operation and Electricity Markets

AI data centers introduce substantial short-term operational and electricity-market challenges because their tens-to-hundreds-of-megawatts demand is large and bursty. These effects increase balancing, reserve, capacity, and cost-allocation pressures.

  • Short-Term Operation and Electricity Markets: Tens-to-hundreds-of-megawatts AI data-center loads create substantial hourly-to-weekly operational and market challenges.
  • Short-Term Operation and Electricity Markets: Unit commitment, economic dispatch, and reserve scheduling must maintain power balance while accommodating immense and bursty AI demand.
  • Short-Term Operation and Electricity Markets: $329.17/MW-day was the PJM capacity-market clearing price for 2026-2027, over ten times the $28.92/MW-day price for 2024-2025.
  • Short-Term Operation and Electricity Markets: Market design must preserve resource adequacy, incentivize grid-supportive operation, and allocate AI-driven infrastructure costs fairly.

1) Grid Disturbance Ride-Through:

AI data centers’ converter-based interfaces and workload dynamics shape their response to voltage disturbances and rapid power changes. Ride-through, converter controls, energy buffering, and coordinated stabilization are therefore central to limiting grid impacts.

  • Grid Disturbance Ride-Through: Voltage sags and swells can induce power oscillations or disconnect AI data centers, making adequate fault ride-through essential.
  • Grid Disturbance Ride-Through: For a given computing load, reduced grid voltage requires higher current, increasing the likelihood of converter current limits and protection actions.
  • Grid Disturbance Ride-Through: During sustained voltage sags, the simulated data center transitions from bypass to double-conversion and then battery-backup mode before reconnecting after recovery.
  • Grid Disturbance Ride-Through: Larger electronic-load trips produce more severe frequency excursions, indicating that insufficient ride-through can worsen grid disturbances.
  • Power Fluctuations and Grid Stability: Synchronous GPU computation and communication phases can create rapid power swings that challenge frequency stability and excite oscillations.
  • Power Fluctuations and Grid Stability: The study uses PSCAD EMT simulations to relate measured device-level IT profiles to facility-level grid-side power under two rectifier-control cases.
  • Power Fluctuations and Grid Stability: High-damping rectifier control smooths fluctuations, whereas low damping can amplify rapid variations and transient overshoot before they reach the grid.
  • Power Fluctuations and Grid Stability: Higher damping reduces oscillations but can slow disturbance tracking, while larger DC-link capacitance buffers fluctuations at greater cost and size.

3) Power Quality Issues:

AI data centers create cross-cutting cyber-physical, decarbonization, and power-quality challenges as concentrated, programmable, power-electronics-based loads expand. Addressing them requires coordinated security, operational safeguards, clean-energy resources, storage, flexibility, and power-quality management.

  • Power Quality Issues: Nonlinear power supplies, AC/DC conversion, and server power-factor correction can introduce harmonic distortion that degrades local power quality and threatens equipment safety.
  • Cybersecurity: Concentrated demand and programmable power electronics expand AI data centers’ cyber-physical attack surface, while intrusions can produce fast physical grid disturbances.
  • Cybersecurity: A 2024 Virginia event disconnected 60 of more than 200 data centers after a protection-system failure, illustrating the scale of correlated disconnection risk.
  • Cybersecurity: Recommended defenses include authenticated communications, network segmentation, joint cyber-physical monitoring, signal validation, ramp limits, staged reconnection, and ride-through requirements.
  • Decarbonization: Global data-center electricity emissions reached 180 Mt in 2024 and are projected to reach 300 Mt by 2035.
  • Decarbonization: Continuous AI electricity demand creates a temporal mismatch with variable renewable generation, requiring firm low-carbon resources, long-duration energy storage, and flexibility for real-time matching.
  • Decarbonization: Decarbonization strategies include co-located renewables, PPAs, RECs, and emerging 24/7 carbon-free energy procurement, but scaling requires additional generation, transmission, storage, and flexibility.

3) Water Consumption:

AI data center expansion creates substantial water demands whose impacts depend on local and seasonal water stress. Cooling and electricity-supply choices, workload scheduling, and water-aware infrastructure can reduce these pressures.

  • Water Consumption: A 100 MW U.S. data center consumes around 2 million liters of water daily, equivalent to about 6,500 households.Water is primarily associated with heat removal from high-density computing equipment.
  • Water Consumption: Data center water impacts depend on where and when water is consumed, not only on total volume.Assessment frameworks account for spatial and temporal variations in water stress.
  • Water Consumption: Dry heat rejection and direct-to-chip, immersion, or hybrid cooling paired with low-water heat rejection can reduce evaporative-cooling reliance.Improved water treatment, concentration management, and reclaimed or non-potable water can further reduce freshwater withdrawals.
  • Water Consumption: Water consumption is a critical constraint on AI data center expansion, especially during summer peaks when water and power systems face simultaneous stress.Nearly two-thirds of new U.S. data centers built since 2022 are located in high water-stress regions.
  • Forecasting and Coordination: Accurate AI load forecasting requires modeling workload schedules, training durations, hardware utilization, compute intensity, and cooling demand.Many relevant predictors, including job queues and cooling states, are internal to data centers and unavailable to grid operators.
  • Forecasting and Coordination: Collaborative forecasting combines private data center telemetry with grid-side weather, congestion, price, and reserve information.Machine-learning methods can use these inputs to predict short-term fluctuations and identify peak-demand events.

2) Dynamic Modeling of AI Data Centers and Available Open-Source Models:

AI data centers require detailed dynamic models because converter controls, UPS systems, protection, cooling, and workload timing shape their grid interactions. Existing modeling practice remains incomplete, motivating multi-timescale representations and tailored demand-response and interconnection requirements.

  • Dynamic Modeling: AI data center dynamics are governed by converter controls, UPS operation, protection logic, cooling dynamics, and workload timing.These models support assessment of ride-through, power quality, stability, and reliability interactions.
  • Dynamic Modeling: Dynamic modeling, validation, and testing for emerging large data center loads remain under development.NERC guidance identifies limitations in existing dynamic load models for representing such loads accurately.
  • Modeling Timescales: EMT models capture fast converter controls, waveform phenomena, unbalanced conditions, protection, and mode transitions, while phasor models support wide-area studies.Fully detailed EMT simulation may be computationally impractical for large-scale system analysis.
  • Demand Response: Price-based demand response may be ineffective because deferred training or inference can be rescheduled and opportunity costs may exceed energy-arbitrage savings.AI data center operators may also be relatively insensitive to electricity prices because of high sunk costs and lucrative workloads.
  • Demand Response: Incentive-, penalty-, and contract-based programs can specify curtailable load, response duration, ramp limits, recovery constraints, and service-quality impacts.These mechanisms exchange committed flexibility for compensation, reduced charges, or priority grid services.
  • Standardization and Regulation: Interconnection standards may require fault ride-through, voltage and frequency tolerances, reactive capability, harmonic limits, ramp-rate limits, staged reconnection, and contingency responses.Proposed UPS specifications include operation at 50–70% of nominal voltage with resynchronization within one second.
  • Workload Flexibility: Temporal and spatial workload shifting can align AI demand with low grid stress, renewable availability, cleaner electricity, lower prices, or reduced congestion.Prior research indicates that even modest spatial shifting can produce significant emissions reductions.

2) Hybrid Energy Storage Solutions:

AI data centers need storage and coordinated on-site resources to manage rapid, large power fluctuations and provide grid flexibility. Complementary storage technologies and computing, cooling, and hardware efficiencies address different parts of their energy footprint.

  • Hybrid Energy Storage: Hybrid energy storage systems combine technologies with complementary characteristics because no single technology meets all AI load requirements.Relevant requirements include rapid response, sufficient power and energy capacity, long cycle life, and cost-effectiveness.
  • Hybrid Energy Storage: Supercapacitors absorb sub-second server- or rack-level spikes, batteries provide longer-duration buffering, and flywheels smooth aggregated short-term fluctuations.The technologies differ in response time, energy density, duration, efficiency, and cycling characteristics.
  • On-Site Resources: On-site generators, renewables, UPS systems, and storage can provide dispatchable demand-side support when coordinated with grid operators.These resources are primarily deployed for reliable operation during grid disturbances.
  • Computing Efficiency: AI computing is the largest electricity-consuming component of AI data centers, making computing efficiency and management central to reducing energy use and stabilizing demand.The paper presents complementary strategies spanning computing, models, hardware, cooling, and heat reuse.
  • Computing Efficiency: Under-clocking a single NVIDIA A100 GPU reduced power consumption by 40% for a specific application while decreasing performance.Energy-aware training and dynamic voltage and frequency scaling are additional computing-efficiency approaches.
  • Model and Hardware Efficiency: Pruning, quantization, and knowledge distillation can reduce computational demand while preserving output quality in smaller task-specific models.The passage cites TinyBERT as achieving over 90% lower energy consumption.
  • Model and Hardware Efficiency: Energy-efficient processors, domain-specific accelerators, low-power CPUs or GPUs, and AI-optimized hardware improve performance per watt.The paper identifies Google TPUs as an example of AI-optimized hardware.
  • Cooling and Heat Reuse: Hybrid air-liquid cooling manages dense AI thermal loads while balancing energy use, complexity, and adaptability; heat reuse can supply nearby thermal demands.Liquid cooling can provide higher-temperature streams that are more suitable for heat recovery.

6) Retrofitting Existing Data Centers for AI Workloads:

Retrofitting existing data centers offers a faster near-term route to AI capacity by leveraging existing infrastructure, but higher AI power density and dynamic behavior make upgrades technically demanding. Energy-aware software, user flexibility, and coordinated demonstrations complement facility-level expansion strategies.

  • Retrofitting Existing Facilities: Retrofitting can shorten deployment timelines and leverage existing land, fiber, security infrastructure, and grid interconnections.It reduces upfront development uncertainty compared with building entirely new facilities.
  • Retrofitting Existing Facilities: AI retrofit projects must assess structural, electrical, mechanical, thermal, and operational constraints because AI servers impose higher density and stronger dynamic profiles.The passage also highlights stronger thermal gradients than conventional workloads.
  • Energy-Aware Usage: Energy-aware prompting and appropriately scaled models can reduce per-query computation while preserving task objectives or comparable performance.Longer prompts increase token processing and energy use, while lightweight task-specific models can cost a small fraction of foundation models.
  • Energy-Aware Usage: Providers can use prompt optimization, model routing, energy metrics, labels, pricing incentives, credits, and delayed-processing discounts to support energy-aware behavior.The AI Energy Score project is cited as an initiative for benchmarking energy cost per query.
  • User Flexibility and AI Demand Response: AI Demand Response uses users’ tolerance for delayed results to shift query processing toward power-system needs.The paper proposes AI-DR as an analogue of electric demand response for user-side workload flexibility.
  • User Flexibility and AI Demand Response: AI-DR mechanisms must account for service quality, privacy, user acceptance, hardware reliability, incentive compatibility, and grid coordination.Latency-sensitive requests should remain distinct from delay-tolerant workloads that can be shifted or assigned to lower-power resources.
  • Demonstrations: Industry demonstrations organize computing-electricity coordination around workload flexibility, including shifting delay-tolerant workloads across time and locations.Google’s demonstrations include carbon-aware and demand-response computing.

E. Integrated Solution Framework and Practical Deployment

AI data center deployment requires coordinated solutions across end-user demand, data center infrastructure, and power-system planning and operation. The paper emphasizes flexibility, fair incentives, adaptive investment, and multidisciplinary co-planning, while noting uncertainty in future demand and gaps in grid-friendliness metrics.

  • Hierarchical demand chain: AI end-user requests shape the timing, volume, and computational intensity of demand, which data centers translate into computing, cooling, and power-infrastructure requirements.The resulting demand chain couples end-user preferences, facility operation, and grid conditions.
  • Planning and deployment: Future AI electricity demand is highly uncertain because model architectures, hardware efficiency, inference demand, workload strategies, and provider business decisions change rapidly.This uncertainty complicates decisions about when, where, and at what scale grid upgrades should be built.
  • Metrics and standardization: Robust, validated metrics beyond PUE and WUE are needed to capture flexibility, reliability impacts, carbon footprint, water use, and local-grid interactions.Developing such metrics remains an important direction for research and standardization.
  • Technical flexibility: AI data centers contain flexibility through storage, dispatchable resources, converter control, cooling, computing workloads, and end-user demand.With suitable standards, controls, and grid coordination, this flexibility can support disturbance ride-through and mitigate fast power fluctuations.
  • Markets and regulation: Fair cost allocation and flexibility incentives can help electricity markets accommodate large AI data center loads while encouraging grid-supportive operation.Proposed incentives include shifting delay-tolerant workloads, reducing demand during system stress, dispatching on-site resources, and using storage for peak shaving and ramping mitigation.
  • Multidisciplinary coordination: Planning and operation span computing, AI algorithms, power electronics, grid infrastructure, markets, environmental constraints, policy, economics, and engineering.The paper calls for integrated co-planning of the compute-electricity-water-carbon nexus and collaboration across the relevant stakeholder community.

APPENDIX A GLOBAL AI DATA CENTER LOADS

Global AI data center electricity demand is expanding across the United States, China, and Europe, prompting region-specific planning and sustainability responses. China emphasizes coordinated relocation and energy planning, while Europe combines growth with climate-neutrality and resource-efficiency objectives.

  • Global overview: Global data centers consumed around 415 TWh in 2024, representing about 1.5% of total global electricity consumption.The United States accounted for 45% of this demand, followed by China at 25% and Europe at 15%.
  • United States: U.S. data center energy consumption reached 176 TWh in 2023 and is projected to reach 325-580 TWh by 2028 under different growth scenarios.U.S. data center load is also projected to rise from 25 GW in 2024 to over 80 GW by 2030, largely driven by AI workloads.
  • China: China’s data center electricity consumption is estimated at 100-150 TWh in 2024 and projected to rise to 400-600 TWh annually by 2030.The Eastern Data, Western Computing initiative aligns computing deployment with regions having renewable resources, favorable cooling, and grid-development advantages.
  • China: China’s coordinated computing-hub strategy is expected to reduce data center-sector emissions by 16-20% by 2030 through cleaner electricity, natural cooling, and lower-cost resources.The country has approved eight national computing hubs and ten national data center clusters.
  • Europe: European data center IT load is projected to increase from approximately 10 GW in 2024 to 35 GW by 2030, while annual electricity consumption rises from 62 TWh to over 150 TWh.This would represent about 5% of Europe’s total electricity use by 2030.
  • Europe: European sustainability efforts address climate neutrality, energy efficiency, carbon-free procurement, water conservation, equipment circularity, and heat reuse.Northern and Nordic regions are attractive because of low-carbon electricity, cooler climates, and potential waste-heat utilization.
Loading 2509.07218v6…