{"id":88444,"topic":"ai","source":"Nature","title":"Artificial intelligence data centers could reach one percent of global electricity demand by 2030 - Nature","url":"https://www.nature.com/articles/s44458-026-00152-5","url_hash":"1d7fa503ae177bc69cce173d1cec1d5999a11c47","author":"","summary":"<a href=\"https://news.google.com/rss/articles/CBMiX0FVX3lxTE52LWVBcEJFVm5Dcm80UWNfaVNhRy04Qmt0OGhyNklYcGxGYVpCaE1NWFFGak0wWndxWW5FS0RZZU1HaTNKSmdjWnZDbGZKZWZWMkNRN2RBSG9QMmhQWVFj?oc=5\" target=\"_blank\">Artificial intelligence data centers could reach one percent of global electricity demand by 2030</a>&nbsp;&nbsp;<font color=\"#6f6f6f\">Nature</font>","content":"Abstract\nThe rapid growth of generative artificial intelligence is increasing global electricity demand and placing new pressure on power systems. Here we show that electricity use by data centers built for artificial intelligence could rise from about 118 terawatt-hours in 2024 to between 239 and 295 terawatt-hours by 2030, or about 1% of global electricity demand. New computing infrastructure is highly concentrated in North America, Western Europe, and the Asia-Pacific, which together account for more than 90% of projected computing capacity. Some regions, including Oregon, Ireland, and Iowa, face greater pressure from concentrated data-center loads, whereas larger systems such as Texas can absorb new demand more effectively. These results indicate that artificial intelligence infrastructure is becoming a structural part of power-system dynamics. The study combines large language model analysis of corporate, policy, and media sources with scenario-based projections of future electricity demand.\nSimilar content being viewed by others\nIntroduction\nThe rapid emergence of generative artificial intelligence (AI) and large-scale data analytics has driven a sharp expansion of computational demand1,2. As AI models grow exponentially in size and complexity, their training and inference require vast computing power and data throughput, driving record investment in high-performance data centers and digital infrastructure (Figs. S1 and S2)3,4,5,6,7. According to Bain & Company’s Global Technology Report 2025, sustaining the computational requirements of AI expansion could generate nearly US$2 trillion in annual revenues by 2030–equivalent to the combined GDP of the world’s ten largest emerging economies8. This “AI infrastructure boom” is transforming data centers into the industrial backbone of the digital economy and, increasingly, a major source of electricity demand9,10.\nElectricity use has risen sharply in parallel with the digital transition. The International Energy Agency (IEA, 2025) projects that global data-center electricity consumption will more than double—from about 415 TWh in 2024 to roughly 945 TWh by 2030—with the United States and China accounting for almost 80% of the increase11,12. AI-specific facilities rely on GPU-based computation, which enables large-scale parallel processing but consumes up to six times more power than conventional racks, elevating both cooling intensity and peak-load requirements. These facilities are increasingly concentrated in regions with abundant renewable resources, low electricity prices, and favorable climates, yet such clustering also amplifies local grid stress and transmission constraints. As AI campuses scale up from megawatt to gigawatt levels, ensuring a reliable and low-carbon electricity supply has become a strategic challenge for utilities, regulators, and technology developers13,14.\nA growing body of literature has examined the energy footprint of digital infrastructure. Early studies quantified electricity consumption and emissions associated with information and communication technologies15,16,17, while more recent work highlights that the growth of AI workloads—particularly large-scale GPU clusters and generative models—may substantially increase electricity demand and associated environmental impacts18,19,20,21,22.\nTo quantify these impacts, prior research has developed two main classes of modeling approaches. Top-down methods estimate electricity consumption using aggregate indicators such as data traffic, computing capacity, or ICT statistics15,16,17. While these approaches provide useful global benchmarks, they often struggle to capture the heterogeneous and rapidly evolving workloads associated with AI computing. Bottom-up engineering models, in contrast, estimate energy use from facility-level characteristics, including server hardware configurations, cooling systems, and operational efficiency metrics19,23. However, such models are frequently constrained by limited data availability and incomplete knowledge of proprietary hyperscale infrastructure deployments20. Recent studies attempt to bridge these perspectives by linking computational workloads to energy demand through task-based or model-level accounting frameworks20,24,25, complemented by scenario analyses such as the IEA’s Energy and AI report11.\nDespite these advances, existing approaches remain largely focused on aggregate demand estimation or facility-level energy accounting, providing limited insight into how firm-level infrastructure investment, spatial clustering of AI data centers, and regional electricity-system characteristics jointly shape the geography and magnitude of AI-related energy demand.\nThis study addresses this gap by developing an integrated analytical framework that links AI infrastructure deployment, spatial siting patterns, and regional power-system impacts. Unlike conventional approaches that assume stable relationships between computing intensity, hardware efficiency, and electricity use—an assumption that breaks down under rapidly evolving AI workloads26,27,28—our framework employs a large language model (LLM)-based semantic retrieval and inference to dynamically extract firm-level strategic signals from heterogeneous corporate disclosures. These signals are combined with scenario-based electricity-demand projections to assess how AI data-center expansion translates into regional electricity demand and grid pressure. This approach moves beyond aggregate demand projections and facility-level efficiency metrics to capture how firm-level AI deployment strategies and spatial clustering of data centers generate localized electricity-system pressures.\nEmpirically, we focus on six leading technology firms—Amazon, Microsoft, Google, Meta, Oracle, and Apple—which collectively account for ~70−75% of global hyperscale and cloud-linked data-center electricity demand (Table S1). Their global footprint and relatively transparent reporting enable consistent identification of siting patterns and electricity-demand trajectories across regions.\nUnderstanding and managing this emerging compute-energy nexus is critical to ensuring that the rapid expansion of AI infrastructure evolves in tandem with the development of adequate and reliable electricity systems14. In this study, the compute-energy nexus is modeled as a demand-side relationship: we quantify and spatially allocate the electricity consumption generated by AI data-center expansion. Bidirectional interactions—such as demand response, flexible load scheduling, or participation in electricity markets—are recognized as important dimensions of this nexus but fall outside the scope of the current framework, and are identified as directions for future research. Accordingly, this study addresses three central research questions: (i) what spatial factors shape the siting patterns of large-scale AI data centers; (ii) how firm-level infrastructure investments translate into electricity demand trajectories; and (iii) what regional electricity-system pressures arise from the spatial clustering of AI data-center loads.\nAddressing these questions, we find that AI-oriented infrastructure is concentrating in a limited set of regions—chiefly North America, Western Europe, and the Asia-Pacific, which together account for more than 90% of projected compute capacity—and that aggregate electricity consumption by six leading operators is projected to rise from roughly 118 TWh in 2024 to between 239 and 295 TWh by 2030, equivalent to about 1% of projected global electricity demand. This concentration is associated with markedly higher relative electricity-demand pressure in some host regions, such as Oregon, Ireland, and Iowa, than in regions with larger electricity systems, such as Texas. Together, these findings indicate that the geography of AI compute is becoming a structural feature of regional power-system dynamics rather than a peripheral digital load, underscoring the value of anticipatory planning that aligns computational growth with electricity-system development.\nResults\nGlobal clustering and heterogeneity in AI data-center siting\nAI data-center locations are identified using a large language model (LLM)-based inference approach that extracts and classifies site-level information from public disclosures, assigning each candidate site a probability of AI-oriented deployment (see Methods). Reported siting patterns reflect three components: (i) outputs of the LLM/RAG framework, specifically the site-level probabilities and determinant intensity scores derived from semantic retrieval and sentiment analysis; (ii) evidence contained in the underlying source documents—corporate announcements, SEC filings, sustainability reports, and media coverage—from which the LLM extracts signals; and (iii) the authors’ interpretation of these outputs in the context of the broader literature, which synthesizes these signals into regional and firm-level characterizations. Figure 1 illustrates the global spatial evolution and underlying determinants of AI data-center siting, with legacy infrastructure and AI-specific siting layers presented separately in Supplementary Fig. S3.\nLegacy data centers (blue markers) are relatively dispersed, reflecting historical priorities such as latency reduction and proximity to users. In contrast, newly identified AI-specific data centers (color-coded circles) exhibit strong geographic concentration, forming dense clusters in a limited number of regions. These clusters are most pronounced in North America, Europe, and the Asia-Pacific, which together account for over 90% of projected AI compute capacity among leading firms. The dominant corridors align with regions characterized by strong energy availability, mature grid and fiber-optic infrastructure, favorable climatic conditions, and supportive policy environments29,30.\nDistinct regional configurations highlight the heterogeneity of siting determinants. The following characterizations are based on signals extracted by the LLM framework from input documents (corporate filings, sustainability disclosures, and policy sources); the synthesis of these signals into regional interpretations represents the authors’ reading of those extracted patterns. In North America, clusters are concentrated in regions such as Virginia, Texas, Ohio, and North Carolina, which coincide with established hyperscale infrastructure, large electricity markets, and documented policy support for data-center development11,31. These regions have historically attracted cloud infrastructure investment, although underlying conditions such as energy costs, regulatory incentives, and climatic factors vary substantially across locations. In Europe, data-center development is concentrated in the Netherlands, the United Kingdom, and Italy. Renewable-energy availability, regulatory frameworks, and digital infrastructure jointly shape these patterns although their relative importance varies across countries. Nordic regions, in particular, provide favorable conditions due to hydro and wind resources and naturally cool climates30. In the Asia-Pacific region, clusters in Singapore, Taipei, Malaysia, and Japan align with strong digital demand and policy-supported infrastructure expansion. Similarly, emerging sites in the Middle East are associated with state-led investment strategies and land availability.\nFirm-level differentiation further reinforces these spatial dynamics (Fig. 1a, b). To interpret the firm-determinant intensity matrix (Fig. 1b), we classify siting drivers into six dimensions: corporate integration, energy access, policy environment, market demand, infrastructure maturity, and network connectivity. Higher scores indicate that a determinant category appears more frequently and with stronger positive sentiment in retrieved documents associated with a given firm. Among globally scaling firms, Amazon exhibits the broadest geographic diversification (Fig. 1a), expanding across regions with strong market demand and mature infrastructure. Microsoft and Oracle primarily follow policy- and incentive-driven corridors, while Google and Meta anchor their siting strategies in regions with strong energy availability and high network connectivity to reduce latency and operational constraints. By contrast, domestically oriented firms such as Apple adopt a more vertically integrated strategy, concentrating deployments within US regions-particularly those with favorable energy conditions–to optimize operational efficiency and alignment with their broader manufacturing and service ecosystems.\nOverall, these results reveal a structural transition from historically dispersed siting toward geographically concentrated AI infrastructure, with clustering patterns shaped jointly by regional characteristics and firm-level strategies. Energy availability, policy alignment, and network connectivity emerge as dominant determinants of siting decisions, driving the formation of high-density infrastructure corridors.\nElectricity consumption projection: rapid scaling of AI data centers elevates system-level energy demand\nTo capture uncertainty in expansion pace and AI workload adoption, we evaluate three forward scenarios for 2025–2030: a conservative scenario (15% annual growth), a neutral scenario (25%), and an optimistic scenario (35%). These growth rates span the range of recent estimates of data-center and AI-related electricity demand growth reported in the literature and industry forecasts11,31. In all scenarios, baseline data-center electricity consumption grows at 10% annually, while the share of AI-intensive workloads increases over time, reflecting the rising importance of generative-AI computation. Electricity demand is estimated by combining firm-level data-center expansion scenarios with projected AI workload intensity and baseline energy-use trends (see Methods). Figure 2 presents the projected electricity consumption of AI data centers under the three growth scenarios from 2025 to 2030. Figure 2a–f illustrates firm-level trajectories for Amazon, Microsoft, Google, Meta, Oracle, and Apple across conservative, neutral, and optimistic assumptions. The energy demand forecasts for all firms show consistent upward trends, reflecting the intensification of AI workloads and the proliferation of large-model training clusters. Aggregate electricity use among hyperscale operators is projected to rise by ~40−70% over the period, though the pace and scale vary considerably by firm. Amazon, Microsoft, and Google are forecasted to maintain the highest absolute demand, consistent with their extensive cloud-service portfolios and broad geographic reach. Meta follows comparable but more moderate forecast paths, supported by continuous improvements in model efficiency and renewable procurement strategies32,33,34,35,36,37. Apple and Oracle exhibit lower projected energy demand, aligning with their smaller operational footprints and vertically integrated architectures. Across all firms, the optimistic scenario accelerates notably after 2027, coinciding with large-scale deployment of generative-AI infrastructure and training workloads38,39.\nFigure 3 aggregates these firm-level forecasts to depict the global evolution of electricity demand associated with leading AI operators from 2018 to 2030. The solid gray line represents historical estimates derived from industrial disclosures and utility data, while the colored dashed lines correspond to scenario-based projections. Total electricity consumption by these leading firms’ data centers is expected to rise from ~118 TWh in 2024 to between 239 TWh (conservative) and 295 TWh (optimistic) by 2030, implying a compound annual growth rate (CAGR) of roughly 13−17%. This increase is equivalent to adding the electricity demand of a medium-sized national market every two years, or comparable to the annual electricity use of about 27 million US households. The projected magnitude is quantitatively consistent with the International Energy Agency’s (IEA) forecast that total global data center electricity demand will reach around 945 TWh by 2030 (~3% of global power consumption)11, compared with about 1.5% in 2024 (see Methods). These projections collectively illustrate the accelerating energy footprint of AI infrastructure within the broader global power system.\nRegional electricity demand pressure implications of AI data-center clustering\nFigure 4 characterizes the regional distribution of data-center electricity demand and the relative concentration of AI-related loads within regional electricity systems in 2030. Figure 4 displays projected electricity demand by firm and region, revealing a highly uneven spatial distribution of AI-related loads. Fewer than ten regions account for nearly two-thirds of total projected demand, indicating a strong concentration of infrastructure in a limited number of states and countries. Oregon, Virginia, Iowa, Ohio and Ireland emerge as dominant hubs, each exceeding 15 TWh of annual demand under the neutral scenario. These regions are associated with favorable conditions—such as established hyperscale clusters, access to energy resources, and supportive policy environments—that are consistent with large-scale AI deployment. Second-tier regions, including the Texas, Netherlands, and Washington, exhibit moderate but rapidly increasing demand, often driven by firm-specific expansion. Scenario ranges indicate that uncertainty in deployment pace and firm strategy could alter local electricity demand by up to 30%, underscoring the challenges of forecasting and planning for digital-infrastructure growth.\nTo characterize the relative concentration of data-center electricity demand across regions, we construct an electricity demand pressure index (EDPI), defined as the ratio of projected data-center electricity demand to regional total available electricity supply (see Methods). Figure 5a presents the cross-regional evolution of EDPI between 2025 and 2030 under the neutral scenario, revealing a pronounced and spatially uneven increase in demand concentration. A small number of regions—including Oregon, Ireland and Iowa—display markedly elevated EDPI levels in both years, indicating a persistently high concentration of data-center electricity demand relative to available supply. Virginia, Nebraska, and Ohio occupy the upper-middle range of the distribution, reflecting growing demand concentration. The absolute change in EDPI between 2025 and 2030 (ΔEDPI, expressed in percentage points) further reveals a self-reinforcing spatial pattern: regions with already elevated demand concentration tend to experience the largest absolute increases. In contrast, regions with larger electricity systems—such as Texas—exhibit comparatively small ΔEDPI values, reflecting greater available-supply adequacy to absorb incremental AI-related demand (Fig. 5a–c).\nOverall, these results reveal a pronounced spatial asymmetry in the distribution of data-center electricity demand, with a limited number of regions accounting for a disproportionate share of demand relative to their available-electricity adequacy.\nDiscussion\nThis study links AI data-center siting patterns, electricity-demand projections, and regional electricity demand concentration. By combining large language model (LLM)-based infrastructure identification with scenario-based electricity-demand projections, we assess how the spatial expansion of AI computing infrastructure may influence electricity demand and the relative burden on regional electricity systems. A key methodological contribution of this framework is the fusion of qualitative corporate intelligence—extracted by LLMs from financial disclosures, sustainability reports, and strategic announcements—with quantitative electricity demand projection. This combination enables inference of firms’ implicit strategic intent from language patterns in corporate communications, providing a more interpretable and forward-looking basis for demand forecasting than statistical extrapolation from historical consumption trends alone11,20.\nOur results indicate that AI-driven data-center expansion is emerging as an increasingly influential component of global electricity systems7,40. Across firms and regions, the analysis reveals a consistent sequence linking spatial siting patterns, rapidly increasing electricity demand, and rising regional power stress. This pattern suggests that the digital infrastructure supporting AI development is becoming closely intertwined with electricity-system dynamics rather than remaining an exogenous source of demand. As AI-related computing capacity expands and concentrates geographically, it may influence regional load growth, infrastructure investment priorities, and decarbonization pathways41.\nThe clustering of AI data centers in a limited number of favorable regions—primarily in North America, Western Europe, and parts of the Asia-Pacific—creates pronounced spatial asymmetries in load distribution. Regions such as Oregon, Ireland and Iowa—which exhibit elevated EDPI values—face a high relative concentration of data-center electricity demand within their electricity systems, alongside rapid AI infrastructure expansion. In contrast, regions with larger electricity systems, such as Texas, show lower demand concentration relative to total available electricity supply. These patterns highlight how regional electricity-system size, cross-boundary trade flows, and the spatial footprint of AI infrastructure jointly determine the relative exposure of local electricity systems to AI-driven demand growth. The projected growth in electricity demand from AI data centers therefore presents both challenges and opportunities for electricity systems. On one hand, concentrated digital loads could intensify transmission congestion, complicate renewable integration, and increase electricity price volatility if system expansion does not keep pace with demand. On the other hand, large and stable computational loads may provide opportunities for greater integration with electricity systems Data centers can function as anchor customers for long-term renewable power purchase agreements (PPAs), potential participants in flexible demand programs, and focal points for co-located energy storage or hybrid renewable generation42,43. The extent to which these opportunities materialize will depend on the interaction between electricity market design, infrastructure investment, and regulatory frameworks.\nAt the policy level, these findings highlight the potential value of anticipatory frameworks that explicitly incorporate AI-related demand into national energy modeling and long-term planning. Conventional forecasts that treat digital infrastructure as exogenous commercial load risk underestimating the spatial concentration of new demand and its implications for annual energy balancing. Integrating firm-level digital-load projections into planning processes could enable regulators and utilities to better anticipate load growth in high-concentration regions. In emerging markets, policy design may need to balance the benefits of digital-infrastructure development against the risk of exacerbating existing demand-supply imbalances.\nImportantly, the firm-level granularity of this analysis offers a policy-relevant tool that complements aggregate demand forecasts. Because AI-related electricity demand is highly concentrated among a small number of identifiable corporate actors, regulators and utilities can use firm-level demand trajectories to engage with specific operators rather than with a generic “technology sector.” This granularity may support more targeted policy interventions—such as renewable procurement requirements, locational guidance, or mandatory load disclosure—calibrated to the expansion strategies and geographical footprints of individual firms.\nSeveral limitations of the current framework warrant explicit acknowledgment. First, the LLM-based inference relies on corporate sustainability reports, press releases, and annual filings, which may contain optimistic bias regarding efficiency trajectories, renewable integration commitments, or future investment plans. To the extent that such documents overstate actual or planned improvements, the sentiment-based siting probabilities may be biased toward high-profile or low-carbon locations, potentially understating deployment in less prominently disclosed regions. Moreover, regions where AI infrastructure projects are less frequently disclosed (such as Africa or Australia), less systematically reported, or still at an early stage of development may be underrepresented within the inferred deployment database. Second, the efficiency-gain assumption of 1−3% annually—calibrated to historical hyperscale patterns—reflects a central scenario. Further improvements in chip efficiency, including more energy-efficient accelerators or advances in cooling architectures, could reduce energy intensity below projected levels. Third, the EDPI is constructed from aggregate annual energy quantities—available electricity supply (generation plus positive net imports) and projected data-center demand in the numerator—and does not capture intra-regional transmission network constraints or interconnection queue delays. In systems such as the PJM interconnection (a regional transmission organization coordinating wholesale electricity markets and grid operations across parts of the eastern United States, including Virginia), grid integration bottlenecks and interconnection queues can be binding constraints on the ability to absorb new large loads, even when aggregate available supply appears sufficient. The EDPI may therefore underestimate localized electricity-system pressures in transmission-constrained areas, and future extensions of this framework should incorporate network-level indicators. Fourth, the firm-level analysis covers six technology companies that together account for ~70−75% of global hyperscale and cloud-linked data-center electricity demand. The remaining share comprises a heterogeneous set of operators—including smaller cloud-service providers, colocation facility operators, telecommunications companies with data-center divisions, and regional or national operators—that are excluded primarily because their infrastructure footprints and electricity consumption are less consistently disclosed in public filings and their siting patterns are not readily identifiable by the LLM-based inference approach applied here. Excluding this segment may lead to an underestimate of total AI-related electricity demand. Fifth, available electricity projections are based on CAGR estimates from 2019–2024 with a small sample and pandemic impact, so they may not fully capture future changes in electricity systems, including renewable deployment, grid expansion, and infrastructure investments. In particular, in regions with relatively small electricity systems, such developments could substantially increase future electricity availability beyond historical trends. Likewise, future national and regional energy policies, market reforms, and unforeseen extreme events may significantly reshape electricity-supply trajectories.\nMore broadly, the co-evolution of digitalization and electrification marks a conceptual turning point in energy-system governance13. AI infrastructure is transforming electricity demand from a passive outcome of economic activity into an active determinant of system configuration. Its social and economic value increasingly depends on the sustainability of its power source and its integration within low-carbon grids. Future research should quantify the feedback loops between AI-compute trajectories, carbon intensity, and investment flows across the energy-digital interface. Several specific directions are particularly promising. First, the co-optimization of AI data-center siting with other emerging high-load technologies—such as hydrogen electrolyzers—could help minimize transmission congestion and renewable curtailment; strategic spatial allocation of such loads has been shown to reduce boundary reinforcement costs44. Second, the interaction between AI data-center loads and the electrification of transport warrants coordinated analysis, as both introduce large and potentially concurrent demands on distribution and transmission infrastructure; holistic planning frameworks that integrate gigawatt-scale AI campuses alongside electric-vehicle charging have been identified as an emerging priority45. Third, the integration of data centers into urban energy systems through waste heat recovery for building heating offers a dual benefit of improved cooling efficiency and urban decarbonization; coordinated control strategies for building thermal flexibility informed by data-center waste heat could reduce local distribution network congestion46. Fourth, the application of advanced risk assessment and stochastic optimization techniques—analogous to those used in commodity supply-chain management47—could better account for uncertainties in energy pricing, grid availability, and AI compute demand, supporting more robust infrastructure planning under deep uncertainty. Understanding and governing the coupling between computational growth and electricity systems will be central to achieving a resilient, low-carbon digital economy.\nMethods\nThis study develops a hybrid retrieval-augmented generation (RAG) + large language model (LLM) forecasting framework to predict the geographic and operational evolution of hyperscale data centers operated by major technology companies. The framework integrates structured document retrieval, language sentiment analysis, and LLM-based contextual reasoning to infer (i) potential expansion locations, (ii) baseline technical parameters, and (iii) multi-year energy trajectories.\nThe workflow (Fig. 6) consists of six main stages: (1) construction of a RAG knowledge base from multi-year corporate and regional sources, (2) sentiment-aware identification of likely future data center locations, (3) retrieval of location-specific contextual information, (4) extraction of baseline operational parameters for the current year, (5) multi-year forecasting of key technical and environmental metrics using an LLM-guided model and physically interpretable energy equations, and (6) construction of the Electricity Demand Pressure Index (EDPI) that compares projected data center demand against the available regional electricity supply capacity.\nKnowledge-base construction\nA multi-source text corpus was constructed for each firm, including press releases, annual earnings reports, environmental sustainability disclosures, infrastructure investment announcements, government policy documents, and major news articles from 2015 to 2025. All texts were preprocessed and embedded using the Hugging Face all-MiniLM-L6-v2 model, then indexed in a FAISS (Facebook AI similarity search) vector database for semantic retrieval. Each entry retained metadata such as publication date, geographic reference, and document type, enabling targeted retrieval (e.g., “AWS Taiwan 2025 expansion energy infrastructure”).\nLocation Identification via RAG + LLM\nTo infer likely future expansion sites between 2025 and 2030, the framework explicitly prompts the LLM to predict which geographic regions a firm is most likely to expand into. This prediction is based on contextual evidence retrieved from the RAG knowledge base, which includes firm sustainability reports, capital expenditure statements, and government or media coverage.\nThe RAG system is first queried using firm-specific and temporal prompts, such as {firm} 2025–2030 data center investment or construction plans. The retrieved documents contain both qualitative statements (e.g., “strong regional demand growth,” “strategic infrastructure partnership”) and quantitative indicators (e.g., “$2B allocated to APAC expansion”).\nEach document is analyzed by the LLM not only for content relevance but also for its sentiment and linguistic tone. Positive sentiment toward investment, expansion, or infrastructure development—particularly in earnings calls or annual reports—acts as an implicit prior that increases the model’s belief in a region’s likelihood of receiving new data center capacity. Conversely, negative sentiment (e.g., mentions of cost control, divestment, or policy uncertainty) decreases the associated probability weight.\nThe LLM then synthesizes these cues into a structured probability distribution over candidate locations, producing outputs of the form: P(expansion at site i ∣evidence) = fLLM(retrieved text embeddings, sentiment scores).\nThis probabilistic representation is used in later stages to weight forecasts of capacity growth and energy consumption at each site.\nBaseline parameter extraction\nFor each confirmed or high-probability site, the system used a RAG-conditioned LLM prompt to extract current-year baseline parameters describing the data center’s IT and facility operations. The baseline operational parameters used in the forecasting framework are summarized in Table 1, including accelerator counts, average power draw, utilization factors, operating hours, and power usage effectiveness (PUE).\nThe model was instructed to reason explicitly about local grid mix, cooling efficiency, and infrastructure maturity, and to avoid copying assumptions between regions. The extracted parameters included: {Ntrain, Ninference, Pavg,train, Pavg,inference, utrain, uinference, Htrain, Hinference, PUE, g}\nAll outputs were formatted as structured JSON (JavaScript Object Notation) objects, accompanied by short notes summarizing assumptions and references. Each variable corresponds to a measurable operational attribute of the local compute fleet.\nForecasting future parameters\nA second RAG + LLM chain projected the time evolution of these parameters over a 5-year horizon. The prompt incorporated both historical operational trends and retrieved regional information, constraining LLM reasoning with empirically observed hyperscale patterns: 10−20% annual accelerator capacity growth, 1−3% annual efficiency gains, and gradual improvements in utilization. These parameters characterize the supply-side evolution of infrastructure, specifically capturing the growth in installed accelerator capacity and incremental improvements in energy efficiency at the facility level. The model generated structured forecasts for each year t ∈ [2025, 2030], each annotated with explanatory notes and relevant document sources.\nEnergy computation\nForecasted operational parameters were post-processed in Python to derive the physical energy demand of each site. The IT load energy (EIT) represents the direct electrical consumption of computing equipment, while the estimated AI data center energy demand (EDC) includes both IT and non-IT overheads such as cooling and power distribution losses.\nHere, N, Pavg, u, and H denote the number of accelerators, average power draw (kW), utilization factor, and annual operating hours, respectively. EIT is thus measured in MWh, and EDC scales this load by the facility’s power usage effectiveness (PUE), capturing site-specific cooling and infrastructure efficiency.\nImplementation\nThe full pipeline was implemented in Python using the LangChain framework for composable RAG pipelines, FAISS for vector retrieval, and the OpenAI GPT-4o-mini model as the reasoning component (temperature = 0.3). All modules were orchestrated via reproducible runnables with strict JSON parsing for downstream numerical analysis.\nSentiment-aware expansion probability modeling\nThe sentiment analysis stage produced, for each candidate region i, an expansion likelihood score Pi defined as:\nwhere Si is the normalized sentiment score for region i (ranging from −1 to +1), and Ri is the retrieval relevance score (cosine similarity in the embedding space). These probabilities informed the sampling or prioritization of forecast targets in subsequent steps. Regions frequently mentioned in corporate reports and described with positive financial tone thus received elevated probabilities of selection.\nThis coupling of semantic retrieval and sentiment tone weighting enables the model to capture implicit strategic intent in corporate communications—an important predictor of near-term data center expansion behavior.\nBy fusing semantic retrieval, sentiment weighting, and physical energy modeling, this hybrid framework bridges qualitative corporate signals and quantitative operational forecasting. The approach captures how firms’ language choices in financial and environmental disclosures correlate with real-world infrastructure trajectories, enabling a more interpretable and evidence-grounded forecast of global data center development.\nScenario design and aggregation\nTo capture uncertainty in both firm-level expansion intensity and technological efficiency, we constructed three forward scenarios s-conservative, neutral, and optimistic-for 2025–2030. Each scenario jointly controls (i) the rate of new AI data-center additions and (ii) the efficiency gains of all operating sites. The annual growth rate of new AI load is set to gnew,s ∈ {0.15, 0.25, 0.35} for scenario s among the three scenarios, while existing stock consumption for each firm grows at gstock = 0.10. Specifically, the three forward scenarios for AI-driven data-center expansion (15, 25, and 35% annual growth) are designed to span conservative, central, and high-growth trajectories of AI workload expansion. These values are informed by recent industry and research evidence on the growth of AI compute demand and infrastructure. AI training workloads are estimated to grow at around 22% annually, while inference workloads may expand at rates approaching 35% over the next five years48. Other estimates indicate that overall AI-related infrastructure demand may grow in the range of 20−25% annually in baseline scenarios49. Based on this range of evidence, the 15% scenario reflects a lower-bound trajectory consistent with moderate expansion in AI deployment, the 25% scenario captures the central tendency of observed infrastructure growth rates, and the 35% scenario represents an upper-tail case aligned with rapid scaling of inference workloads and AI-ready capacity. These scenarios are designed to capture the empirically observed range of AI compute and infrastructure growth, while also reflecting uncertainty in future AI adoption intensity and technological scaling.\nFirm-specific AI shares pAI(f, t) increase gradually over time to reflect the rising share of AI workloads in total compute demand. Firm-specific AI shares pAI(f, t) were defined to represent the evolving proportion of AI-related compute workloads within each operator’s total data-center activity. For each firm f, pAI(f, t) increases over time according to a discrete schedule calibrated to reflect the firm’s relative intensity of AI adoption between 2025 and 2030:\nwhere p1(f), p2(f), and p3(f) correspond to the firm’s baseline (2025), mid-phase (2026–2027), and mature-phase (2028–2030) AI workload shares, respectively. Formally, this can be expressed as:\nThe parameter set {p1(f), p2(f), p3(f)} is firm-specific, capturing heterogeneity in AI adoption trajectories across major hyperscale operators:\nThis stepwise formulation approximates each firm’s increasing allocation of computational resources to AI model training and inference. The heterogeneity in pAI(f, t) reflects differences in investment magnitude, infrastructure build-out, and strategic focus among major operators. Firms with large-scale and early commitments to generative AI—such as Amazon, Google, Meta, and Microsoft—exhibit steeper trajectories, consistent with multibillion-dollar data-center expansions, dedicated AI accelerator deployment (e.g., Trainium, H100, and TPUv5), and vertically integrated AI ecosystems. By contrast, vertically integrated or enterprise-focused operators such as Apple and Oracle show slower growth, consistent with smaller-scale AI infrastructure investment, reliance on consumer-device optimization rather than large-model training, and a later pivot toward cloud-based AI services. Accordingly, the evolution of pAI(f, t) captures each firm’s position along the spectrum of AI-driven digital-infrastructure transformation, serving as a proxy for the relative share of compute capacity devoted to AI workloads within their overall fleet.\nFor each firm f, the existing stock electricity consumption evolves as\nNew-site electricity demand follows compounding growth of AI workloads and the firm-specific AI share schedule:\nThe total firm-level electricity demand is then\nTo represent uncertainty arising from multiple independent siting pathways, we simulate a set of Pt,s plausible firm-region configurations for each year t and scenario s. Global electricity demand in year t and scenario s is the ensemble mean:\nand the range\ndefines the shaded uncertainty bands in the scenario plots, where \\({E}_{t,s,p}^{{{\\rm{glob}}}}\\) denotes the global total electricity demand in year t under scenario s for projection path p ∈ {1, …, Pt,s}. The 2024 observed total serves as the common anchor point for all trajectories.\nRegional allocation of energy\nFirm-level electricity demand is spatially distributed according to two complementary weights: (i) AI siting probabilities derived from the LLM siting model, and (ii) historical stock weights from the existing data-center inventory.\nFor AI-related new sites, firm-year location weights are proportional to modeled AI energy at each location ℓ:\nHere \\({E}_{f,t,\\ell }^{AI,\\,{\\mbox{loc}}\\,}\\) denotes the modeled AI-related electricity demand assigned to firm f at location ℓ in year t, as determined from the LLM-based siting model and corresponding regional AI workload intensity.\nFor legacy and non-AI sites, we use historical weights based on observed facility counts. For legacy and non-AI sites, we use historical weights based on observed facility counts:\nLegacy (historical) data-center counts refer to the number of distinct operating campuses under each firm’s control as of end-2024, including both owned and long-term leased facilities. These are derived from the observed 2015–2024 dataset and serve as proxies for baseline spatial capacity prior to the AI expansion. Since 2024, the global data-center landscape has entered a rapid expansion phase driven by generative-AI workloads, with AI infrastructure investment reaching a record $57 billion globally50. This “AI data-center boom” marks a structural inflection in digital-infrastructure growth.\nLet r(ℓ) map locations to regions. Regional electricity use for firm f and scenario s is:\nSumming over all firms yields regional totals:\nThis approach preserves firm-level energy consistency while ensuring that AI-driven growth concentrates in high-probability siting regions identified by the LLM model, and that legacy loads follow empirically observed footprints.\nCross-validation\nAggregate electricity consumption by the six leading hyperscale operators is projected to rise from ~118 TWh in 2024 to 239–295 TWh by 2030. This projection is broadly consistent with the International Energy Agency’s (IEA) estimate that global data-center electricity demand will reach roughly 945 TWh by 2030 (IEA, 2025). To benchmark the modeled values, we derive an implied 2030 consumption level for the six leading firms based on the IEA’s global forecast. Assuming (i) that hyperscale data centers account for 70% of global data-center electricity use (Synergy Research Group, 2024), and (ii) that the top six operators collectively represent 40% of hyperscale activity, the corresponding implied electricity consumption is:\nThis cross-validation demonstrates that the projected range of 239–295 TWh lies well within the values implied by independent international forecasts. The close correspondence reinforces confidence in the representativeness of the modeled global totals and supports the credibility of the upper-bound scenario in light of established external benchmarks. Overall, the modeled range remains broadly aligned with independent global projections and reflects plausible hyperscale market dynamics.\nElectricity demand pressure index (EDPI)\nTo assess the relative demand burden associated with data-center expansion, we construct an electricity demand pressure index (EDPI), defined as the ratio of annual data-center electricity demand to the regional available electricity supply—comprising in-region generation plus positive net electricity imports. For each region r and year t:\nwhere \\({E}_{r,t}^{{\\mathrm{Avail}}} = {E}_{r,t}^{{\\mathrm{Gen}}} + {\\max} \\left(0, {{E}^{{\\mathrm{NetImport}}}_{r,t}}\\right), {{E}^{{\\mathrm{DC}}}_{r,t}}\\) denotes total annual data-center electricity demand (in TWh), with \\({E}_{r,t}^{Gen}\\) denoting regional electricity generation (TWh), \\({E}_{r,t}^{{\\mathrm{Gen}}}\\) denoting net electricity imports, and \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) denoting total available electricity supply (TWh). Positive net imports are included because imported electricity contributes directly to regional electricity supply. For net-exporting regions, historical electricity exports are treated as adjustable rather than permanently fixed. Accordingly, negative net imports are set to zero in the denominator, reflecting the assumption that part of the electricity historically exported could instead be retained to serve additional local electricity demand associated with AI data-center expansion. This approximation avoids overstating electricity demand pressure in regions that are persistent net electricity exporters while providing a consistent basis for cross-regional comparison.\nFor US states, \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) is derived from EIA state electricity profiles as total retail electricity sales, which by the energy balance identity equals in-state net generation plus net interstate electricity receipts. For countries, \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) is the sum of total electricity generation and positive net electricity imports, both drawn from the Ember Yearly Full Release dataset (2019–2024). The EDPI is interpreted as a demand-share indicator capturing the relative contribution of data-center loads to regional electricity supply.\nRegional available electricity supply is projected using historical trends of available electricity as a baseline approximation. Specifically, available electricity is extrapolated from the period 2019–2024 using a region-specific compound annual growth rate (CAGR), consistent with the five-year window used in the CAGR formula below.\nRegional available electricity baselines\n(\\({E}_{r,2024}^{{\\mathrm{Avail}}}\\)) were obtained from EIA State Electricity Profiles (US states) and the Ember Yearly Full Release dataset (countries), covering 2019–2024, expressed in terawatt-hours (TWh). Available electricity data were extrapolated to 2030 using a compound annual growth rate (CAGR):\nwith \\({{\\mbox{CAGR}}}_{avail,r}={\\left(\\frac{{E}_{r,2024}^{Avail}}{{E}_{r,2019}^{Avail}}\\right)}^{1/5}-1\\).\nRegional electricity available supply is projected using region-specific compound annual growth rates (CAGR) estimated from historical observations during 2019–2024. This approach assumes that recent historical trends provide a reasonable baseline approximation of future electricity-system evolution. The resulting projections should therefore be interpreted as baseline-trend estimates rather than policy-conditioned forecasts.\nData availability\nThe publicly available electricity data used in this study were obtained from the Ember Yearly Electricity Data Explorer (https://ember-energy.org/data/yearly-electricity-data/) and the US Energy Information Administration State Electricity Profiles (https://www.eia.gov/electricity/state/). The International Energy Agency’s Energy and AI report (https://www.iea.org/reports/energy-and-ai) was used for external benchmarking. Additional public corporate and contextual information was obtained from SEC EDGAR filings (https://www.sec.gov/edgar/search/), corporate annual and sustainability reports, press releases, government and grid-operator publications, and media reports. Historical facility-level data for Amazon, Microsoft, Google, Meta, Oracle, and Apple were obtained from S&P Capital IQ (https://www.capitaliq.com/). The processed data and model outputs underlying the data-center siting estimates, electricity-demand projections, electricity demand pressure index calculations, figures, and tables are publicly available at https://github.com/dchen219/ai-data-center-electricity.\nCode availability\nThe custom code used for document processing, LLM-based data-center siting inference, electricity-demand projection, electricity demand pressure index calculation, and figure generation is publicly available at https://github.com/dchen219/ai-data-center-electricity. Analyses were conducted in Python using LangChain, FAISS, the Hugging Face all-MiniLM-L6-v2 embedding model, and OpenAI’s GPT-4o-mini model.\nReferences\n- Stylianou, N. et al. Inside the relentless race for AI capacity. Financial Times. https://ig.ft.com/ai-data-centres/ (2025). \n- Storey, V. C., Yue, W. T., Zhao, J. L. & L., R. Generative artificial intelligence: evolving technology, growing societal impact, and opportunities for information systems research. Inform. Syst. Front. https://doi.org/10.1007/s10796-025-10581-7 (2025). \n- Cho, J. & Moon, J. H. Numerical coupling of energy efficiency and thermal performance for cold plate cooling optimization in high-density compute AI data centers. Energy Build. 348, 116441 (2025). \n- Pilz, K. F., Mahmood, Y. & Heim, L. AI’s Power Requirements under Exponential Growth: Extrapolating ai Data Center Power Demand and Assessing its Potential Impact on U.S. Competitiveness. Tech. Rep. https://www.rand.org/pubs/research_reports/RRA3572-1.html (RAND Corporation, 2025). \n- Oxford Institute for Energy Studies. Artificial Intelligence and its Implications for Electricity Systems. Tech. Rep. 145. https://www.oxfordenergy.org/wpcms/wp-content/uploads/2025/05/OEF-145.pdf (Oxford Institute for Energy Studies, 2025)) \n- Chen, X., Wang, X., Colacelli, A., Lee, M. & Xie, L. Electricity demand and grid impacts of AI data centers: Challenges and prospects. Nexus 3, 100162 https://doi.org/10.1016/j.ynexs.2026.100162 (2026) \n- Stojkovic, J., Zhang, C., Goiri, Íñigo., Torrellas, J. & Choukse, E. DynamoLLM: designing LLM inference clusters for performance and energy efficiency. In 2025 IEEE International Symposium on High Performance Computer Architecture (HPCA) 1348–1362 (2025). \n- Crawford, D., Schallehn, M., Renno, P., Hanbury, P. & Cannarsi, A. How can We meet AI’s Insatiable Demand for Compute Power? https://www.bain.com/insights/how-can-we-meet-ais-insatiable-demand-for-compute-power-technology-report-2025/ (Bain & Company, 2025). \n- Ngata, W. M. et al. The cloud next door: Investigating the environmental and socioeconomic strain of datacenters on local communities. In Proc. 2025 ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies, COMPASS ’25 769–774 (Association for Computing Machinery, 2025). \n- Li, P., Yang, J., Islam, M. A. & Ren, S. Making AI less ‘Thirsty’. Commun. ACM 68, 54–61 (2025). \n- International Energy Agency. Energy and AI. https://www.iea.org/reports/energy-and-ai (2025). \n- Bogmans, C. et al. Power hungry: how AI will drive energy demand. IMF Working Paper 2025/081, International Monetary Fund (2025). \n- Wu, C.-J. et al. Sustainable AI: environmental implications, challenges, and opportunities. In Proc. Machine Learning and Systems 795–813. https://proceedings.mlsys.org/paper_files/paper/2022/hash/462211f67c7d858f663355eff93b745e-Abstract.html (2022). \n- Chen, Z.-M. et al. AI carbon footprint in China sets to double post-2030 carbon peaking. Energy Econ. 150, 108880 (2025). \n- Belkhir, L. & Elmeligi, A. Assessing ICT global emissions footprint: trends to 2040 and recommendations. J. Clean. Prod. 177, 448–463 (2018). \n- Masanet, E., Shehabi, A., Lei, N., Smith, S. & Koomey, J. Recalibrating global data center energy-use estimates. Science 367, 984–986 (2020). \n- Malmodin, J., Lövehagen, N., Bergmark, P. & Lundén, D. ICT sector electricity consumption and greenhouse gas emissions–2020 outcome. Telecommun. Policy 48, 102701 (2024). \n- Debus, C., Piraud, M., Streit, A., Theis, F. & Götz, M. Reporting electricity consumption is essential for sustainable AI. Nat. Mach. Intell. 5, 1176–1178 (2023). \n- Cao, Z. et al. Data center sustainability: revisits and outlooks. IEEE Trans. Sustain. Comput. 9, 236–248 (2024). \n- Xiao, T., Nerini, F. F., Matthews, H. D., Tavoni, M. & You, F. Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA. Nat. Sustain. 8, 1541–1553 (2025). \n- Zhang, Y., Tang, H., Li, H. & Wang, S. Integration and interaction of next-generation AI-focused data centers with smart grids and district energy systems: the state-of-the-art, opportunities and challenges. Renew. Sustain. Energy Rev. 224, 116097 (2025). \n- Han, J. et al. Data-driven distributionally robust optimization of low-carbon data center energy systems considering multi-task response and renewable energy uncertainty. J. Build. Eng. 102, 111937 (2025). \n- Nøland, J. K., Hjelmeland, M. & Korpås, M. Will energy-hungry AI create a baseload power demand boom? IEEE Access 12, 110353–110360 (2024). \n- Luccioni, A. S., Viguier, S. & Ligozat, A.-L. Estimating the carbon footprint of BLOOM, a 176B parameter language model. J. Mach. Learn. Res. 24, 1–15 (2023). \n- Tomlinson, B., Black, R. W., Patterson, D. J. & Torrance, A. W. The carbon emissions of writing and illustrating are lower for AI than for humans. Sci. Rep. 14, 3732 (2024). \n- Ahmed, K. M. U., Bollen, M. H. J. & Alvarez, M. A review of data centers energy consumption and reliability modeling. IEEE Access 9, 152536–152563 (2021). \n- Liu, Y. et al. Energy consumption and emission mitigation prediction based on data center traffic and PUE for global data centers. Glob. Energy Interconnect. 3, 272–282 https://doi.org/10.1016/j.gloei.2020.07.008 (2020). \n- Siddik, M. A. B., Shehabi, A. & Marston, L. The environmental footprint of data centers in the United States. Environ. Res. Lett. https://doi.org/10.1088/1748-9326/abfba1(2021). \n- Cao, Z., Zhou, X., Hu, H., Wang, Z. & Wen, Y. Toward a systematic survey for carbon neutral data centers. IEEE Commun. Surv. Tutor. 24, 895–936 https://doi.org/10.1109/COMST.2022.3161275 (2022). \n- Hyvönen, J. et al. Potential of solar photovoltaics and waste heat utilization in cold climate data centers: case study: Finland and northern Japan. Renew. Sustain. Energy Rev. https://doi.org/10.1016/j.rser.2024.114619 (2024). \n- Shehabi, A. et al. 2024 United States data center energy usage report. Tech. Rep. LBNL-2001637. https://doi.org/10.71468/P1WC7Q (Lawrence Berkeley National Laboratory, 2024). \n- Synergy Research Group. Hyperscale data center count hits 1136; average size increases; US accounts for 54% of total capacity. https://www.srgresearch.com/articles/hyperscale-data-center-count-hits-1136-average-size-increases-us-accounts-for-54-of-total-capacity (2025). \n- Donnellan, D. et al. Uptime Institute global data center survey results 2024. Tech. Rep. https://uptimeinstitute.com/resources/research-and-reports/uptime-institute-global-data-center-survey-results-2024 (Uptime Institute, 2024). \n- Khan, O. Scale your AI transformation with a powerful, secure, and adaptive cloud infrastructure. Microsoft Azure Blog. https://azure.microsoft.com/en-us/blog/scale-your-ai-transformation-with-a-powerful-secure-and-adaptive-cloud-infrastructure/ (2024). \n- Google Sustainability. 2023 Environmental report. https://sustainability.google/reports/google-2023-environmental-report/ (2023) \n- Meta Platforms, Inc. Annual report on Form 10-K for the fiscal year ended december 31, 2024. Tech. Rep. SEC Accession No. 0001326801-25-000017, U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/1326801/000132680125000017/meta-20241231.htm (2025). \n- Sivasubramanian, S. Welcome to a new era of building in the cloud with generative AI on AWS. AWS Machine Learning Blog. https://aws.amazon.com/blogs/machine-learning/welcome-to-a-new-era-of-building-in-the-cloud-with-generative-ai-on-aws/ (2023). \n- Apple Inc. Annual report on Form 10-K for the fiscal year ended september 28, 2024. U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/320193/000032019324000123/aapl-20240928.htm (2024). \n- Oracle Corporation. Oracle offers first zettascale cloud computing cluster. Oracle News press release. https://www.oracle.com/news/announcement/ocw24-oracle-offers-first-zettascale-cloud-computing-cluster-2024-09-11/ (2024). \n- Lambert, S., Caron, E., Lefèvre, L. & Grivel, R. Consolidation of virtual machines to reduce energy consumption of data centers by using ballooning, sharing and swapping mechanisms. Future Gener. Comput. Syst. 174, 107968 https://doi.org/10.1016/j.future.2025.107968 (2026). \n- Bolón-Canedo, V., Morán-Fernández, L., Cancela, B. & Alonso-Betanzos, A. A review of green artificial intelligence: towards a more sustainable future. Neurocomputing 599, 128096 https://doi.org/10.1016/j.neucom.2024.128096 (2024). \n- de Roucy-Rochegonde, L. & Buffard, A. AI, Data Centers and Energy Demand: Reassessing and Exploring the Trends. Tech. Rep. https://www.ifri.org/en/papers/ai-data-centers-and-energy-demand-reassessing-and-exploring-trends-0 (Institut français des relations internationales (Ifri), 2025). \n- Goldman Sachs Research. GS SUSTAIN: Generational Growth: AI, Data Centers and the Coming US Power Demand Surge. Tech. Rep. https://www.goldmansachs.com/pdfs/insights/pages/generational-growth-ai-data-centers-and-the-coming-us-power-surge/report.pdf (Goldman Sachs, 2024). \n- Giannelos, S., Konstantelos, I., Pudjianto, D. & Strbac, G. The impact of electrolyser allocation on Great Britain’s electricity transmission system in 2050. Int. J. Hydrog. Energy 202, 153097 https://doi.org/10.1016/j.ijhydene.2025.153097 (2026). \n- Amann, G. et al. E-mobility Deployment and Impact on Grids: Impact of EV and Charging Infrastructure on European T&D Grids: Innovation needs. Technical Report MJ-09-22-246-EN-N (Publications Office of the European Union, 2022). \n- Dong, Z., Zhang, X., Zhang, L., Giannelos, S. & Strbac, G. Flexibility enhancement of urban energy systems through coordinated space heating aggregation of numerous buildings. Appl. Energy 374, 123971 (2024). \n- Giannelos, S., Konstantelos, I. & Strbac, G. Optimal supply chain design using machine learning, risk assessment and optimisation applied to coal distribution. EURO J. Decis. Process. 13, 100062 (2025). \n- Arora, C., Sorel, M. & Sachdeva, P. The next big shifts in AI workloads and hyperscaler strategies. McKinsey & Company. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-next-big-shifts-in-ai-workloads-and-hyperscaler-strategies (2025) \n- Barth, A., Arora, C., Shenai, G., Noffsinger, J. & Sachdeva, P. The data center balance: how US states can navigate the opportunities and challenges. McKinsey & Company. https://www.mckinsey.com/industries/public-sector/our-insights/the-data-center-balance-how-us-states-can-navigate-the-opportunities-and-challenges (2025). \n- Saavedra, R. & Seaward, S. 2025 Data center marketplace: balancing unprecedented opportunity with strategic risk. U.S. Research Report, Colliers. https://www.colliers.com/en/research/nrep-usdc-data-center-marketplace-2025 (2025). \nAcknowledgements\nNot applicable.\nFunding\nAuthors declare no relevant funding.\nAuthor information\nAuthors and Affiliations\nContributions\nD.C. and Y.C. conceived the study. D.C. designed the analytical framework, collected and processed the data, implemented the LLM/RAG-based analysis, conducted the electricity-demand projections and EDPI analysis, prepared the figures and tables, and drafted the manuscript. Z.Z. contributed to data collection, data validation, and interpretation of results. J.Q. and L.C. contributed to methodological development, results visualization, and validation of the analysis. A.K. and Y.C. supervised the study, provided conceptual guidance, and contributed to the interpretation and revision of the manuscript. All authors reviewed, edited, and approved the final manuscript.\nCorresponding author\nEthics declarations\nCompeting interests\nThe authors declare no competing interests.\nPeer review\nPeer review information\nCommunications Sustainability thanks Rahman Khorramfar and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Nandita Basu. A peer review file is available.\nAdditional information\nPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.\nRights and permissions\nOpen Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.\nAbout this article\nCite this article\nChen, D., Zhou, Z., Cai, Y. et al. Artificial intelligence data centers could reach one percent of global electricity demand by 2030. Commun. Sustain. 1, 147 (2026). https://doi.org/10.1038/s44458-026-00152-5\n- Received: \n- Accepted: \n- Published: \n- Version of record: \n- DOI: https://doi.org/10.1038/s44458-026-00152-5","image_url":"https://media.springernature.com/m685/springer-static/image/art%3A10.1038%2Fs44458-026-00152-5/MediaObjects/44458_2026_152_Fig1_HTML.png","lang":"en","published_at":"2026-09-21T11:04:25+00:00","fetched_at":"2026-09-23T03:15:03+00:00","status":"read","starred":0,"extract_state":"ok","summary_auto":"Abstract\nThe rapid growth of generative artificial intelligence is increasing global electricity demand and placing new pressure on power systems. Here we show that electricity use by data centers built for artificial intelligence could rise from about 118 terawatt-hours in 2024 to between 239 and 295 terawatt-hours by 2030, or about 1% of global electricity demand.","cluster_id":null,"extract_retries":0,"extract_error":null,"contract_version":"news_item.v1","format_contract_version":"news_item_formats.v1","dedup_url":"https://www.nature.com/articles/s44458-026-00152-5","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 62509 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":62509,"summary_length":368,"usable_text_length":62509,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":62509,"summary_length":368}},"news_item":{"id":88444,"canonical_url":"https://www.nature.com/articles/s44458-026-00152-5","source_url":"https://www.nature.com/articles/s44458-026-00152-5","title":"Artificial intelligence data centers could reach one percent of global electricity demand by 2030 - Nature","source_name":"Nature","author":null,"published_at":"2026-09-21T11:04:25+00:00","locale":"en","topic":"ai","tags":[],"rss_summary":"<a href=\"https://news.google.com/rss/articles/CBMiX0FVX3lxTE52LWVBcEJFVm5Dcm80UWNfaVNhRy04Qmt0OGhyNklYcGxGYVpCaE1NWFFGak0wWndxWW5FS0RZZU1HaTNKSmdjWnZDbGZKZWZWMkNRN2RBSG9QMmhQWVFj?oc=5\" target=\"_blank\">Artificial intelligence data centers could reach one percent of global electricity demand by 2030</a>&nbsp;&nbsp;<font color=\"#6f6f6f\">Nature</font>","full_text":"Abstract\nThe rapid growth of generative artificial intelligence is increasing global electricity demand and placing new pressure on power systems. Here we show that electricity use by data centers built for artificial intelligence could rise from about 118 terawatt-hours in 2024 to between 239 and 295 terawatt-hours by 2030, or about 1% of global electricity demand. New computing infrastructure is highly concentrated in North America, Western Europe, and the Asia-Pacific, which together account for more than 90% of projected computing capacity. Some regions, including Oregon, Ireland, and Iowa, face greater pressure from concentrated data-center loads, whereas larger systems such as Texas can absorb new demand more effectively. These results indicate that artificial intelligence infrastructure is becoming a structural part of power-system dynamics. The study combines large language model analysis of corporate, policy, and media sources with scenario-based projections of future electricity demand.\nSimilar content being viewed by others\nIntroduction\nThe rapid emergence of generative artificial intelligence (AI) and large-scale data analytics has driven a sharp expansion of computational demand1,2. As AI models grow exponentially in size and complexity, their training and inference require vast computing power and data throughput, driving record investment in high-performance data centers and digital infrastructure (Figs. S1 and S2)3,4,5,6,7. According to Bain & Company’s Global Technology Report 2025, sustaining the computational requirements of AI expansion could generate nearly US$2 trillion in annual revenues by 2030–equivalent to the combined GDP of the world’s ten largest emerging economies8. This “AI infrastructure boom” is transforming data centers into the industrial backbone of the digital economy and, increasingly, a major source of electricity demand9,10.\nElectricity use has risen sharply in parallel with the digital transition. The International Energy Agency (IEA, 2025) projects that global data-center electricity consumption will more than double—from about 415 TWh in 2024 to roughly 945 TWh by 2030—with the United States and China accounting for almost 80% of the increase11,12. AI-specific facilities rely on GPU-based computation, which enables large-scale parallel processing but consumes up to six times more power than conventional racks, elevating both cooling intensity and peak-load requirements. These facilities are increasingly concentrated in regions with abundant renewable resources, low electricity prices, and favorable climates, yet such clustering also amplifies local grid stress and transmission constraints. As AI campuses scale up from megawatt to gigawatt levels, ensuring a reliable and low-carbon electricity supply has become a strategic challenge for utilities, regulators, and technology developers13,14.\nA growing body of literature has examined the energy footprint of digital infrastructure. Early studies quantified electricity consumption and emissions associated with information and communication technologies15,16,17, while more recent work highlights that the growth of AI workloads—particularly large-scale GPU clusters and generative models—may substantially increase electricity demand and associated environmental impacts18,19,20,21,22.\nTo quantify these impacts, prior research has developed two main classes of modeling approaches. Top-down methods estimate electricity consumption using aggregate indicators such as data traffic, computing capacity, or ICT statistics15,16,17. While these approaches provide useful global benchmarks, they often struggle to capture the heterogeneous and rapidly evolving workloads associated with AI computing. Bottom-up engineering models, in contrast, estimate energy use from facility-level characteristics, including server hardware configurations, cooling systems, and operational efficiency metrics19,23. However, such models are frequently constrained by limited data availability and incomplete knowledge of proprietary hyperscale infrastructure deployments20. Recent studies attempt to bridge these perspectives by linking computational workloads to energy demand through task-based or model-level accounting frameworks20,24,25, complemented by scenario analyses such as the IEA’s Energy and AI report11.\nDespite these advances, existing approaches remain largely focused on aggregate demand estimation or facility-level energy accounting, providing limited insight into how firm-level infrastructure investment, spatial clustering of AI data centers, and regional electricity-system characteristics jointly shape the geography and magnitude of AI-related energy demand.\nThis study addresses this gap by developing an integrated analytical framework that links AI infrastructure deployment, spatial siting patterns, and regional power-system impacts. Unlike conventional approaches that assume stable relationships between computing intensity, hardware efficiency, and electricity use—an assumption that breaks down under rapidly evolving AI workloads26,27,28—our framework employs a large language model (LLM)-based semantic retrieval and inference to dynamically extract firm-level strategic signals from heterogeneous corporate disclosures. These signals are combined with scenario-based electricity-demand projections to assess how AI data-center expansion translates into regional electricity demand and grid pressure. This approach moves beyond aggregate demand projections and facility-level efficiency metrics to capture how firm-level AI deployment strategies and spatial clustering of data centers generate localized electricity-system pressures.\nEmpirically, we focus on six leading technology firms—Amazon, Microsoft, Google, Meta, Oracle, and Apple—which collectively account for ~70−75% of global hyperscale and cloud-linked data-center electricity demand (Table S1). Their global footprint and relatively transparent reporting enable consistent identification of siting patterns and electricity-demand trajectories across regions.\nUnderstanding and managing this emerging compute-energy nexus is critical to ensuring that the rapid expansion of AI infrastructure evolves in tandem with the development of adequate and reliable electricity systems14. In this study, the compute-energy nexus is modeled as a demand-side relationship: we quantify and spatially allocate the electricity consumption generated by AI data-center expansion. Bidirectional interactions—such as demand response, flexible load scheduling, or participation in electricity markets—are recognized as important dimensions of this nexus but fall outside the scope of the current framework, and are identified as directions for future research. Accordingly, this study addresses three central research questions: (i) what spatial factors shape the siting patterns of large-scale AI data centers; (ii) how firm-level infrastructure investments translate into electricity demand trajectories; and (iii) what regional electricity-system pressures arise from the spatial clustering of AI data-center loads.\nAddressing these questions, we find that AI-oriented infrastructure is concentrating in a limited set of regions—chiefly North America, Western Europe, and the Asia-Pacific, which together account for more than 90% of projected compute capacity—and that aggregate electricity consumption by six leading operators is projected to rise from roughly 118 TWh in 2024 to between 239 and 295 TWh by 2030, equivalent to about 1% of projected global electricity demand. This concentration is associated with markedly higher relative electricity-demand pressure in some host regions, such as Oregon, Ireland, and Iowa, than in regions with larger electricity systems, such as Texas. Together, these findings indicate that the geography of AI compute is becoming a structural feature of regional power-system dynamics rather than a peripheral digital load, underscoring the value of anticipatory planning that aligns computational growth with electricity-system development.\nResults\nGlobal clustering and heterogeneity in AI data-center siting\nAI data-center locations are identified using a large language model (LLM)-based inference approach that extracts and classifies site-level information from public disclosures, assigning each candidate site a probability of AI-oriented deployment (see Methods). Reported siting patterns reflect three components: (i) outputs of the LLM/RAG framework, specifically the site-level probabilities and determinant intensity scores derived from semantic retrieval and sentiment analysis; (ii) evidence contained in the underlying source documents—corporate announcements, SEC filings, sustainability reports, and media coverage—from which the LLM extracts signals; and (iii) the authors’ interpretation of these outputs in the context of the broader literature, which synthesizes these signals into regional and firm-level characterizations. Figure 1 illustrates the global spatial evolution and underlying determinants of AI data-center siting, with legacy infrastructure and AI-specific siting layers presented separately in Supplementary Fig. S3.\nLegacy data centers (blue markers) are relatively dispersed, reflecting historical priorities such as latency reduction and proximity to users. In contrast, newly identified AI-specific data centers (color-coded circles) exhibit strong geographic concentration, forming dense clusters in a limited number of regions. These clusters are most pronounced in North America, Europe, and the Asia-Pacific, which together account for over 90% of projected AI compute capacity among leading firms. The dominant corridors align with regions characterized by strong energy availability, mature grid and fiber-optic infrastructure, favorable climatic conditions, and supportive policy environments29,30.\nDistinct regional configurations highlight the heterogeneity of siting determinants. The following characterizations are based on signals extracted by the LLM framework from input documents (corporate filings, sustainability disclosures, and policy sources); the synthesis of these signals into regional interpretations represents the authors’ reading of those extracted patterns. In North America, clusters are concentrated in regions such as Virginia, Texas, Ohio, and North Carolina, which coincide with established hyperscale infrastructure, large electricity markets, and documented policy support for data-center development11,31. These regions have historically attracted cloud infrastructure investment, although underlying conditions such as energy costs, regulatory incentives, and climatic factors vary substantially across locations. In Europe, data-center development is concentrated in the Netherlands, the United Kingdom, and Italy. Renewable-energy availability, regulatory frameworks, and digital infrastructure jointly shape these patterns although their relative importance varies across countries. Nordic regions, in particular, provide favorable conditions due to hydro and wind resources and naturally cool climates30. In the Asia-Pacific region, clusters in Singapore, Taipei, Malaysia, and Japan align with strong digital demand and policy-supported infrastructure expansion. Similarly, emerging sites in the Middle East are associated with state-led investment strategies and land availability.\nFirm-level differentiation further reinforces these spatial dynamics (Fig. 1a, b). To interpret the firm-determinant intensity matrix (Fig. 1b), we classify siting drivers into six dimensions: corporate integration, energy access, policy environment, market demand, infrastructure maturity, and network connectivity. Higher scores indicate that a determinant category appears more frequently and with stronger positive sentiment in retrieved documents associated with a given firm. Among globally scaling firms, Amazon exhibits the broadest geographic diversification (Fig. 1a), expanding across regions with strong market demand and mature infrastructure. Microsoft and Oracle primarily follow policy- and incentive-driven corridors, while Google and Meta anchor their siting strategies in regions with strong energy availability and high network connectivity to reduce latency and operational constraints. By contrast, domestically oriented firms such as Apple adopt a more vertically integrated strategy, concentrating deployments within US regions-particularly those with favorable energy conditions–to optimize operational efficiency and alignment with their broader manufacturing and service ecosystems.\nOverall, these results reveal a structural transition from historically dispersed siting toward geographically concentrated AI infrastructure, with clustering patterns shaped jointly by regional characteristics and firm-level strategies. Energy availability, policy alignment, and network connectivity emerge as dominant determinants of siting decisions, driving the formation of high-density infrastructure corridors.\nElectricity consumption projection: rapid scaling of AI data centers elevates system-level energy demand\nTo capture uncertainty in expansion pace and AI workload adoption, we evaluate three forward scenarios for 2025–2030: a conservative scenario (15% annual growth), a neutral scenario (25%), and an optimistic scenario (35%). These growth rates span the range of recent estimates of data-center and AI-related electricity demand growth reported in the literature and industry forecasts11,31. In all scenarios, baseline data-center electricity consumption grows at 10% annually, while the share of AI-intensive workloads increases over time, reflecting the rising importance of generative-AI computation. Electricity demand is estimated by combining firm-level data-center expansion scenarios with projected AI workload intensity and baseline energy-use trends (see Methods). Figure 2 presents the projected electricity consumption of AI data centers under the three growth scenarios from 2025 to 2030. Figure 2a–f illustrates firm-level trajectories for Amazon, Microsoft, Google, Meta, Oracle, and Apple across conservative, neutral, and optimistic assumptions. The energy demand forecasts for all firms show consistent upward trends, reflecting the intensification of AI workloads and the proliferation of large-model training clusters. Aggregate electricity use among hyperscale operators is projected to rise by ~40−70% over the period, though the pace and scale vary considerably by firm. Amazon, Microsoft, and Google are forecasted to maintain the highest absolute demand, consistent with their extensive cloud-service portfolios and broad geographic reach. Meta follows comparable but more moderate forecast paths, supported by continuous improvements in model efficiency and renewable procurement strategies32,33,34,35,36,37. Apple and Oracle exhibit lower projected energy demand, aligning with their smaller operational footprints and vertically integrated architectures. Across all firms, the optimistic scenario accelerates notably after 2027, coinciding with large-scale deployment of generative-AI infrastructure and training workloads38,39.\nFigure 3 aggregates these firm-level forecasts to depict the global evolution of electricity demand associated with leading AI operators from 2018 to 2030. The solid gray line represents historical estimates derived from industrial disclosures and utility data, while the colored dashed lines correspond to scenario-based projections. Total electricity consumption by these leading firms’ data centers is expected to rise from ~118 TWh in 2024 to between 239 TWh (conservative) and 295 TWh (optimistic) by 2030, implying a compound annual growth rate (CAGR) of roughly 13−17%. This increase is equivalent to adding the electricity demand of a medium-sized national market every two years, or comparable to the annual electricity use of about 27 million US households. The projected magnitude is quantitatively consistent with the International Energy Agency’s (IEA) forecast that total global data center electricity demand will reach around 945 TWh by 2030 (~3% of global power consumption)11, compared with about 1.5% in 2024 (see Methods). These projections collectively illustrate the accelerating energy footprint of AI infrastructure within the broader global power system.\nRegional electricity demand pressure implications of AI data-center clustering\nFigure 4 characterizes the regional distribution of data-center electricity demand and the relative concentration of AI-related loads within regional electricity systems in 2030. Figure 4 displays projected electricity demand by firm and region, revealing a highly uneven spatial distribution of AI-related loads. Fewer than ten regions account for nearly two-thirds of total projected demand, indicating a strong concentration of infrastructure in a limited number of states and countries. Oregon, Virginia, Iowa, Ohio and Ireland emerge as dominant hubs, each exceeding 15 TWh of annual demand under the neutral scenario. These regions are associated with favorable conditions—such as established hyperscale clusters, access to energy resources, and supportive policy environments—that are consistent with large-scale AI deployment. Second-tier regions, including the Texas, Netherlands, and Washington, exhibit moderate but rapidly increasing demand, often driven by firm-specific expansion. Scenario ranges indicate that uncertainty in deployment pace and firm strategy could alter local electricity demand by up to 30%, underscoring the challenges of forecasting and planning for digital-infrastructure growth.\nTo characterize the relative concentration of data-center electricity demand across regions, we construct an electricity demand pressure index (EDPI), defined as the ratio of projected data-center electricity demand to regional total available electricity supply (see Methods). Figure 5a presents the cross-regional evolution of EDPI between 2025 and 2030 under the neutral scenario, revealing a pronounced and spatially uneven increase in demand concentration. A small number of regions—including Oregon, Ireland and Iowa—display markedly elevated EDPI levels in both years, indicating a persistently high concentration of data-center electricity demand relative to available supply. Virginia, Nebraska, and Ohio occupy the upper-middle range of the distribution, reflecting growing demand concentration. The absolute change in EDPI between 2025 and 2030 (ΔEDPI, expressed in percentage points) further reveals a self-reinforcing spatial pattern: regions with already elevated demand concentration tend to experience the largest absolute increases. In contrast, regions with larger electricity systems—such as Texas—exhibit comparatively small ΔEDPI values, reflecting greater available-supply adequacy to absorb incremental AI-related demand (Fig. 5a–c).\nOverall, these results reveal a pronounced spatial asymmetry in the distribution of data-center electricity demand, with a limited number of regions accounting for a disproportionate share of demand relative to their available-electricity adequacy.\nDiscussion\nThis study links AI data-center siting patterns, electricity-demand projections, and regional electricity demand concentration. By combining large language model (LLM)-based infrastructure identification with scenario-based electricity-demand projections, we assess how the spatial expansion of AI computing infrastructure may influence electricity demand and the relative burden on regional electricity systems. A key methodological contribution of this framework is the fusion of qualitative corporate intelligence—extracted by LLMs from financial disclosures, sustainability reports, and strategic announcements—with quantitative electricity demand projection. This combination enables inference of firms’ implicit strategic intent from language patterns in corporate communications, providing a more interpretable and forward-looking basis for demand forecasting than statistical extrapolation from historical consumption trends alone11,20.\nOur results indicate that AI-driven data-center expansion is emerging as an increasingly influential component of global electricity systems7,40. Across firms and regions, the analysis reveals a consistent sequence linking spatial siting patterns, rapidly increasing electricity demand, and rising regional power stress. This pattern suggests that the digital infrastructure supporting AI development is becoming closely intertwined with electricity-system dynamics rather than remaining an exogenous source of demand. As AI-related computing capacity expands and concentrates geographically, it may influence regional load growth, infrastructure investment priorities, and decarbonization pathways41.\nThe clustering of AI data centers in a limited number of favorable regions—primarily in North America, Western Europe, and parts of the Asia-Pacific—creates pronounced spatial asymmetries in load distribution. Regions such as Oregon, Ireland and Iowa—which exhibit elevated EDPI values—face a high relative concentration of data-center electricity demand within their electricity systems, alongside rapid AI infrastructure expansion. In contrast, regions with larger electricity systems, such as Texas, show lower demand concentration relative to total available electricity supply. These patterns highlight how regional electricity-system size, cross-boundary trade flows, and the spatial footprint of AI infrastructure jointly determine the relative exposure of local electricity systems to AI-driven demand growth. The projected growth in electricity demand from AI data centers therefore presents both challenges and opportunities for electricity systems. On one hand, concentrated digital loads could intensify transmission congestion, complicate renewable integration, and increase electricity price volatility if system expansion does not keep pace with demand. On the other hand, large and stable computational loads may provide opportunities for greater integration with electricity systems Data centers can function as anchor customers for long-term renewable power purchase agreements (PPAs), potential participants in flexible demand programs, and focal points for co-located energy storage or hybrid renewable generation42,43. The extent to which these opportunities materialize will depend on the interaction between electricity market design, infrastructure investment, and regulatory frameworks.\nAt the policy level, these findings highlight the potential value of anticipatory frameworks that explicitly incorporate AI-related demand into national energy modeling and long-term planning. Conventional forecasts that treat digital infrastructure as exogenous commercial load risk underestimating the spatial concentration of new demand and its implications for annual energy balancing. Integrating firm-level digital-load projections into planning processes could enable regulators and utilities to better anticipate load growth in high-concentration regions. In emerging markets, policy design may need to balance the benefits of digital-infrastructure development against the risk of exacerbating existing demand-supply imbalances.\nImportantly, the firm-level granularity of this analysis offers a policy-relevant tool that complements aggregate demand forecasts. Because AI-related electricity demand is highly concentrated among a small number of identifiable corporate actors, regulators and utilities can use firm-level demand trajectories to engage with specific operators rather than with a generic “technology sector.” This granularity may support more targeted policy interventions—such as renewable procurement requirements, locational guidance, or mandatory load disclosure—calibrated to the expansion strategies and geographical footprints of individual firms.\nSeveral limitations of the current framework warrant explicit acknowledgment. First, the LLM-based inference relies on corporate sustainability reports, press releases, and annual filings, which may contain optimistic bias regarding efficiency trajectories, renewable integration commitments, or future investment plans. To the extent that such documents overstate actual or planned improvements, the sentiment-based siting probabilities may be biased toward high-profile or low-carbon locations, potentially understating deployment in less prominently disclosed regions. Moreover, regions where AI infrastructure projects are less frequently disclosed (such as Africa or Australia), less systematically reported, or still at an early stage of development may be underrepresented within the inferred deployment database. Second, the efficiency-gain assumption of 1−3% annually—calibrated to historical hyperscale patterns—reflects a central scenario. Further improvements in chip efficiency, including more energy-efficient accelerators or advances in cooling architectures, could reduce energy intensity below projected levels. Third, the EDPI is constructed from aggregate annual energy quantities—available electricity supply (generation plus positive net imports) and projected data-center demand in the numerator—and does not capture intra-regional transmission network constraints or interconnection queue delays. In systems such as the PJM interconnection (a regional transmission organization coordinating wholesale electricity markets and grid operations across parts of the eastern United States, including Virginia), grid integration bottlenecks and interconnection queues can be binding constraints on the ability to absorb new large loads, even when aggregate available supply appears sufficient. The EDPI may therefore underestimate localized electricity-system pressures in transmission-constrained areas, and future extensions of this framework should incorporate network-level indicators. Fourth, the firm-level analysis covers six technology companies that together account for ~70−75% of global hyperscale and cloud-linked data-center electricity demand. The remaining share comprises a heterogeneous set of operators—including smaller cloud-service providers, colocation facility operators, telecommunications companies with data-center divisions, and regional or national operators—that are excluded primarily because their infrastructure footprints and electricity consumption are less consistently disclosed in public filings and their siting patterns are not readily identifiable by the LLM-based inference approach applied here. Excluding this segment may lead to an underestimate of total AI-related electricity demand. Fifth, available electricity projections are based on CAGR estimates from 2019–2024 with a small sample and pandemic impact, so they may not fully capture future changes in electricity systems, including renewable deployment, grid expansion, and infrastructure investments. In particular, in regions with relatively small electricity systems, such developments could substantially increase future electricity availability beyond historical trends. Likewise, future national and regional energy policies, market reforms, and unforeseen extreme events may significantly reshape electricity-supply trajectories.\nMore broadly, the co-evolution of digitalization and electrification marks a conceptual turning point in energy-system governance13. AI infrastructure is transforming electricity demand from a passive outcome of economic activity into an active determinant of system configuration. Its social and economic value increasingly depends on the sustainability of its power source and its integration within low-carbon grids. Future research should quantify the feedback loops between AI-compute trajectories, carbon intensity, and investment flows across the energy-digital interface. Several specific directions are particularly promising. First, the co-optimization of AI data-center siting with other emerging high-load technologies—such as hydrogen electrolyzers—could help minimize transmission congestion and renewable curtailment; strategic spatial allocation of such loads has been shown to reduce boundary reinforcement costs44. Second, the interaction between AI data-center loads and the electrification of transport warrants coordinated analysis, as both introduce large and potentially concurrent demands on distribution and transmission infrastructure; holistic planning frameworks that integrate gigawatt-scale AI campuses alongside electric-vehicle charging have been identified as an emerging priority45. Third, the integration of data centers into urban energy systems through waste heat recovery for building heating offers a dual benefit of improved cooling efficiency and urban decarbonization; coordinated control strategies for building thermal flexibility informed by data-center waste heat could reduce local distribution network congestion46. Fourth, the application of advanced risk assessment and stochastic optimization techniques—analogous to those used in commodity supply-chain management47—could better account for uncertainties in energy pricing, grid availability, and AI compute demand, supporting more robust infrastructure planning under deep uncertainty. Understanding and governing the coupling between computational growth and electricity systems will be central to achieving a resilient, low-carbon digital economy.\nMethods\nThis study develops a hybrid retrieval-augmented generation (RAG) + large language model (LLM) forecasting framework to predict the geographic and operational evolution of hyperscale data centers operated by major technology companies. The framework integrates structured document retrieval, language sentiment analysis, and LLM-based contextual reasoning to infer (i) potential expansion locations, (ii) baseline technical parameters, and (iii) multi-year energy trajectories.\nThe workflow (Fig. 6) consists of six main stages: (1) construction of a RAG knowledge base from multi-year corporate and regional sources, (2) sentiment-aware identification of likely future data center locations, (3) retrieval of location-specific contextual information, (4) extraction of baseline operational parameters for the current year, (5) multi-year forecasting of key technical and environmental metrics using an LLM-guided model and physically interpretable energy equations, and (6) construction of the Electricity Demand Pressure Index (EDPI) that compares projected data center demand against the available regional electricity supply capacity.\nKnowledge-base construction\nA multi-source text corpus was constructed for each firm, including press releases, annual earnings reports, environmental sustainability disclosures, infrastructure investment announcements, government policy documents, and major news articles from 2015 to 2025. All texts were preprocessed and embedded using the Hugging Face all-MiniLM-L6-v2 model, then indexed in a FAISS (Facebook AI similarity search) vector database for semantic retrieval. Each entry retained metadata such as publication date, geographic reference, and document type, enabling targeted retrieval (e.g., “AWS Taiwan 2025 expansion energy infrastructure”).\nLocation Identification via RAG + LLM\nTo infer likely future expansion sites between 2025 and 2030, the framework explicitly prompts the LLM to predict which geographic regions a firm is most likely to expand into. This prediction is based on contextual evidence retrieved from the RAG knowledge base, which includes firm sustainability reports, capital expenditure statements, and government or media coverage.\nThe RAG system is first queried using firm-specific and temporal prompts, such as {firm} 2025–2030 data center investment or construction plans. The retrieved documents contain both qualitative statements (e.g., “strong regional demand growth,” “strategic infrastructure partnership”) and quantitative indicators (e.g., “$2B allocated to APAC expansion”).\nEach document is analyzed by the LLM not only for content relevance but also for its sentiment and linguistic tone. Positive sentiment toward investment, expansion, or infrastructure development—particularly in earnings calls or annual reports—acts as an implicit prior that increases the model’s belief in a region’s likelihood of receiving new data center capacity. Conversely, negative sentiment (e.g., mentions of cost control, divestment, or policy uncertainty) decreases the associated probability weight.\nThe LLM then synthesizes these cues into a structured probability distribution over candidate locations, producing outputs of the form: P(expansion at site i ∣evidence) = fLLM(retrieved text embeddings, sentiment scores).\nThis probabilistic representation is used in later stages to weight forecasts of capacity growth and energy consumption at each site.\nBaseline parameter extraction\nFor each confirmed or high-probability site, the system used a RAG-conditioned LLM prompt to extract current-year baseline parameters describing the data center’s IT and facility operations. The baseline operational parameters used in the forecasting framework are summarized in Table 1, including accelerator counts, average power draw, utilization factors, operating hours, and power usage effectiveness (PUE).\nThe model was instructed to reason explicitly about local grid mix, cooling efficiency, and infrastructure maturity, and to avoid copying assumptions between regions. The extracted parameters included: {Ntrain, Ninference, Pavg,train, Pavg,inference, utrain, uinference, Htrain, Hinference, PUE, g}\nAll outputs were formatted as structured JSON (JavaScript Object Notation) objects, accompanied by short notes summarizing assumptions and references. Each variable corresponds to a measurable operational attribute of the local compute fleet.\nForecasting future parameters\nA second RAG + LLM chain projected the time evolution of these parameters over a 5-year horizon. The prompt incorporated both historical operational trends and retrieved regional information, constraining LLM reasoning with empirically observed hyperscale patterns: 10−20% annual accelerator capacity growth, 1−3% annual efficiency gains, and gradual improvements in utilization. These parameters characterize the supply-side evolution of infrastructure, specifically capturing the growth in installed accelerator capacity and incremental improvements in energy efficiency at the facility level. The model generated structured forecasts for each year t ∈ [2025, 2030], each annotated with explanatory notes and relevant document sources.\nEnergy computation\nForecasted operational parameters were post-processed in Python to derive the physical energy demand of each site. The IT load energy (EIT) represents the direct electrical consumption of computing equipment, while the estimated AI data center energy demand (EDC) includes both IT and non-IT overheads such as cooling and power distribution losses.\nHere, N, Pavg, u, and H denote the number of accelerators, average power draw (kW), utilization factor, and annual operating hours, respectively. EIT is thus measured in MWh, and EDC scales this load by the facility’s power usage effectiveness (PUE), capturing site-specific cooling and infrastructure efficiency.\nImplementation\nThe full pipeline was implemented in Python using the LangChain framework for composable RAG pipelines, FAISS for vector retrieval, and the OpenAI GPT-4o-mini model as the reasoning component (temperature = 0.3). All modules were orchestrated via reproducible runnables with strict JSON parsing for downstream numerical analysis.\nSentiment-aware expansion probability modeling\nThe sentiment analysis stage produced, for each candidate region i, an expansion likelihood score Pi defined as:\nwhere Si is the normalized sentiment score for region i (ranging from −1 to +1), and Ri is the retrieval relevance score (cosine similarity in the embedding space). These probabilities informed the sampling or prioritization of forecast targets in subsequent steps. Regions frequently mentioned in corporate reports and described with positive financial tone thus received elevated probabilities of selection.\nThis coupling of semantic retrieval and sentiment tone weighting enables the model to capture implicit strategic intent in corporate communications—an important predictor of near-term data center expansion behavior.\nBy fusing semantic retrieval, sentiment weighting, and physical energy modeling, this hybrid framework bridges qualitative corporate signals and quantitative operational forecasting. The approach captures how firms’ language choices in financial and environmental disclosures correlate with real-world infrastructure trajectories, enabling a more interpretable and evidence-grounded forecast of global data center development.\nScenario design and aggregation\nTo capture uncertainty in both firm-level expansion intensity and technological efficiency, we constructed three forward scenarios s-conservative, neutral, and optimistic-for 2025–2030. Each scenario jointly controls (i) the rate of new AI data-center additions and (ii) the efficiency gains of all operating sites. The annual growth rate of new AI load is set to gnew,s ∈ {0.15, 0.25, 0.35} for scenario s among the three scenarios, while existing stock consumption for each firm grows at gstock = 0.10. Specifically, the three forward scenarios for AI-driven data-center expansion (15, 25, and 35% annual growth) are designed to span conservative, central, and high-growth trajectories of AI workload expansion. These values are informed by recent industry and research evidence on the growth of AI compute demand and infrastructure. AI training workloads are estimated to grow at around 22% annually, while inference workloads may expand at rates approaching 35% over the next five years48. Other estimates indicate that overall AI-related infrastructure demand may grow in the range of 20−25% annually in baseline scenarios49. Based on this range of evidence, the 15% scenario reflects a lower-bound trajectory consistent with moderate expansion in AI deployment, the 25% scenario captures the central tendency of observed infrastructure growth rates, and the 35% scenario represents an upper-tail case aligned with rapid scaling of inference workloads and AI-ready capacity. These scenarios are designed to capture the empirically observed range of AI compute and infrastructure growth, while also reflecting uncertainty in future AI adoption intensity and technological scaling.\nFirm-specific AI shares pAI(f, t) increase gradually over time to reflect the rising share of AI workloads in total compute demand. Firm-specific AI shares pAI(f, t) were defined to represent the evolving proportion of AI-related compute workloads within each operator’s total data-center activity. For each firm f, pAI(f, t) increases over time according to a discrete schedule calibrated to reflect the firm’s relative intensity of AI adoption between 2025 and 2030:\nwhere p1(f), p2(f), and p3(f) correspond to the firm’s baseline (2025), mid-phase (2026–2027), and mature-phase (2028–2030) AI workload shares, respectively. Formally, this can be expressed as:\nThe parameter set {p1(f), p2(f), p3(f)} is firm-specific, capturing heterogeneity in AI adoption trajectories across major hyperscale operators:\nThis stepwise formulation approximates each firm’s increasing allocation of computational resources to AI model training and inference. The heterogeneity in pAI(f, t) reflects differences in investment magnitude, infrastructure build-out, and strategic focus among major operators. Firms with large-scale and early commitments to generative AI—such as Amazon, Google, Meta, and Microsoft—exhibit steeper trajectories, consistent with multibillion-dollar data-center expansions, dedicated AI accelerator deployment (e.g., Trainium, H100, and TPUv5), and vertically integrated AI ecosystems. By contrast, vertically integrated or enterprise-focused operators such as Apple and Oracle show slower growth, consistent with smaller-scale AI infrastructure investment, reliance on consumer-device optimization rather than large-model training, and a later pivot toward cloud-based AI services. Accordingly, the evolution of pAI(f, t) captures each firm’s position along the spectrum of AI-driven digital-infrastructure transformation, serving as a proxy for the relative share of compute capacity devoted to AI workloads within their overall fleet.\nFor each firm f, the existing stock electricity consumption evolves as\nNew-site electricity demand follows compounding growth of AI workloads and the firm-specific AI share schedule:\nThe total firm-level electricity demand is then\nTo represent uncertainty arising from multiple independent siting pathways, we simulate a set of Pt,s plausible firm-region configurations for each year t and scenario s. Global electricity demand in year t and scenario s is the ensemble mean:\nand the range\ndefines the shaded uncertainty bands in the scenario plots, where \\({E}_{t,s,p}^{{{\\rm{glob}}}}\\) denotes the global total electricity demand in year t under scenario s for projection path p ∈ {1, …, Pt,s}. The 2024 observed total serves as the common anchor point for all trajectories.\nRegional allocation of energy\nFirm-level electricity demand is spatially distributed according to two complementary weights: (i) AI siting probabilities derived from the LLM siting model, and (ii) historical stock weights from the existing data-center inventory.\nFor AI-related new sites, firm-year location weights are proportional to modeled AI energy at each location ℓ:\nHere \\({E}_{f,t,\\ell }^{AI,\\,{\\mbox{loc}}\\,}\\) denotes the modeled AI-related electricity demand assigned to firm f at location ℓ in year t, as determined from the LLM-based siting model and corresponding regional AI workload intensity.\nFor legacy and non-AI sites, we use historical weights based on observed facility counts. For legacy and non-AI sites, we use historical weights based on observed facility counts:\nLegacy (historical) data-center counts refer to the number of distinct operating campuses under each firm’s control as of end-2024, including both owned and long-term leased facilities. These are derived from the observed 2015–2024 dataset and serve as proxies for baseline spatial capacity prior to the AI expansion. Since 2024, the global data-center landscape has entered a rapid expansion phase driven by generative-AI workloads, with AI infrastructure investment reaching a record $57 billion globally50. This “AI data-center boom” marks a structural inflection in digital-infrastructure growth.\nLet r(ℓ) map locations to regions. Regional electricity use for firm f and scenario s is:\nSumming over all firms yields regional totals:\nThis approach preserves firm-level energy consistency while ensuring that AI-driven growth concentrates in high-probability siting regions identified by the LLM model, and that legacy loads follow empirically observed footprints.\nCross-validation\nAggregate electricity consumption by the six leading hyperscale operators is projected to rise from ~118 TWh in 2024 to 239–295 TWh by 2030. This projection is broadly consistent with the International Energy Agency’s (IEA) estimate that global data-center electricity demand will reach roughly 945 TWh by 2030 (IEA, 2025). To benchmark the modeled values, we derive an implied 2030 consumption level for the six leading firms based on the IEA’s global forecast. Assuming (i) that hyperscale data centers account for 70% of global data-center electricity use (Synergy Research Group, 2024), and (ii) that the top six operators collectively represent 40% of hyperscale activity, the corresponding implied electricity consumption is:\nThis cross-validation demonstrates that the projected range of 239–295 TWh lies well within the values implied by independent international forecasts. The close correspondence reinforces confidence in the representativeness of the modeled global totals and supports the credibility of the upper-bound scenario in light of established external benchmarks. Overall, the modeled range remains broadly aligned with independent global projections and reflects plausible hyperscale market dynamics.\nElectricity demand pressure index (EDPI)\nTo assess the relative demand burden associated with data-center expansion, we construct an electricity demand pressure index (EDPI), defined as the ratio of annual data-center electricity demand to the regional available electricity supply—comprising in-region generation plus positive net electricity imports. For each region r and year t:\nwhere \\({E}_{r,t}^{{\\mathrm{Avail}}} = {E}_{r,t}^{{\\mathrm{Gen}}} + {\\max} \\left(0, {{E}^{{\\mathrm{NetImport}}}_{r,t}}\\right), {{E}^{{\\mathrm{DC}}}_{r,t}}\\) denotes total annual data-center electricity demand (in TWh), with \\({E}_{r,t}^{Gen}\\) denoting regional electricity generation (TWh), \\({E}_{r,t}^{{\\mathrm{Gen}}}\\) denoting net electricity imports, and \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) denoting total available electricity supply (TWh). Positive net imports are included because imported electricity contributes directly to regional electricity supply. For net-exporting regions, historical electricity exports are treated as adjustable rather than permanently fixed. Accordingly, negative net imports are set to zero in the denominator, reflecting the assumption that part of the electricity historically exported could instead be retained to serve additional local electricity demand associated with AI data-center expansion. This approximation avoids overstating electricity demand pressure in regions that are persistent net electricity exporters while providing a consistent basis for cross-regional comparison.\nFor US states, \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) is derived from EIA state electricity profiles as total retail electricity sales, which by the energy balance identity equals in-state net generation plus net interstate electricity receipts. For countries, \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) is the sum of total electricity generation and positive net electricity imports, both drawn from the Ember Yearly Full Release dataset (2019–2024). The EDPI is interpreted as a demand-share indicator capturing the relative contribution of data-center loads to regional electricity supply.\nRegional available electricity supply is projected using historical trends of available electricity as a baseline approximation. Specifically, available electricity is extrapolated from the period 2019–2024 using a region-specific compound annual growth rate (CAGR), consistent with the five-year window used in the CAGR formula below.\nRegional available electricity baselines\n(\\({E}_{r,2024}^{{\\mathrm{Avail}}}\\)) were obtained from EIA State Electricity Profiles (US states) and the Ember Yearly Full Release dataset (countries), covering 2019–2024, expressed in terawatt-hours (TWh). Available electricity data were extrapolated to 2030 using a compound annual growth rate (CAGR):\nwith \\({{\\mbox{CAGR}}}_{avail,r}={\\left(\\frac{{E}_{r,2024}^{Avail}}{{E}_{r,2019}^{Avail}}\\right)}^{1/5}-1\\).\nRegional electricity available supply is projected using region-specific compound annual growth rates (CAGR) estimated from historical observations during 2019–2024. This approach assumes that recent historical trends provide a reasonable baseline approximation of future electricity-system evolution. The resulting projections should therefore be interpreted as baseline-trend estimates rather than policy-conditioned forecasts.\nData availability\nThe publicly available electricity data used in this study were obtained from the Ember Yearly Electricity Data Explorer (https://ember-energy.org/data/yearly-electricity-data/) and the US Energy Information Administration State Electricity Profiles (https://www.eia.gov/electricity/state/). The International Energy Agency’s Energy and AI report (https://www.iea.org/reports/energy-and-ai) was used for external benchmarking. Additional public corporate and contextual information was obtained from SEC EDGAR filings (https://www.sec.gov/edgar/search/), corporate annual and sustainability reports, press releases, government and grid-operator publications, and media reports. Historical facility-level data for Amazon, Microsoft, Google, Meta, Oracle, and Apple were obtained from S&P Capital IQ (https://www.capitaliq.com/). The processed data and model outputs underlying the data-center siting estimates, electricity-demand projections, electricity demand pressure index calculations, figures, and tables are publicly available at https://github.com/dchen219/ai-data-center-electricity.\nCode availability\nThe custom code used for document processing, LLM-based data-center siting inference, electricity-demand projection, electricity demand pressure index calculation, and figure generation is publicly available at https://github.com/dchen219/ai-data-center-electricity. Analyses were conducted in Python using LangChain, FAISS, the Hugging Face all-MiniLM-L6-v2 embedding model, and OpenAI’s GPT-4o-mini model.\nReferences\n- Stylianou, N. et al. Inside the relentless race for AI capacity. Financial Times. https://ig.ft.com/ai-data-centres/ (2025). \n- Storey, V. C., Yue, W. T., Zhao, J. L. & L., R. Generative artificial intelligence: evolving technology, growing societal impact, and opportunities for information systems research. Inform. Syst. Front. https://doi.org/10.1007/s10796-025-10581-7 (2025). \n- Cho, J. & Moon, J. H. Numerical coupling of energy efficiency and thermal performance for cold plate cooling optimization in high-density compute AI data centers. Energy Build. 348, 116441 (2025). \n- Pilz, K. F., Mahmood, Y. & Heim, L. AI’s Power Requirements under Exponential Growth: Extrapolating ai Data Center Power Demand and Assessing its Potential Impact on U.S. Competitiveness. Tech. Rep. https://www.rand.org/pubs/research_reports/RRA3572-1.html (RAND Corporation, 2025). \n- Oxford Institute for Energy Studies. Artificial Intelligence and its Implications for Electricity Systems. Tech. Rep. 145. https://www.oxfordenergy.org/wpcms/wp-content/uploads/2025/05/OEF-145.pdf (Oxford Institute for Energy Studies, 2025)) \n- Chen, X., Wang, X., Colacelli, A., Lee, M. & Xie, L. Electricity demand and grid impacts of AI data centers: Challenges and prospects. Nexus 3, 100162 https://doi.org/10.1016/j.ynexs.2026.100162 (2026) \n- Stojkovic, J., Zhang, C., Goiri, Íñigo., Torrellas, J. & Choukse, E. DynamoLLM: designing LLM inference clusters for performance and energy efficiency. In 2025 IEEE International Symposium on High Performance Computer Architecture (HPCA) 1348–1362 (2025). \n- Crawford, D., Schallehn, M., Renno, P., Hanbury, P. & Cannarsi, A. How can We meet AI’s Insatiable Demand for Compute Power? https://www.bain.com/insights/how-can-we-meet-ais-insatiable-demand-for-compute-power-technology-report-2025/ (Bain & Company, 2025). \n- Ngata, W. M. et al. The cloud next door: Investigating the environmental and socioeconomic strain of datacenters on local communities. In Proc. 2025 ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies, COMPASS ’25 769–774 (Association for Computing Machinery, 2025). \n- Li, P., Yang, J., Islam, M. A. & Ren, S. Making AI less ‘Thirsty’. Commun. ACM 68, 54–61 (2025). \n- International Energy Agency. Energy and AI. https://www.iea.org/reports/energy-and-ai (2025). \n- Bogmans, C. et al. Power hungry: how AI will drive energy demand. IMF Working Paper 2025/081, International Monetary Fund (2025). \n- Wu, C.-J. et al. Sustainable AI: environmental implications, challenges, and opportunities. In Proc. Machine Learning and Systems 795–813. https://proceedings.mlsys.org/paper_files/paper/2022/hash/462211f67c7d858f663355eff93b745e-Abstract.html (2022). \n- Chen, Z.-M. et al. AI carbon footprint in China sets to double post-2030 carbon peaking. Energy Econ. 150, 108880 (2025). \n- Belkhir, L. & Elmeligi, A. Assessing ICT global emissions footprint: trends to 2040 and recommendations. J. Clean. Prod. 177, 448–463 (2018). \n- Masanet, E., Shehabi, A., Lei, N., Smith, S. & Koomey, J. Recalibrating global data center energy-use estimates. Science 367, 984–986 (2020). \n- Malmodin, J., Lövehagen, N., Bergmark, P. & Lundén, D. ICT sector electricity consumption and greenhouse gas emissions–2020 outcome. Telecommun. Policy 48, 102701 (2024). \n- Debus, C., Piraud, M., Streit, A., Theis, F. & Götz, M. Reporting electricity consumption is essential for sustainable AI. Nat. Mach. Intell. 5, 1176–1178 (2023). \n- Cao, Z. et al. Data center sustainability: revisits and outlooks. IEEE Trans. Sustain. Comput. 9, 236–248 (2024). \n- Xiao, T., Nerini, F. F., Matthews, H. D., Tavoni, M. & You, F. Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA. Nat. Sustain. 8, 1541–1553 (2025). \n- Zhang, Y., Tang, H., Li, H. & Wang, S. Integration and interaction of next-generation AI-focused data centers with smart grids and district energy systems: the state-of-the-art, opportunities and challenges. Renew. Sustain. Energy Rev. 224, 116097 (2025). \n- Han, J. et al. Data-driven distributionally robust optimization of low-carbon data center energy systems considering multi-task response and renewable energy uncertainty. J. Build. Eng. 102, 111937 (2025). \n- Nøland, J. K., Hjelmeland, M. & Korpås, M. Will energy-hungry AI create a baseload power demand boom? IEEE Access 12, 110353–110360 (2024). \n- Luccioni, A. S., Viguier, S. & Ligozat, A.-L. Estimating the carbon footprint of BLOOM, a 176B parameter language model. J. Mach. Learn. Res. 24, 1–15 (2023). \n- Tomlinson, B., Black, R. W., Patterson, D. J. & Torrance, A. W. The carbon emissions of writing and illustrating are lower for AI than for humans. Sci. Rep. 14, 3732 (2024). \n- Ahmed, K. M. U., Bollen, M. H. J. & Alvarez, M. A review of data centers energy consumption and reliability modeling. IEEE Access 9, 152536–152563 (2021). \n- Liu, Y. et al. Energy consumption and emission mitigation prediction based on data center traffic and PUE for global data centers. Glob. Energy Interconnect. 3, 272–282 https://doi.org/10.1016/j.gloei.2020.07.008 (2020). \n- Siddik, M. A. B., Shehabi, A. & Marston, L. The environmental footprint of data centers in the United States. Environ. Res. Lett. https://doi.org/10.1088/1748-9326/abfba1(2021). \n- Cao, Z., Zhou, X., Hu, H., Wang, Z. & Wen, Y. Toward a systematic survey for carbon neutral data centers. IEEE Commun. Surv. Tutor. 24, 895–936 https://doi.org/10.1109/COMST.2022.3161275 (2022). \n- Hyvönen, J. et al. Potential of solar photovoltaics and waste heat utilization in cold climate data centers: case study: Finland and northern Japan. Renew. Sustain. Energy Rev. https://doi.org/10.1016/j.rser.2024.114619 (2024). \n- Shehabi, A. et al. 2024 United States data center energy usage report. Tech. Rep. LBNL-2001637. https://doi.org/10.71468/P1WC7Q (Lawrence Berkeley National Laboratory, 2024). \n- Synergy Research Group. Hyperscale data center count hits 1136; average size increases; US accounts for 54% of total capacity. https://www.srgresearch.com/articles/hyperscale-data-center-count-hits-1136-average-size-increases-us-accounts-for-54-of-total-capacity (2025). \n- Donnellan, D. et al. Uptime Institute global data center survey results 2024. Tech. Rep. https://uptimeinstitute.com/resources/research-and-reports/uptime-institute-global-data-center-survey-results-2024 (Uptime Institute, 2024). \n- Khan, O. Scale your AI transformation with a powerful, secure, and adaptive cloud infrastructure. Microsoft Azure Blog. https://azure.microsoft.com/en-us/blog/scale-your-ai-transformation-with-a-powerful-secure-and-adaptive-cloud-infrastructure/ (2024). \n- Google Sustainability. 2023 Environmental report. https://sustainability.google/reports/google-2023-environmental-report/ (2023) \n- Meta Platforms, Inc. Annual report on Form 10-K for the fiscal year ended december 31, 2024. Tech. Rep. SEC Accession No. 0001326801-25-000017, U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/1326801/000132680125000017/meta-20241231.htm (2025). \n- Sivasubramanian, S. Welcome to a new era of building in the cloud with generative AI on AWS. AWS Machine Learning Blog. https://aws.amazon.com/blogs/machine-learning/welcome-to-a-new-era-of-building-in-the-cloud-with-generative-ai-on-aws/ (2023). \n- Apple Inc. Annual report on Form 10-K for the fiscal year ended september 28, 2024. U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/320193/000032019324000123/aapl-20240928.htm (2024). \n- Oracle Corporation. Oracle offers first zettascale cloud computing cluster. Oracle News press release. https://www.oracle.com/news/announcement/ocw24-oracle-offers-first-zettascale-cloud-computing-cluster-2024-09-11/ (2024). \n- Lambert, S., Caron, E., Lefèvre, L. & Grivel, R. Consolidation of virtual machines to reduce energy consumption of data centers by using ballooning, sharing and swapping mechanisms. Future Gener. Comput. Syst. 174, 107968 https://doi.org/10.1016/j.future.2025.107968 (2026). \n- Bolón-Canedo, V., Morán-Fernández, L., Cancela, B. & Alonso-Betanzos, A. A review of green artificial intelligence: towards a more sustainable future. Neurocomputing 599, 128096 https://doi.org/10.1016/j.neucom.2024.128096 (2024). \n- de Roucy-Rochegonde, L. & Buffard, A. AI, Data Centers and Energy Demand: Reassessing and Exploring the Trends. Tech. Rep. https://www.ifri.org/en/papers/ai-data-centers-and-energy-demand-reassessing-and-exploring-trends-0 (Institut français des relations internationales (Ifri), 2025). \n- Goldman Sachs Research. GS SUSTAIN: Generational Growth: AI, Data Centers and the Coming US Power Demand Surge. Tech. Rep. https://www.goldmansachs.com/pdfs/insights/pages/generational-growth-ai-data-centers-and-the-coming-us-power-surge/report.pdf (Goldman Sachs, 2024). \n- Giannelos, S., Konstantelos, I., Pudjianto, D. & Strbac, G. The impact of electrolyser allocation on Great Britain’s electricity transmission system in 2050. Int. J. Hydrog. Energy 202, 153097 https://doi.org/10.1016/j.ijhydene.2025.153097 (2026). \n- Amann, G. et al. E-mobility Deployment and Impact on Grids: Impact of EV and Charging Infrastructure on European T&D Grids: Innovation needs. Technical Report MJ-09-22-246-EN-N (Publications Office of the European Union, 2022). \n- Dong, Z., Zhang, X., Zhang, L., Giannelos, S. & Strbac, G. Flexibility enhancement of urban energy systems through coordinated space heating aggregation of numerous buildings. Appl. Energy 374, 123971 (2024). \n- Giannelos, S., Konstantelos, I. & Strbac, G. Optimal supply chain design using machine learning, risk assessment and optimisation applied to coal distribution. EURO J. Decis. Process. 13, 100062 (2025). \n- Arora, C., Sorel, M. & Sachdeva, P. The next big shifts in AI workloads and hyperscaler strategies. McKinsey & Company. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-next-big-shifts-in-ai-workloads-and-hyperscaler-strategies (2025) \n- Barth, A., Arora, C., Shenai, G., Noffsinger, J. & Sachdeva, P. The data center balance: how US states can navigate the opportunities and challenges. McKinsey & Company. https://www.mckinsey.com/industries/public-sector/our-insights/the-data-center-balance-how-us-states-can-navigate-the-opportunities-and-challenges (2025). \n- Saavedra, R. & Seaward, S. 2025 Data center marketplace: balancing unprecedented opportunity with strategic risk. U.S. Research Report, Colliers. https://www.colliers.com/en/research/nrep-usdc-data-center-marketplace-2025 (2025). \nAcknowledgements\nNot applicable.\nFunding\nAuthors declare no relevant funding.\nAuthor information\nAuthors and Affiliations\nContributions\nD.C. and Y.C. conceived the study. D.C. designed the analytical framework, collected and processed the data, implemented the LLM/RAG-based analysis, conducted the electricity-demand projections and EDPI analysis, prepared the figures and tables, and drafted the manuscript. Z.Z. contributed to data collection, data validation, and interpretation of results. J.Q. and L.C. contributed to methodological development, results visualization, and validation of the analysis. A.K. and Y.C. supervised the study, provided conceptual guidance, and contributed to the interpretation and revision of the manuscript. All authors reviewed, edited, and approved the final manuscript.\nCorresponding author\nEthics declarations\nCompeting interests\nThe authors declare no competing interests.\nPeer review\nPeer review information\nCommunications Sustainability thanks Rahman Khorramfar and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Nandita Basu. A peer review file is available.\nAdditional information\nPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.\nRights and permissions\nOpen Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.\nAbout this article\nCite this article\nChen, D., Zhou, Z., Cai, Y. et al. Artificial intelligence data centers could reach one percent of global electricity demand by 2030. Commun. Sustain. 1, 147 (2026). https://doi.org/10.1038/s44458-026-00152-5\n- Received: \n- Accepted: \n- Published: \n- Version of record: \n- DOI: https://doi.org/10.1038/s44458-026-00152-5","excerpt":"Abstract\nThe rapid growth of generative artificial intelligence is increasing global electricity demand and placing new pressure on power systems. Here we show that electricity use by data centers built for artificial intelligence could rise from about 118 terawatt-hours in 2024 to between 239 and 295 terawatt-hours by 2030, or about 1% of global electricity demand.","extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 62509 characters.","diagnostics_url":"/api/diagnose?url=https%3A//www.nature.com/articles/s44458-026-00152-5","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 62509 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":62509,"summary_length":368,"usable_text_length":62509,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":62509,"summary_length":368}}},"display_formats":["compact","card","full","digest_section","json"]},"daily_stack_record":{"title":"Artificial intelligence data centers could reach one percent of global electricity demand by 2030 - Nature","url":"https://www.nature.com/articles/s44458-026-00152-5","summary":"Abstract\nThe rapid growth of generative artificial intelligence is increasing global electricity demand and placing new pressure on power systems. Here we show that electricity use by data centers built for artificial intelligence could rise from about 118 terawatt-hours in 2024 to between 239 and 295 terawatt-hours by 2030, or about 1% of global electricity demand.","source":"Nature","date":"2026-09-21T11:04:25+00:00","content":"Abstract\nThe rapid growth of generative artificial intelligence is increasing global electricity demand and placing new pressure on power systems. Here we show that electricity use by data centers built for artificial intelligence could rise from about 118 terawatt-hours in 2024 to between 239 and 295 terawatt-hours by 2030, or about 1% of global electricity demand. New computing infrastructure is highly concentrated in North America, Western Europe, and the Asia-Pacific, which together account for more than 90% of projected computing capacity. Some regions, including Oregon, Ireland, and Iowa, face greater pressure from concentrated data-center loads, whereas larger systems such as Texas can absorb new demand more effectively. These results indicate that artificial intelligence infrastructure is becoming a structural part of power-system dynamics. The study combines large language model analysis of corporate, policy, and media sources with scenario-based projections of future electricity demand.\nSimilar content being viewed by others\nIntroduction\nThe rapid emergence of generative artificial intelligence (AI) and large-scale data analytics has driven a sharp expansion of computational demand1,2. As AI models grow exponentially in size and complexity, their training and inference require vast computing power and data throughput, driving record investment in high-performance data centers and digital infrastructure (Figs. S1 and S2)3,4,5,6,7. According to Bain & Company’s Global Technology Report 2025, sustaining the computational requirements of AI expansion could generate nearly US$2 trillion in annual revenues by 2030–equivalent to the combined GDP of the world’s ten largest emerging economies8. This “AI infrastructure boom” is transforming data centers into the industrial backbone of the digital economy and, increasingly, a major source of electricity demand9,10.\nElectricity use has risen sharply in parallel with the digital transition. The International Energy Agency (IEA, 2025) projects that global data-center electricity consumption will more than double—from about 415 TWh in 2024 to roughly 945 TWh by 2030—with the United States and China accounting for almost 80% of the increase11,12. AI-specific facilities rely on GPU-based computation, which enables large-scale parallel processing but consumes up to six times more power than conventional racks, elevating both cooling intensity and peak-load requirements. These facilities are increasingly concentrated in regions with abundant renewable resources, low electricity prices, and favorable climates, yet such clustering also amplifies local grid stress and transmission constraints. As AI campuses scale up from megawatt to gigawatt levels, ensuring a reliable and low-carbon electricity supply has become a strategic challenge for utilities, regulators, and technology developers13,14.\nA growing body of literature has examined the energy footprint of digital infrastructure. Early studies quantified electricity consumption and emissions associated with information and communication technologies15,16,17, while more recent work highlights that the growth of AI workloads—particularly large-scale GPU clusters and generative models—may substantially increase electricity demand and associated environmental impacts18,19,20,21,22.\nTo quantify these impacts, prior research has developed two main classes of modeling approaches. Top-down methods estimate electricity consumption using aggregate indicators such as data traffic, computing capacity, or ICT statistics15,16,17. While these approaches provide useful global benchmarks, they often struggle to capture the heterogeneous and rapidly evolving workloads associated with AI computing. Bottom-up engineering models, in contrast, estimate energy use from facility-level characteristics, including server hardware configurations, cooling systems, and operational efficiency metrics19,23. However, such models are frequently constrained by limited data availability and incomplete knowledge of proprietary hyperscale infrastructure deployments20. Recent studies attempt to bridge these perspectives by linking computational workloads to energy demand through task-based or model-level accounting frameworks20,24,25, complemented by scenario analyses such as the IEA’s Energy and AI report11.\nDespite these advances, existing approaches remain largely focused on aggregate demand estimation or facility-level energy accounting, providing limited insight into how firm-level infrastructure investment, spatial clustering of AI data centers, and regional electricity-system characteristics jointly shape the geography and magnitude of AI-related energy demand.\nThis study addresses this gap by developing an integrated analytical framework that links AI infrastructure deployment, spatial siting patterns, and regional power-system impacts. Unlike conventional approaches that assume stable relationships between computing intensity, hardware efficiency, and electricity use—an assumption that breaks down under rapidly evolving AI workloads26,27,28—our framework employs a large language model (LLM)-based semantic retrieval and inference to dynamically extract firm-level strategic signals from heterogeneous corporate disclosures. These signals are combined with scenario-based electricity-demand projections to assess how AI data-center expansion translates into regional electricity demand and grid pressure. This approach moves beyond aggregate demand projections and facility-level efficiency metrics to capture how firm-level AI deployment strategies and spatial clustering of data centers generate localized electricity-system pressures.\nEmpirically, we focus on six leading technology firms—Amazon, Microsoft, Google, Meta, Oracle, and Apple—which collectively account for ~70−75% of global hyperscale and cloud-linked data-center electricity demand (Table S1). Their global footprint and relatively transparent reporting enable consistent identification of siting patterns and electricity-demand trajectories across regions.\nUnderstanding and managing this emerging compute-energy nexus is critical to ensuring that the rapid expansion of AI infrastructure evolves in tandem with the development of adequate and reliable electricity systems14. In this study, the compute-energy nexus is modeled as a demand-side relationship: we quantify and spatially allocate the electricity consumption generated by AI data-center expansion. Bidirectional interactions—such as demand response, flexible load scheduling, or participation in electricity markets—are recognized as important dimensions of this nexus but fall outside the scope of the current framework, and are identified as directions for future research. Accordingly, this study addresses three central research questions: (i) what spatial factors shape the siting patterns of large-scale AI data centers; (ii) how firm-level infrastructure investments translate into electricity demand trajectories; and (iii) what regional electricity-system pressures arise from the spatial clustering of AI data-center loads.\nAddressing these questions, we find that AI-oriented infrastructure is concentrating in a limited set of regions—chiefly North America, Western Europe, and the Asia-Pacific, which together account for more than 90% of projected compute capacity—and that aggregate electricity consumption by six leading operators is projected to rise from roughly 118 TWh in 2024 to between 239 and 295 TWh by 2030, equivalent to about 1% of projected global electricity demand. This concentration is associated with markedly higher relative electricity-demand pressure in some host regions, such as Oregon, Ireland, and Iowa, than in regions with larger electricity systems, such as Texas. Together, these findings indicate that the geography of AI compute is becoming a structural feature of regional power-system dynamics rather than a peripheral digital load, underscoring the value of anticipatory planning that aligns computational growth with electricity-system development.\nResults\nGlobal clustering and heterogeneity in AI data-center siting\nAI data-center locations are identified using a large language model (LLM)-based inference approach that extracts and classifies site-level information from public disclosures, assigning each candidate site a probability of AI-oriented deployment (see Methods). Reported siting patterns reflect three components: (i) outputs of the LLM/RAG framework, specifically the site-level probabilities and determinant intensity scores derived from semantic retrieval and sentiment analysis; (ii) evidence contained in the underlying source documents—corporate announcements, SEC filings, sustainability reports, and media coverage—from which the LLM extracts signals; and (iii) the authors’ interpretation of these outputs in the context of the broader literature, which synthesizes these signals into regional and firm-level characterizations. Figure 1 illustrates the global spatial evolution and underlying determinants of AI data-center siting, with legacy infrastructure and AI-specific siting layers presented separately in Supplementary Fig. S3.\nLegacy data centers (blue markers) are relatively dispersed, reflecting historical priorities such as latency reduction and proximity to users. In contrast, newly identified AI-specific data centers (color-coded circles) exhibit strong geographic concentration, forming dense clusters in a limited number of regions. These clusters are most pronounced in North America, Europe, and the Asia-Pacific, which together account for over 90% of projected AI compute capacity among leading firms. The dominant corridors align with regions characterized by strong energy availability, mature grid and fiber-optic infrastructure, favorable climatic conditions, and supportive policy environments29,30.\nDistinct regional configurations highlight the heterogeneity of siting determinants. The following characterizations are based on signals extracted by the LLM framework from input documents (corporate filings, sustainability disclosures, and policy sources); the synthesis of these signals into regional interpretations represents the authors’ reading of those extracted patterns. In North America, clusters are concentrated in regions such as Virginia, Texas, Ohio, and North Carolina, which coincide with established hyperscale infrastructure, large electricity markets, and documented policy support for data-center development11,31. These regions have historically attracted cloud infrastructure investment, although underlying conditions such as energy costs, regulatory incentives, and climatic factors vary substantially across locations. In Europe, data-center development is concentrated in the Netherlands, the United Kingdom, and Italy. Renewable-energy availability, regulatory frameworks, and digital infrastructure jointly shape these patterns although their relative importance varies across countries. Nordic regions, in particular, provide favorable conditions due to hydro and wind resources and naturally cool climates30. In the Asia-Pacific region, clusters in Singapore, Taipei, Malaysia, and Japan align with strong digital demand and policy-supported infrastructure expansion. Similarly, emerging sites in the Middle East are associated with state-led investment strategies and land availability.\nFirm-level differentiation further reinforces these spatial dynamics (Fig. 1a, b). To interpret the firm-determinant intensity matrix (Fig. 1b), we classify siting drivers into six dimensions: corporate integration, energy access, policy environment, market demand, infrastructure maturity, and network connectivity. Higher scores indicate that a determinant category appears more frequently and with stronger positive sentiment in retrieved documents associated with a given firm. Among globally scaling firms, Amazon exhibits the broadest geographic diversification (Fig. 1a), expanding across regions with strong market demand and mature infrastructure. Microsoft and Oracle primarily follow policy- and incentive-driven corridors, while Google and Meta anchor their siting strategies in regions with strong energy availability and high network connectivity to reduce latency and operational constraints. By contrast, domestically oriented firms such as Apple adopt a more vertically integrated strategy, concentrating deployments within US regions-particularly those with favorable energy conditions–to optimize operational efficiency and alignment with their broader manufacturing and service ecosystems.\nOverall, these results reveal a structural transition from historically dispersed siting toward geographically concentrated AI infrastructure, with clustering patterns shaped jointly by regional characteristics and firm-level strategies. Energy availability, policy alignment, and network connectivity emerge as dominant determinants of siting decisions, driving the formation of high-density infrastructure corridors.\nElectricity consumption projection: rapid scaling of AI data centers elevates system-level energy demand\nTo capture uncertainty in expansion pace and AI workload adoption, we evaluate three forward scenarios for 2025–2030: a conservative scenario (15% annual growth), a neutral scenario (25%), and an optimistic scenario (35%). These growth rates span the range of recent estimates of data-center and AI-related electricity demand growth reported in the literature and industry forecasts11,31. In all scenarios, baseline data-center electricity consumption grows at 10% annually, while the share of AI-intensive workloads increases over time, reflecting the rising importance of generative-AI computation. Electricity demand is estimated by combining firm-level data-center expansion scenarios with projected AI workload intensity and baseline energy-use trends (see Methods). Figure 2 presents the projected electricity consumption of AI data centers under the three growth scenarios from 2025 to 2030. Figure 2a–f illustrates firm-level trajectories for Amazon, Microsoft, Google, Meta, Oracle, and Apple across conservative, neutral, and optimistic assumptions. The energy demand forecasts for all firms show consistent upward trends, reflecting the intensification of AI workloads and the proliferation of large-model training clusters. Aggregate electricity use among hyperscale operators is projected to rise by ~40−70% over the period, though the pace and scale vary considerably by firm. Amazon, Microsoft, and Google are forecasted to maintain the highest absolute demand, consistent with their extensive cloud-service portfolios and broad geographic reach. Meta follows comparable but more moderate forecast paths, supported by continuous improvements in model efficiency and renewable procurement strategies32,33,34,35,36,37. Apple and Oracle exhibit lower projected energy demand, aligning with their smaller operational footprints and vertically integrated architectures. Across all firms, the optimistic scenario accelerates notably after 2027, coinciding with large-scale deployment of generative-AI infrastructure and training workloads38,39.\nFigure 3 aggregates these firm-level forecasts to depict the global evolution of electricity demand associated with leading AI operators from 2018 to 2030. The solid gray line represents historical estimates derived from industrial disclosures and utility data, while the colored dashed lines correspond to scenario-based projections. Total electricity consumption by these leading firms’ data centers is expected to rise from ~118 TWh in 2024 to between 239 TWh (conservative) and 295 TWh (optimistic) by 2030, implying a compound annual growth rate (CAGR) of roughly 13−17%. This increase is equivalent to adding the electricity demand of a medium-sized national market every two years, or comparable to the annual electricity use of about 27 million US households. The projected magnitude is quantitatively consistent with the International Energy Agency’s (IEA) forecast that total global data center electricity demand will reach around 945 TWh by 2030 (~3% of global power consumption)11, compared with about 1.5% in 2024 (see Methods). These projections collectively illustrate the accelerating energy footprint of AI infrastructure within the broader global power system.\nRegional electricity demand pressure implications of AI data-center clustering\nFigure 4 characterizes the regional distribution of data-center electricity demand and the relative concentration of AI-related loads within regional electricity systems in 2030. Figure 4 displays projected electricity demand by firm and region, revealing a highly uneven spatial distribution of AI-related loads. Fewer than ten regions account for nearly two-thirds of total projected demand, indicating a strong concentration of infrastructure in a limited number of states and countries. Oregon, Virginia, Iowa, Ohio and Ireland emerge as dominant hubs, each exceeding 15 TWh of annual demand under the neutral scenario. These regions are associated with favorable conditions—such as established hyperscale clusters, access to energy resources, and supportive policy environments—that are consistent with large-scale AI deployment. Second-tier regions, including the Texas, Netherlands, and Washington, exhibit moderate but rapidly increasing demand, often driven by firm-specific expansion. Scenario ranges indicate that uncertainty in deployment pace and firm strategy could alter local electricity demand by up to 30%, underscoring the challenges of forecasting and planning for digital-infrastructure growth.\nTo characterize the relative concentration of data-center electricity demand across regions, we construct an electricity demand pressure index (EDPI), defined as the ratio of projected data-center electricity demand to regional total available electricity supply (see Methods). Figure 5a presents the cross-regional evolution of EDPI between 2025 and 2030 under the neutral scenario, revealing a pronounced and spatially uneven increase in demand concentration. A small number of regions—including Oregon, Ireland and Iowa—display markedly elevated EDPI levels in both years, indicating a persistently high concentration of data-center electricity demand relative to available supply. Virginia, Nebraska, and Ohio occupy the upper-middle range of the distribution, reflecting growing demand concentration. The absolute change in EDPI between 2025 and 2030 (ΔEDPI, expressed in percentage points) further reveals a self-reinforcing spatial pattern: regions with already elevated demand concentration tend to experience the largest absolute increases. In contrast, regions with larger electricity systems—such as Texas—exhibit comparatively small ΔEDPI values, reflecting greater available-supply adequacy to absorb incremental AI-related demand (Fig. 5a–c).\nOverall, these results reveal a pronounced spatial asymmetry in the distribution of data-center electricity demand, with a limited number of regions accounting for a disproportionate share of demand relative to their available-electricity adequacy.\nDiscussion\nThis study links AI data-center siting patterns, electricity-demand projections, and regional electricity demand concentration. By combining large language model (LLM)-based infrastructure identification with scenario-based electricity-demand projections, we assess how the spatial expansion of AI computing infrastructure may influence electricity demand and the relative burden on regional electricity systems. A key methodological contribution of this framework is the fusion of qualitative corporate intelligence—extracted by LLMs from financial disclosures, sustainability reports, and strategic announcements—with quantitative electricity demand projection. This combination enables inference of firms’ implicit strategic intent from language patterns in corporate communications, providing a more interpretable and forward-looking basis for demand forecasting than statistical extrapolation from historical consumption trends alone11,20.\nOur results indicate that AI-driven data-center expansion is emerging as an increasingly influential component of global electricity systems7,40. Across firms and regions, the analysis reveals a consistent sequence linking spatial siting patterns, rapidly increasing electricity demand, and rising regional power stress. This pattern suggests that the digital infrastructure supporting AI development is becoming closely intertwined with electricity-system dynamics rather than remaining an exogenous source of demand. As AI-related computing capacity expands and concentrates geographically, it may influence regional load growth, infrastructure investment priorities, and decarbonization pathways41.\nThe clustering of AI data centers in a limited number of favorable regions—primarily in North America, Western Europe, and parts of the Asia-Pacific—creates pronounced spatial asymmetries in load distribution. Regions such as Oregon, Ireland and Iowa—which exhibit elevated EDPI values—face a high relative concentration of data-center electricity demand within their electricity systems, alongside rapid AI infrastructure expansion. In contrast, regions with larger electricity systems, such as Texas, show lower demand concentration relative to total available electricity supply. These patterns highlight how regional electricity-system size, cross-boundary trade flows, and the spatial footprint of AI infrastructure jointly determine the relative exposure of local electricity systems to AI-driven demand growth. The projected growth in electricity demand from AI data centers therefore presents both challenges and opportunities for electricity systems. On one hand, concentrated digital loads could intensify transmission congestion, complicate renewable integration, and increase electricity price volatility if system expansion does not keep pace with demand. On the other hand, large and stable computational loads may provide opportunities for greater integration with electricity systems Data centers can function as anchor customers for long-term renewable power purchase agreements (PPAs), potential participants in flexible demand programs, and focal points for co-located energy storage or hybrid renewable generation42,43. The extent to which these opportunities materialize will depend on the interaction between electricity market design, infrastructure investment, and regulatory frameworks.\nAt the policy level, these findings highlight the potential value of anticipatory frameworks that explicitly incorporate AI-related demand into national energy modeling and long-term planning. Conventional forecasts that treat digital infrastructure as exogenous commercial load risk underestimating the spatial concentration of new demand and its implications for annual energy balancing. Integrating firm-level digital-load projections into planning processes could enable regulators and utilities to better anticipate load growth in high-concentration regions. In emerging markets, policy design may need to balance the benefits of digital-infrastructure development against the risk of exacerbating existing demand-supply imbalances.\nImportantly, the firm-level granularity of this analysis offers a policy-relevant tool that complements aggregate demand forecasts. Because AI-related electricity demand is highly concentrated among a small number of identifiable corporate actors, regulators and utilities can use firm-level demand trajectories to engage with specific operators rather than with a generic “technology sector.” This granularity may support more targeted policy interventions—such as renewable procurement requirements, locational guidance, or mandatory load disclosure—calibrated to the expansion strategies and geographical footprints of individual firms.\nSeveral limitations of the current framework warrant explicit acknowledgment. First, the LLM-based inference relies on corporate sustainability reports, press releases, and annual filings, which may contain optimistic bias regarding efficiency trajectories, renewable integration commitments, or future investment plans. To the extent that such documents overstate actual or planned improvements, the sentiment-based siting probabilities may be biased toward high-profile or low-carbon locations, potentially understating deployment in less prominently disclosed regions. Moreover, regions where AI infrastructure projects are less frequently disclosed (such as Africa or Australia), less systematically reported, or still at an early stage of development may be underrepresented within the inferred deployment database. Second, the efficiency-gain assumption of 1−3% annually—calibrated to historical hyperscale patterns—reflects a central scenario. Further improvements in chip efficiency, including more energy-efficient accelerators or advances in cooling architectures, could reduce energy intensity below projected levels. Third, the EDPI is constructed from aggregate annual energy quantities—available electricity supply (generation plus positive net imports) and projected data-center demand in the numerator—and does not capture intra-regional transmission network constraints or interconnection queue delays. In systems such as the PJM interconnection (a regional transmission organization coordinating wholesale electricity markets and grid operations across parts of the eastern United States, including Virginia), grid integration bottlenecks and interconnection queues can be binding constraints on the ability to absorb new large loads, even when aggregate available supply appears sufficient. The EDPI may therefore underestimate localized electricity-system pressures in transmission-constrained areas, and future extensions of this framework should incorporate network-level indicators. Fourth, the firm-level analysis covers six technology companies that together account for ~70−75% of global hyperscale and cloud-linked data-center electricity demand. The remaining share comprises a heterogeneous set of operators—including smaller cloud-service providers, colocation facility operators, telecommunications companies with data-center divisions, and regional or national operators—that are excluded primarily because their infrastructure footprints and electricity consumption are less consistently disclosed in public filings and their siting patterns are not readily identifiable by the LLM-based inference approach applied here. Excluding this segment may lead to an underestimate of total AI-related electricity demand. Fifth, available electricity projections are based on CAGR estimates from 2019–2024 with a small sample and pandemic impact, so they may not fully capture future changes in electricity systems, including renewable deployment, grid expansion, and infrastructure investments. In particular, in regions with relatively small electricity systems, such developments could substantially increase future electricity availability beyond historical trends. Likewise, future national and regional energy policies, market reforms, and unforeseen extreme events may significantly reshape electricity-supply trajectories.\nMore broadly, the co-evolution of digitalization and electrification marks a conceptual turning point in energy-system governance13. AI infrastructure is transforming electricity demand from a passive outcome of economic activity into an active determinant of system configuration. Its social and economic value increasingly depends on the sustainability of its power source and its integration within low-carbon grids. Future research should quantify the feedback loops between AI-compute trajectories, carbon intensity, and investment flows across the energy-digital interface. Several specific directions are particularly promising. First, the co-optimization of AI data-center siting with other emerging high-load technologies—such as hydrogen electrolyzers—could help minimize transmission congestion and renewable curtailment; strategic spatial allocation of such loads has been shown to reduce boundary reinforcement costs44. Second, the interaction between AI data-center loads and the electrification of transport warrants coordinated analysis, as both introduce large and potentially concurrent demands on distribution and transmission infrastructure; holistic planning frameworks that integrate gigawatt-scale AI campuses alongside electric-vehicle charging have been identified as an emerging priority45. Third, the integration of data centers into urban energy systems through waste heat recovery for building heating offers a dual benefit of improved cooling efficiency and urban decarbonization; coordinated control strategies for building thermal flexibility informed by data-center waste heat could reduce local distribution network congestion46. Fourth, the application of advanced risk assessment and stochastic optimization techniques—analogous to those used in commodity supply-chain management47—could better account for uncertainties in energy pricing, grid availability, and AI compute demand, supporting more robust infrastructure planning under deep uncertainty. Understanding and governing the coupling between computational growth and electricity systems will be central to achieving a resilient, low-carbon digital economy.\nMethods\nThis study develops a hybrid retrieval-augmented generation (RAG) + large language model (LLM) forecasting framework to predict the geographic and operational evolution of hyperscale data centers operated by major technology companies. The framework integrates structured document retrieval, language sentiment analysis, and LLM-based contextual reasoning to infer (i) potential expansion locations, (ii) baseline technical parameters, and (iii) multi-year energy trajectories.\nThe workflow (Fig. 6) consists of six main stages: (1) construction of a RAG knowledge base from multi-year corporate and regional sources, (2) sentiment-aware identification of likely future data center locations, (3) retrieval of location-specific contextual information, (4) extraction of baseline operational parameters for the current year, (5) multi-year forecasting of key technical and environmental metrics using an LLM-guided model and physically interpretable energy equations, and (6) construction of the Electricity Demand Pressure Index (EDPI) that compares projected data center demand against the available regional electricity supply capacity.\nKnowledge-base construction\nA multi-source text corpus was constructed for each firm, including press releases, annual earnings reports, environmental sustainability disclosures, infrastructure investment announcements, government policy documents, and major news articles from 2015 to 2025. All texts were preprocessed and embedded using the Hugging Face all-MiniLM-L6-v2 model, then indexed in a FAISS (Facebook AI similarity search) vector database for semantic retrieval. Each entry retained metadata such as publication date, geographic reference, and document type, enabling targeted retrieval (e.g., “AWS Taiwan 2025 expansion energy infrastructure”).\nLocation Identification via RAG + LLM\nTo infer likely future expansion sites between 2025 and 2030, the framework explicitly prompts the LLM to predict which geographic regions a firm is most likely to expand into. This prediction is based on contextual evidence retrieved from the RAG knowledge base, which includes firm sustainability reports, capital expenditure statements, and government or media coverage.\nThe RAG system is first queried using firm-specific and temporal prompts, such as {firm} 2025–2030 data center investment or construction plans. The retrieved documents contain both qualitative statements (e.g., “strong regional demand growth,” “strategic infrastructure partnership”) and quantitative indicators (e.g., “$2B allocated to APAC expansion”).\nEach document is analyzed by the LLM not only for content relevance but also for its sentiment and linguistic tone. Positive sentiment toward investment, expansion, or infrastructure development—particularly in earnings calls or annual reports—acts as an implicit prior that increases the model’s belief in a region’s likelihood of receiving new data center capacity. Conversely, negative sentiment (e.g., mentions of cost control, divestment, or policy uncertainty) decreases the associated probability weight.\nThe LLM then synthesizes these cues into a structured probability distribution over candidate locations, producing outputs of the form: P(expansion at site i ∣evidence) = fLLM(retrieved text embeddings, sentiment scores).\nThis probabilistic representation is used in later stages to weight forecasts of capacity growth and energy consumption at each site.\nBaseline parameter extraction\nFor each confirmed or high-probability site, the system used a RAG-conditioned LLM prompt to extract current-year baseline parameters describing the data center’s IT and facility operations. The baseline operational parameters used in the forecasting framework are summarized in Table 1, including accelerator counts, average power draw, utilization factors, operating hours, and power usage effectiveness (PUE).\nThe model was instructed to reason explicitly about local grid mix, cooling efficiency, and infrastructure maturity, and to avoid copying assumptions between regions. The extracted parameters included: {Ntrain, Ninference, Pavg,train, Pavg,inference, utrain, uinference, Htrain, Hinference, PUE, g}\nAll outputs were formatted as structured JSON (JavaScript Object Notation) objects, accompanied by short notes summarizing assumptions and references. Each variable corresponds to a measurable operational attribute of the local compute fleet.\nForecasting future parameters\nA second RAG + LLM chain projected the time evolution of these parameters over a 5-year horizon. The prompt incorporated both historical operational trends and retrieved regional information, constraining LLM reasoning with empirically observed hyperscale patterns: 10−20% annual accelerator capacity growth, 1−3% annual efficiency gains, and gradual improvements in utilization. These parameters characterize the supply-side evolution of infrastructure, specifically capturing the growth in installed accelerator capacity and incremental improvements in energy efficiency at the facility level. The model generated structured forecasts for each year t ∈ [2025, 2030], each annotated with explanatory notes and relevant document sources.\nEnergy computation\nForecasted operational parameters were post-processed in Python to derive the physical energy demand of each site. The IT load energy (EIT) represents the direct electrical consumption of computing equipment, while the estimated AI data center energy demand (EDC) includes both IT and non-IT overheads such as cooling and power distribution losses.\nHere, N, Pavg, u, and H denote the number of accelerators, average power draw (kW), utilization factor, and annual operating hours, respectively. EIT is thus measured in MWh, and EDC scales this load by the facility’s power usage effectiveness (PUE), capturing site-specific cooling and infrastructure efficiency.\nImplementation\nThe full pipeline was implemented in Python using the LangChain framework for composable RAG pipelines, FAISS for vector retrieval, and the OpenAI GPT-4o-mini model as the reasoning component (temperature = 0.3). All modules were orchestrated via reproducible runnables with strict JSON parsing for downstream numerical analysis.\nSentiment-aware expansion probability modeling\nThe sentiment analysis stage produced, for each candidate region i, an expansion likelihood score Pi defined as:\nwhere Si is the normalized sentiment score for region i (ranging from −1 to +1), and Ri is the retrieval relevance score (cosine similarity in the embedding space). These probabilities informed the sampling or prioritization of forecast targets in subsequent steps. Regions frequently mentioned in corporate reports and described with positive financial tone thus received elevated probabilities of selection.\nThis coupling of semantic retrieval and sentiment tone weighting enables the model to capture implicit strategic intent in corporate communications—an important predictor of near-term data center expansion behavior.\nBy fusing semantic retrieval, sentiment weighting, and physical energy modeling, this hybrid framework bridges qualitative corporate signals and quantitative operational forecasting. The approach captures how firms’ language choices in financial and environmental disclosures correlate with real-world infrastructure trajectories, enabling a more interpretable and evidence-grounded forecast of global data center development.\nScenario design and aggregation\nTo capture uncertainty in both firm-level expansion intensity and technological efficiency, we constructed three forward scenarios s-conservative, neutral, and optimistic-for 2025–2030. Each scenario jointly controls (i) the rate of new AI data-center additions and (ii) the efficiency gains of all operating sites. The annual growth rate of new AI load is set to gnew,s ∈ {0.15, 0.25, 0.35} for scenario s among the three scenarios, while existing stock consumption for each firm grows at gstock = 0.10. Specifically, the three forward scenarios for AI-driven data-center expansion (15, 25, and 35% annual growth) are designed to span conservative, central, and high-growth trajectories of AI workload expansion. These values are informed by recent industry and research evidence on the growth of AI compute demand and infrastructure. AI training workloads are estimated to grow at around 22% annually, while inference workloads may expand at rates approaching 35% over the next five years48. Other estimates indicate that overall AI-related infrastructure demand may grow in the range of 20−25% annually in baseline scenarios49. Based on this range of evidence, the 15% scenario reflects a lower-bound trajectory consistent with moderate expansion in AI deployment, the 25% scenario captures the central tendency of observed infrastructure growth rates, and the 35% scenario represents an upper-tail case aligned with rapid scaling of inference workloads and AI-ready capacity. These scenarios are designed to capture the empirically observed range of AI compute and infrastructure growth, while also reflecting uncertainty in future AI adoption intensity and technological scaling.\nFirm-specific AI shares pAI(f, t) increase gradually over time to reflect the rising share of AI workloads in total compute demand. Firm-specific AI shares pAI(f, t) were defined to represent the evolving proportion of AI-related compute workloads within each operator’s total data-center activity. For each firm f, pAI(f, t) increases over time according to a discrete schedule calibrated to reflect the firm’s relative intensity of AI adoption between 2025 and 2030:\nwhere p1(f), p2(f), and p3(f) correspond to the firm’s baseline (2025), mid-phase (2026–2027), and mature-phase (2028–2030) AI workload shares, respectively. Formally, this can be expressed as:\nThe parameter set {p1(f), p2(f), p3(f)} is firm-specific, capturing heterogeneity in AI adoption trajectories across major hyperscale operators:\nThis stepwise formulation approximates each firm’s increasing allocation of computational resources to AI model training and inference. The heterogeneity in pAI(f, t) reflects differences in investment magnitude, infrastructure build-out, and strategic focus among major operators. Firms with large-scale and early commitments to generative AI—such as Amazon, Google, Meta, and Microsoft—exhibit steeper trajectories, consistent with multibillion-dollar data-center expansions, dedicated AI accelerator deployment (e.g., Trainium, H100, and TPUv5), and vertically integrated AI ecosystems. By contrast, vertically integrated or enterprise-focused operators such as Apple and Oracle show slower growth, consistent with smaller-scale AI infrastructure investment, reliance on consumer-device optimization rather than large-model training, and a later pivot toward cloud-based AI services. Accordingly, the evolution of pAI(f, t) captures each firm’s position along the spectrum of AI-driven digital-infrastructure transformation, serving as a proxy for the relative share of compute capacity devoted to AI workloads within their overall fleet.\nFor each firm f, the existing stock electricity consumption evolves as\nNew-site electricity demand follows compounding growth of AI workloads and the firm-specific AI share schedule:\nThe total firm-level electricity demand is then\nTo represent uncertainty arising from multiple independent siting pathways, we simulate a set of Pt,s plausible firm-region configurations for each year t and scenario s. Global electricity demand in year t and scenario s is the ensemble mean:\nand the range\ndefines the shaded uncertainty bands in the scenario plots, where \\({E}_{t,s,p}^{{{\\rm{glob}}}}\\) denotes the global total electricity demand in year t under scenario s for projection path p ∈ {1, …, Pt,s}. The 2024 observed total serves as the common anchor point for all trajectories.\nRegional allocation of energy\nFirm-level electricity demand is spatially distributed according to two complementary weights: (i) AI siting probabilities derived from the LLM siting model, and (ii) historical stock weights from the existing data-center inventory.\nFor AI-related new sites, firm-year location weights are proportional to modeled AI energy at each location ℓ:\nHere \\({E}_{f,t,\\ell }^{AI,\\,{\\mbox{loc}}\\,}\\) denotes the modeled AI-related electricity demand assigned to firm f at location ℓ in year t, as determined from the LLM-based siting model and corresponding regional AI workload intensity.\nFor legacy and non-AI sites, we use historical weights based on observed facility counts. For legacy and non-AI sites, we use historical weights based on observed facility counts:\nLegacy (historical) data-center counts refer to the number of distinct operating campuses under each firm’s control as of end-2024, including both owned and long-term leased facilities. These are derived from the observed 2015–2024 dataset and serve as proxies for baseline spatial capacity prior to the AI expansion. Since 2024, the global data-center landscape has entered a rapid expansion phase driven by generative-AI workloads, with AI infrastructure investment reaching a record $57 billion globally50. This “AI data-center boom” marks a structural inflection in digital-infrastructure growth.\nLet r(ℓ) map locations to regions. Regional electricity use for firm f and scenario s is:\nSumming over all firms yields regional totals:\nThis approach preserves firm-level energy consistency while ensuring that AI-driven growth concentrates in high-probability siting regions identified by the LLM model, and that legacy loads follow empirically observed footprints.\nCross-validation\nAggregate electricity consumption by the six leading hyperscale operators is projected to rise from ~118 TWh in 2024 to 239–295 TWh by 2030. This projection is broadly consistent with the International Energy Agency’s (IEA) estimate that global data-center electricity demand will reach roughly 945 TWh by 2030 (IEA, 2025). To benchmark the modeled values, we derive an implied 2030 consumption level for the six leading firms based on the IEA’s global forecast. Assuming (i) that hyperscale data centers account for 70% of global data-center electricity use (Synergy Research Group, 2024), and (ii) that the top six operators collectively represent 40% of hyperscale activity, the corresponding implied electricity consumption is:\nThis cross-validation demonstrates that the projected range of 239–295 TWh lies well within the values implied by independent international forecasts. The close correspondence reinforces confidence in the representativeness of the modeled global totals and supports the credibility of the upper-bound scenario in light of established external benchmarks. Overall, the modeled range remains broadly aligned with independent global projections and reflects plausible hyperscale market dynamics.\nElectricity demand pressure index (EDPI)\nTo assess the relative demand burden associated with data-center expansion, we construct an electricity demand pressure index (EDPI), defined as the ratio of annual data-center electricity demand to the regional available electricity supply—comprising in-region generation plus positive net electricity imports. For each region r and year t:\nwhere \\({E}_{r,t}^{{\\mathrm{Avail}}} = {E}_{r,t}^{{\\mathrm{Gen}}} + {\\max} \\left(0, {{E}^{{\\mathrm{NetImport}}}_{r,t}}\\right), {{E}^{{\\mathrm{DC}}}_{r,t}}\\) denotes total annual data-center electricity demand (in TWh), with \\({E}_{r,t}^{Gen}\\) denoting regional electricity generation (TWh), \\({E}_{r,t}^{{\\mathrm{Gen}}}\\) denoting net electricity imports, and \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) denoting total available electricity supply (TWh). Positive net imports are included because imported electricity contributes directly to regional electricity supply. For net-exporting regions, historical electricity exports are treated as adjustable rather than permanently fixed. Accordingly, negative net imports are set to zero in the denominator, reflecting the assumption that part of the electricity historically exported could instead be retained to serve additional local electricity demand associated with AI data-center expansion. This approximation avoids overstating electricity demand pressure in regions that are persistent net electricity exporters while providing a consistent basis for cross-regional comparison.\nFor US states, \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) is derived from EIA state electricity profiles as total retail electricity sales, which by the energy balance identity equals in-state net generation plus net interstate electricity receipts. For countries, \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) is the sum of total electricity generation and positive net electricity imports, both drawn from the Ember Yearly Full Release dataset (2019–2024). The EDPI is interpreted as a demand-share indicator capturing the relative contribution of data-center loads to regional electricity supply.\nRegional available electricity supply is projected using historical trends of available electricity as a baseline approximation. Specifically, available electricity is extrapolated from the period 2019–2024 using a region-specific compound annual growth rate (CAGR), consistent with the five-year window used in the CAGR formula below.\nRegional available electricity baselines\n(\\({E}_{r,2024}^{{\\mathrm{Avail}}}\\)) were obtained from EIA State Electricity Profiles (US states) and the Ember Yearly Full Release dataset (countries), covering 2019–2024, expressed in terawatt-hours (TWh). Available electricity data were extrapolated to 2030 using a compound annual growth rate (CAGR):\nwith \\({{\\mbox{CAGR}}}_{avail,r}={\\left(\\frac{{E}_{r,2024}^{Avail}}{{E}_{r,2019}^{Avail}}\\right)}^{1/5}-1\\).\nRegional electricity available supply is projected using region-specific compound annual growth rates (CAGR) estimated from historical observations during 2019–2024. This approach assumes that recent historical trends provide a reasonable baseline approximation of future electricity-system evolution. The resulting projections should therefore be interpreted as baseline-trend estimates rather than policy-conditioned forecasts.\nData availability\nThe publicly available electricity data used in this study were obtained from the Ember Yearly Electricity Data Explorer (https://ember-energy.org/data/yearly-electricity-data/) and the US Energy Information Administration State Electricity Profiles (https://www.eia.gov/electricity/state/). The International Energy Agency’s Energy and AI report (https://www.iea.org/reports/energy-and-ai) was used for external benchmarking. Additional public corporate and contextual information was obtained from SEC EDGAR filings (https://www.sec.gov/edgar/search/), corporate annual and sustainability reports, press releases, government and grid-operator publications, and media reports. Historical facility-level data for Amazon, Microsoft, Google, Meta, Oracle, and Apple were obtained from S&P Capital IQ (https://www.capitaliq.com/). The processed data and model outputs underlying the data-center siting estimates, electricity-demand projections, electricity demand pressure index calculations, figures, and tables are publicly available at https://github.com/dchen219/ai-data-center-electricity.\nCode availability\nThe custom code used for document processing, LLM-based data-center siting inference, electricity-demand projection, electricity demand pressure index calculation, and figure generation is publicly available at https://github.com/dchen219/ai-data-center-electricity. Analyses were conducted in Python using LangChain, FAISS, the Hugging Face all-MiniLM-L6-v2 embedding model, and OpenAI’s GPT-4o-mini model.\nReferences\n- Stylianou, N. et al. Inside the relentless race for AI capacity. Financial Times. https://ig.ft.com/ai-data-centres/ (2025). \n- Storey, V. C., Yue, W. T., Zhao, J. L. & L., R. Generative artificial intelligence: evolving technology, growing societal impact, and opportunities for information systems research. Inform. Syst. Front. https://doi.org/10.1007/s10796-025-10581-7 (2025). \n- Cho, J. & Moon, J. H. Numerical coupling of energy efficiency and thermal performance for cold plate cooling optimization in high-density compute AI data centers. Energy Build. 348, 116441 (2025). \n- Pilz, K. F., Mahmood, Y. & Heim, L. AI’s Power Requirements under Exponential Growth: Extrapolating ai Data Center Power Demand and Assessing its Potential Impact on U.S. Competitiveness. Tech. Rep. https://www.rand.org/pubs/research_reports/RRA3572-1.html (RAND Corporation, 2025). \n- Oxford Institute for Energy Studies. Artificial Intelligence and its Implications for Electricity Systems. Tech. Rep. 145. https://www.oxfordenergy.org/wpcms/wp-content/uploads/2025/05/OEF-145.pdf (Oxford Institute for Energy Studies, 2025)) \n- Chen, X., Wang, X., Colacelli, A., Lee, M. & Xie, L. Electricity demand and grid impacts of AI data centers: Challenges and prospects. Nexus 3, 100162 https://doi.org/10.1016/j.ynexs.2026.100162 (2026) \n- Stojkovic, J., Zhang, C., Goiri, Íñigo., Torrellas, J. & Choukse, E. DynamoLLM: designing LLM inference clusters for performance and energy efficiency. In 2025 IEEE International Symposium on High Performance Computer Architecture (HPCA) 1348–1362 (2025). \n- Crawford, D., Schallehn, M., Renno, P., Hanbury, P. & Cannarsi, A. How can We meet AI’s Insatiable Demand for Compute Power? https://www.bain.com/insights/how-can-we-meet-ais-insatiable-demand-for-compute-power-technology-report-2025/ (Bain & Company, 2025). \n- Ngata, W. M. et al. The cloud next door: Investigating the environmental and socioeconomic strain of datacenters on local communities. In Proc. 2025 ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies, COMPASS ’25 769–774 (Association for Computing Machinery, 2025). \n- Li, P., Yang, J., Islam, M. A. & Ren, S. Making AI less ‘Thirsty’. Commun. ACM 68, 54–61 (2025). \n- International Energy Agency. Energy and AI. https://www.iea.org/reports/energy-and-ai (2025). \n- Bogmans, C. et al. Power hungry: how AI will drive energy demand. IMF Working Paper 2025/081, International Monetary Fund (2025). \n- Wu, C.-J. et al. Sustainable AI: environmental implications, challenges, and opportunities. In Proc. Machine Learning and Systems 795–813. https://proceedings.mlsys.org/paper_files/paper/2022/hash/462211f67c7d858f663355eff93b745e-Abstract.html (2022). \n- Chen, Z.-M. et al. AI carbon footprint in China sets to double post-2030 carbon peaking. Energy Econ. 150, 108880 (2025). \n- Belkhir, L. & Elmeligi, A. Assessing ICT global emissions footprint: trends to 2040 and recommendations. J. Clean. Prod. 177, 448–463 (2018). \n- Masanet, E., Shehabi, A., Lei, N., Smith, S. & Koomey, J. Recalibrating global data center energy-use estimates. Science 367, 984–986 (2020). \n- Malmodin, J., Lövehagen, N., Bergmark, P. & Lundén, D. ICT sector electricity consumption and greenhouse gas emissions–2020 outcome. Telecommun. Policy 48, 102701 (2024). \n- Debus, C., Piraud, M., Streit, A., Theis, F. & Götz, M. Reporting electricity consumption is essential for sustainable AI. Nat. Mach. Intell. 5, 1176–1178 (2023). \n- Cao, Z. et al. Data center sustainability: revisits and outlooks. IEEE Trans. Sustain. Comput. 9, 236–248 (2024). \n- Xiao, T., Nerini, F. F., Matthews, H. D., Tavoni, M. & You, F. Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA. Nat. Sustain. 8, 1541–1553 (2025). \n- Zhang, Y., Tang, H., Li, H. & Wang, S. Integration and interaction of next-generation AI-focused data centers with smart grids and district energy systems: the state-of-the-art, opportunities and challenges. Renew. Sustain. Energy Rev. 224, 116097 (2025). \n- Han, J. et al. Data-driven distributionally robust optimization of low-carbon data center energy systems considering multi-task response and renewable energy uncertainty. J. Build. Eng. 102, 111937 (2025). \n- Nøland, J. K., Hjelmeland, M. & Korpås, M. Will energy-hungry AI create a baseload power demand boom? IEEE Access 12, 110353–110360 (2024). \n- Luccioni, A. S., Viguier, S. & Ligozat, A.-L. Estimating the carbon footprint of BLOOM, a 176B parameter language model. J. Mach. Learn. Res. 24, 1–15 (2023). \n- Tomlinson, B., Black, R. W., Patterson, D. J. & Torrance, A. W. The carbon emissions of writing and illustrating are lower for AI than for humans. Sci. Rep. 14, 3732 (2024). \n- Ahmed, K. M. U., Bollen, M. H. J. & Alvarez, M. A review of data centers energy consumption and reliability modeling. IEEE Access 9, 152536–152563 (2021). \n- Liu, Y. et al. Energy consumption and emission mitigation prediction based on data center traffic and PUE for global data centers. Glob. Energy Interconnect. 3, 272–282 https://doi.org/10.1016/j.gloei.2020.07.008 (2020). \n- Siddik, M. A. B., Shehabi, A. & Marston, L. The environmental footprint of data centers in the United States. Environ. Res. Lett. https://doi.org/10.1088/1748-9326/abfba1(2021). \n- Cao, Z., Zhou, X., Hu, H., Wang, Z. & Wen, Y. Toward a systematic survey for carbon neutral data centers. IEEE Commun. Surv. Tutor. 24, 895–936 https://doi.org/10.1109/COMST.2022.3161275 (2022). \n- Hyvönen, J. et al. Potential of solar photovoltaics and waste heat utilization in cold climate data centers: case study: Finland and northern Japan. Renew. Sustain. Energy Rev. https://doi.org/10.1016/j.rser.2024.114619 (2024). \n- Shehabi, A. et al. 2024 United States data center energy usage report. Tech. Rep. LBNL-2001637. https://doi.org/10.71468/P1WC7Q (Lawrence Berkeley National Laboratory, 2024). \n- Synergy Research Group. Hyperscale data center count hits 1136; average size increases; US accounts for 54% of total capacity. https://www.srgresearch.com/articles/hyperscale-data-center-count-hits-1136-average-size-increases-us-accounts-for-54-of-total-capacity (2025). \n- Donnellan, D. et al. Uptime Institute global data center survey results 2024. Tech. Rep. https://uptimeinstitute.com/resources/research-and-reports/uptime-institute-global-data-center-survey-results-2024 (Uptime Institute, 2024). \n- Khan, O. Scale your AI transformation with a powerful, secure, and adaptive cloud infrastructure. Microsoft Azure Blog. https://azure.microsoft.com/en-us/blog/scale-your-ai-transformation-with-a-powerful-secure-and-adaptive-cloud-infrastructure/ (2024). \n- Google Sustainability. 2023 Environmental report. https://sustainability.google/reports/google-2023-environmental-report/ (2023) \n- Meta Platforms, Inc. Annual report on Form 10-K for the fiscal year ended december 31, 2024. Tech. Rep. SEC Accession No. 0001326801-25-000017, U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/1326801/000132680125000017/meta-20241231.htm (2025). \n- Sivasubramanian, S. Welcome to a new era of building in the cloud with generative AI on AWS. AWS Machine Learning Blog. https://aws.amazon.com/blogs/machine-learning/welcome-to-a-new-era-of-building-in-the-cloud-with-generative-ai-on-aws/ (2023). \n- Apple Inc. Annual report on Form 10-K for the fiscal year ended september 28, 2024. U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/320193/000032019324000123/aapl-20240928.htm (2024). \n- Oracle Corporation. Oracle offers first zettascale cloud computing cluster. Oracle News press release. https://www.oracle.com/news/announcement/ocw24-oracle-offers-first-zettascale-cloud-computing-cluster-2024-09-11/ (2024). \n- Lambert, S., Caron, E., Lefèvre, L. & Grivel, R. Consolidation of virtual machines to reduce energy consumption of data centers by using ballooning, sharing and swapping mechanisms. Future Gener. Comput. Syst. 174, 107968 https://doi.org/10.1016/j.future.2025.107968 (2026). \n- Bolón-Canedo, V., Morán-Fernández, L., Cancela, B. & Alonso-Betanzos, A. A review of green artificial intelligence: towards a more sustainable future. Neurocomputing 599, 128096 https://doi.org/10.1016/j.neucom.2024.128096 (2024). \n- de Roucy-Rochegonde, L. & Buffard, A. AI, Data Centers and Energy Demand: Reassessing and Exploring the Trends. Tech. Rep. https://www.ifri.org/en/papers/ai-data-centers-and-energy-demand-reassessing-and-exploring-trends-0 (Institut français des relations internationales (Ifri), 2025). \n- Goldman Sachs Research. GS SUSTAIN: Generational Growth: AI, Data Centers and the Coming US Power Demand Surge. Tech. Rep. https://www.goldmansachs.com/pdfs/insights/pages/generational-growth-ai-data-centers-and-the-coming-us-power-surge/report.pdf (Goldman Sachs, 2024). \n- Giannelos, S., Konstantelos, I., Pudjianto, D. & Strbac, G. The impact of electrolyser allocation on Great Britain’s electricity transmission system in 2050. Int. J. Hydrog. Energy 202, 153097 https://doi.org/10.1016/j.ijhydene.2025.153097 (2026). \n- Amann, G. et al. E-mobility Deployment and Impact on Grids: Impact of EV and Charging Infrastructure on European T&D Grids: Innovation needs. Technical Report MJ-09-22-246-EN-N (Publications Office of the European Union, 2022). \n- Dong, Z., Zhang, X., Zhang, L., Giannelos, S. & Strbac, G. Flexibility enhancement of urban energy systems through coordinated space heating aggregation of numerous buildings. Appl. Energy 374, 123971 (2024). \n- Giannelos, S., Konstantelos, I. & Strbac, G. Optimal supply chain design using machine learning, risk assessment and optimisation applied to coal distribution. EURO J. Decis. Process. 13, 100062 (2025). \n- Arora, C., Sorel, M. & Sachdeva, P. The next big shifts in AI workloads and hyperscaler strategies. McKinsey & Company. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-next-big-shifts-in-ai-workloads-and-hyperscaler-strategies (2025) \n- Barth, A., Arora, C., Shenai, G., Noffsinger, J. & Sachdeva, P. The data center balance: how US states can navigate the opportunities and challenges. McKinsey & Company. https://www.mckinsey.com/industries/public-sector/our-insights/the-data-center-balance-how-us-states-can-navigate-the-opportunities-and-challenges (2025). \n- Saavedra, R. & Seaward, S. 2025 Data center marketplace: balancing unprecedented opportunity with strategic risk. U.S. Research Report, Colliers. https://www.colliers.com/en/research/nrep-usdc-data-center-marketplace-2025 (2025). \nAcknowledgements\nNot applicable.\nFunding\nAuthors declare no relevant funding.\nAuthor information\nAuthors and Affiliations\nContributions\nD.C. and Y.C. conceived the study. D.C. designed the analytical framework, collected and processed the data, implemented the LLM/RAG-based analysis, conducted the electricity-demand projections and EDPI analysis, prepared the figures and tables, and drafted the manuscript. Z.Z. contributed to data collection, data validation, and interpretation of results. J.Q. and L.C. contributed to methodological development, results visualization, and validation of the analysis. A.K. and Y.C. supervised the study, provided conceptual guidance, and contributed to the interpretation and revision of the manuscript. All authors reviewed, edited, and approved the final manuscript.\nCorresponding author\nEthics declarations\nCompeting interests\nThe authors declare no competing interests.\nPeer review\nPeer review information\nCommunications Sustainability thanks Rahman Khorramfar and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Nandita Basu. A peer review file is available.\nAdditional information\nPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.\nRights and permissions\nOpen Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.\nAbout this article\nCite this article\nChen, D., Zhou, Z., Cai, Y. et al. Artificial intelligence data centers could reach one percent of global electricity demand by 2030. Commun. Sustain. 1, 147 (2026). https://doi.org/10.1038/s44458-026-00152-5\n- Received: \n- Accepted: \n- Published: \n- Version of record: \n- DOI: https://doi.org/10.1038/s44458-026-00152-5","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//www.nature.com/articles/s44458-026-00152-5","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 62509 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 62509 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":62509,"summary_length":368,"usable_text_length":62509,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":62509,"summary_length":368}},"tags":[]},"fallback_formats":["markdown","json","html"],"actions":{"read":"/item/88444","export_markdown":"/api/items/88444/export?format=markdown","export_json":"/api/items/88444/export?format=json","diagnose":"/api/diagnose?url=https%3A//www.nature.com/articles/s44458-026-00152-5"},"formats":{"full":{"id":88444,"title":"Artificial intelligence data centers could reach one percent of global electricity demand by 2030 - Nature","url":"https://www.nature.com/articles/s44458-026-00152-5","source":"Nature","author":null,"published_at":"2026-09-21T11:04:25+00:00","locale":"en","topic":"ai","tags":[],"excerpt":"Abstract\nThe rapid growth of generative artificial intelligence is increasing global electricity demand and placing new pressure on power systems. Here we show that electricity use by data centers built for artificial intelligence could rise from about 118 terawatt-hours in 2024 to between 239 and 295 terawatt-hours by 2030, or about 1% of global electricity demand.","full_text":"Abstract\nThe rapid growth of generative artificial intelligence is increasing global electricity demand and placing new pressure on power systems. Here we show that electricity use by data centers built for artificial intelligence could rise from about 118 terawatt-hours in 2024 to between 239 and 295 terawatt-hours by 2030, or about 1% of global electricity demand. New computing infrastructure is highly concentrated in North America, Western Europe, and the Asia-Pacific, which together account for more than 90% of projected computing capacity. Some regions, including Oregon, Ireland, and Iowa, face greater pressure from concentrated data-center loads, whereas larger systems such as Texas can absorb new demand more effectively. These results indicate that artificial intelligence infrastructure is becoming a structural part of power-system dynamics. The study combines large language model analysis of corporate, policy, and media sources with scenario-based projections of future electricity demand.\nSimilar content being viewed by others\nIntroduction\nThe rapid emergence of generative artificial intelligence (AI) and large-scale data analytics has driven a sharp expansion of computational demand1,2. As AI models grow exponentially in size and complexity, their training and inference require vast computing power and data throughput, driving record investment in high-performance data centers and digital infrastructure (Figs. S1 and S2)3,4,5,6,7. According to Bain & Company’s Global Technology Report 2025, sustaining the computational requirements of AI expansion could generate nearly US$2 trillion in annual revenues by 2030–equivalent to the combined GDP of the world’s ten largest emerging economies8. This “AI infrastructure boom” is transforming data centers into the industrial backbone of the digital economy and, increasingly, a major source of electricity demand9,10.\nElectricity use has risen sharply in parallel with the digital transition. The International Energy Agency (IEA, 2025) projects that global data-center electricity consumption will more than double—from about 415 TWh in 2024 to roughly 945 TWh by 2030—with the United States and China accounting for almost 80% of the increase11,12. AI-specific facilities rely on GPU-based computation, which enables large-scale parallel processing but consumes up to six times more power than conventional racks, elevating both cooling intensity and peak-load requirements. These facilities are increasingly concentrated in regions with abundant renewable resources, low electricity prices, and favorable climates, yet such clustering also amplifies local grid stress and transmission constraints. As AI campuses scale up from megawatt to gigawatt levels, ensuring a reliable and low-carbon electricity supply has become a strategic challenge for utilities, regulators, and technology developers13,14.\nA growing body of literature has examined the energy footprint of digital infrastructure. Early studies quantified electricity consumption and emissions associated with information and communication technologies15,16,17, while more recent work highlights that the growth of AI workloads—particularly large-scale GPU clusters and generative models—may substantially increase electricity demand and associated environmental impacts18,19,20,21,22.\nTo quantify these impacts, prior research has developed two main classes of modeling approaches. Top-down methods estimate electricity consumption using aggregate indicators such as data traffic, computing capacity, or ICT statistics15,16,17. While these approaches provide useful global benchmarks, they often struggle to capture the heterogeneous and rapidly evolving workloads associated with AI computing. Bottom-up engineering models, in contrast, estimate energy use from facility-level characteristics, including server hardware configurations, cooling systems, and operational efficiency metrics19,23. However, such models are frequently constrained by limited data availability and incomplete knowledge of proprietary hyperscale infrastructure deployments20. Recent studies attempt to bridge these perspectives by linking computational workloads to energy demand through task-based or model-level accounting frameworks20,24,25, complemented by scenario analyses such as the IEA’s Energy and AI report11.\nDespite these advances, existing approaches remain largely focused on aggregate demand estimation or facility-level energy accounting, providing limited insight into how firm-level infrastructure investment, spatial clustering of AI data centers, and regional electricity-system characteristics jointly shape the geography and magnitude of AI-related energy demand.\nThis study addresses this gap by developing an integrated analytical framework that links AI infrastructure deployment, spatial siting patterns, and regional power-system impacts. Unlike conventional approaches that assume stable relationships between computing intensity, hardware efficiency, and electricity use—an assumption that breaks down under rapidly evolving AI workloads26,27,28—our framework employs a large language model (LLM)-based semantic retrieval and inference to dynamically extract firm-level strategic signals from heterogeneous corporate disclosures. These signals are combined with scenario-based electricity-demand projections to assess how AI data-center expansion translates into regional electricity demand and grid pressure. This approach moves beyond aggregate demand projections and facility-level efficiency metrics to capture how firm-level AI deployment strategies and spatial clustering of data centers generate localized electricity-system pressures.\nEmpirically, we focus on six leading technology firms—Amazon, Microsoft, Google, Meta, Oracle, and Apple—which collectively account for ~70−75% of global hyperscale and cloud-linked data-center electricity demand (Table S1). Their global footprint and relatively transparent reporting enable consistent identification of siting patterns and electricity-demand trajectories across regions.\nUnderstanding and managing this emerging compute-energy nexus is critical to ensuring that the rapid expansion of AI infrastructure evolves in tandem with the development of adequate and reliable electricity systems14. In this study, the compute-energy nexus is modeled as a demand-side relationship: we quantify and spatially allocate the electricity consumption generated by AI data-center expansion. Bidirectional interactions—such as demand response, flexible load scheduling, or participation in electricity markets—are recognized as important dimensions of this nexus but fall outside the scope of the current framework, and are identified as directions for future research. Accordingly, this study addresses three central research questions: (i) what spatial factors shape the siting patterns of large-scale AI data centers; (ii) how firm-level infrastructure investments translate into electricity demand trajectories; and (iii) what regional electricity-system pressures arise from the spatial clustering of AI data-center loads.\nAddressing these questions, we find that AI-oriented infrastructure is concentrating in a limited set of regions—chiefly North America, Western Europe, and the Asia-Pacific, which together account for more than 90% of projected compute capacity—and that aggregate electricity consumption by six leading operators is projected to rise from roughly 118 TWh in 2024 to between 239 and 295 TWh by 2030, equivalent to about 1% of projected global electricity demand. This concentration is associated with markedly higher relative electricity-demand pressure in some host regions, such as Oregon, Ireland, and Iowa, than in regions with larger electricity systems, such as Texas. Together, these findings indicate that the geography of AI compute is becoming a structural feature of regional power-system dynamics rather than a peripheral digital load, underscoring the value of anticipatory planning that aligns computational growth with electricity-system development.\nResults\nGlobal clustering and heterogeneity in AI data-center siting\nAI data-center locations are identified using a large language model (LLM)-based inference approach that extracts and classifies site-level information from public disclosures, assigning each candidate site a probability of AI-oriented deployment (see Methods). Reported siting patterns reflect three components: (i) outputs of the LLM/RAG framework, specifically the site-level probabilities and determinant intensity scores derived from semantic retrieval and sentiment analysis; (ii) evidence contained in the underlying source documents—corporate announcements, SEC filings, sustainability reports, and media coverage—from which the LLM extracts signals; and (iii) the authors’ interpretation of these outputs in the context of the broader literature, which synthesizes these signals into regional and firm-level characterizations. Figure 1 illustrates the global spatial evolution and underlying determinants of AI data-center siting, with legacy infrastructure and AI-specific siting layers presented separately in Supplementary Fig. S3.\nLegacy data centers (blue markers) are relatively dispersed, reflecting historical priorities such as latency reduction and proximity to users. In contrast, newly identified AI-specific data centers (color-coded circles) exhibit strong geographic concentration, forming dense clusters in a limited number of regions. These clusters are most pronounced in North America, Europe, and the Asia-Pacific, which together account for over 90% of projected AI compute capacity among leading firms. The dominant corridors align with regions characterized by strong energy availability, mature grid and fiber-optic infrastructure, favorable climatic conditions, and supportive policy environments29,30.\nDistinct regional configurations highlight the heterogeneity of siting determinants. The following characterizations are based on signals extracted by the LLM framework from input documents (corporate filings, sustainability disclosures, and policy sources); the synthesis of these signals into regional interpretations represents the authors’ reading of those extracted patterns. In North America, clusters are concentrated in regions such as Virginia, Texas, Ohio, and North Carolina, which coincide with established hyperscale infrastructure, large electricity markets, and documented policy support for data-center development11,31. These regions have historically attracted cloud infrastructure investment, although underlying conditions such as energy costs, regulatory incentives, and climatic factors vary substantially across locations. In Europe, data-center development is concentrated in the Netherlands, the United Kingdom, and Italy. Renewable-energy availability, regulatory frameworks, and digital infrastructure jointly shape these patterns although their relative importance varies across countries. Nordic regions, in particular, provide favorable conditions due to hydro and wind resources and naturally cool climates30. In the Asia-Pacific region, clusters in Singapore, Taipei, Malaysia, and Japan align with strong digital demand and policy-supported infrastructure expansion. Similarly, emerging sites in the Middle East are associated with state-led investment strategies and land availability.\nFirm-level differentiation further reinforces these spatial dynamics (Fig. 1a, b). To interpret the firm-determinant intensity matrix (Fig. 1b), we classify siting drivers into six dimensions: corporate integration, energy access, policy environment, market demand, infrastructure maturity, and network connectivity. Higher scores indicate that a determinant category appears more frequently and with stronger positive sentiment in retrieved documents associated with a given firm. Among globally scaling firms, Amazon exhibits the broadest geographic diversification (Fig. 1a), expanding across regions with strong market demand and mature infrastructure. Microsoft and Oracle primarily follow policy- and incentive-driven corridors, while Google and Meta anchor their siting strategies in regions with strong energy availability and high network connectivity to reduce latency and operational constraints. By contrast, domestically oriented firms such as Apple adopt a more vertically integrated strategy, concentrating deployments within US regions-particularly those with favorable energy conditions–to optimize operational efficiency and alignment with their broader manufacturing and service ecosystems.\nOverall, these results reveal a structural transition from historically dispersed siting toward geographically concentrated AI infrastructure, with clustering patterns shaped jointly by regional characteristics and firm-level strategies. Energy availability, policy alignment, and network connectivity emerge as dominant determinants of siting decisions, driving the formation of high-density infrastructure corridors.\nElectricity consumption projection: rapid scaling of AI data centers elevates system-level energy demand\nTo capture uncertainty in expansion pace and AI workload adoption, we evaluate three forward scenarios for 2025–2030: a conservative scenario (15% annual growth), a neutral scenario (25%), and an optimistic scenario (35%). These growth rates span the range of recent estimates of data-center and AI-related electricity demand growth reported in the literature and industry forecasts11,31. In all scenarios, baseline data-center electricity consumption grows at 10% annually, while the share of AI-intensive workloads increases over time, reflecting the rising importance of generative-AI computation. Electricity demand is estimated by combining firm-level data-center expansion scenarios with projected AI workload intensity and baseline energy-use trends (see Methods). Figure 2 presents the projected electricity consumption of AI data centers under the three growth scenarios from 2025 to 2030. Figure 2a–f illustrates firm-level trajectories for Amazon, Microsoft, Google, Meta, Oracle, and Apple across conservative, neutral, and optimistic assumptions. The energy demand forecasts for all firms show consistent upward trends, reflecting the intensification of AI workloads and the proliferation of large-model training clusters. Aggregate electricity use among hyperscale operators is projected to rise by ~40−70% over the period, though the pace and scale vary considerably by firm. Amazon, Microsoft, and Google are forecasted to maintain the highest absolute demand, consistent with their extensive cloud-service portfolios and broad geographic reach. Meta follows comparable but more moderate forecast paths, supported by continuous improvements in model efficiency and renewable procurement strategies32,33,34,35,36,37. Apple and Oracle exhibit lower projected energy demand, aligning with their smaller operational footprints and vertically integrated architectures. Across all firms, the optimistic scenario accelerates notably after 2027, coinciding with large-scale deployment of generative-AI infrastructure and training workloads38,39.\nFigure 3 aggregates these firm-level forecasts to depict the global evolution of electricity demand associated with leading AI operators from 2018 to 2030. The solid gray line represents historical estimates derived from industrial disclosures and utility data, while the colored dashed lines correspond to scenario-based projections. Total electricity consumption by these leading firms’ data centers is expected to rise from ~118 TWh in 2024 to between 239 TWh (conservative) and 295 TWh (optimistic) by 2030, implying a compound annual growth rate (CAGR) of roughly 13−17%. This increase is equivalent to adding the electricity demand of a medium-sized national market every two years, or comparable to the annual electricity use of about 27 million US households. The projected magnitude is quantitatively consistent with the International Energy Agency’s (IEA) forecast that total global data center electricity demand will reach around 945 TWh by 2030 (~3% of global power consumption)11, compared with about 1.5% in 2024 (see Methods). These projections collectively illustrate the accelerating energy footprint of AI infrastructure within the broader global power system.\nRegional electricity demand pressure implications of AI data-center clustering\nFigure 4 characterizes the regional distribution of data-center electricity demand and the relative concentration of AI-related loads within regional electricity systems in 2030. Figure 4 displays projected electricity demand by firm and region, revealing a highly uneven spatial distribution of AI-related loads. Fewer than ten regions account for nearly two-thirds of total projected demand, indicating a strong concentration of infrastructure in a limited number of states and countries. Oregon, Virginia, Iowa, Ohio and Ireland emerge as dominant hubs, each exceeding 15 TWh of annual demand under the neutral scenario. These regions are associated with favorable conditions—such as established hyperscale clusters, access to energy resources, and supportive policy environments—that are consistent with large-scale AI deployment. Second-tier regions, including the Texas, Netherlands, and Washington, exhibit moderate but rapidly increasing demand, often driven by firm-specific expansion. Scenario ranges indicate that uncertainty in deployment pace and firm strategy could alter local electricity demand by up to 30%, underscoring the challenges of forecasting and planning for digital-infrastructure growth.\nTo characterize the relative concentration of data-center electricity demand across regions, we construct an electricity demand pressure index (EDPI), defined as the ratio of projected data-center electricity demand to regional total available electricity supply (see Methods). Figure 5a presents the cross-regional evolution of EDPI between 2025 and 2030 under the neutral scenario, revealing a pronounced and spatially uneven increase in demand concentration. A small number of regions—including Oregon, Ireland and Iowa—display markedly elevated EDPI levels in both years, indicating a persistently high concentration of data-center electricity demand relative to available supply. Virginia, Nebraska, and Ohio occupy the upper-middle range of the distribution, reflecting growing demand concentration. The absolute change in EDPI between 2025 and 2030 (ΔEDPI, expressed in percentage points) further reveals a self-reinforcing spatial pattern: regions with already elevated demand concentration tend to experience the largest absolute increases. In contrast, regions with larger electricity systems—such as Texas—exhibit comparatively small ΔEDPI values, reflecting greater available-supply adequacy to absorb incremental AI-related demand (Fig. 5a–c).\nOverall, these results reveal a pronounced spatial asymmetry in the distribution of data-center electricity demand, with a limited number of regions accounting for a disproportionate share of demand relative to their available-electricity adequacy.\nDiscussion\nThis study links AI data-center siting patterns, electricity-demand projections, and regional electricity demand concentration. By combining large language model (LLM)-based infrastructure identification with scenario-based electricity-demand projections, we assess how the spatial expansion of AI computing infrastructure may influence electricity demand and the relative burden on regional electricity systems. A key methodological contribution of this framework is the fusion of qualitative corporate intelligence—extracted by LLMs from financial disclosures, sustainability reports, and strategic announcements—with quantitative electricity demand projection. This combination enables inference of firms’ implicit strategic intent from language patterns in corporate communications, providing a more interpretable and forward-looking basis for demand forecasting than statistical extrapolation from historical consumption trends alone11,20.\nOur results indicate that AI-driven data-center expansion is emerging as an increasingly influential component of global electricity systems7,40. Across firms and regions, the analysis reveals a consistent sequence linking spatial siting patterns, rapidly increasing electricity demand, and rising regional power stress. This pattern suggests that the digital infrastructure supporting AI development is becoming closely intertwined with electricity-system dynamics rather than remaining an exogenous source of demand. As AI-related computing capacity expands and concentrates geographically, it may influence regional load growth, infrastructure investment priorities, and decarbonization pathways41.\nThe clustering of AI data centers in a limited number of favorable regions—primarily in North America, Western Europe, and parts of the Asia-Pacific—creates pronounced spatial asymmetries in load distribution. Regions such as Oregon, Ireland and Iowa—which exhibit elevated EDPI values—face a high relative concentration of data-center electricity demand within their electricity systems, alongside rapid AI infrastructure expansion. In contrast, regions with larger electricity systems, such as Texas, show lower demand concentration relative to total available electricity supply. These patterns highlight how regional electricity-system size, cross-boundary trade flows, and the spatial footprint of AI infrastructure jointly determine the relative exposure of local electricity systems to AI-driven demand growth. The projected growth in electricity demand from AI data centers therefore presents both challenges and opportunities for electricity systems. On one hand, concentrated digital loads could intensify transmission congestion, complicate renewable integration, and increase electricity price volatility if system expansion does not keep pace with demand. On the other hand, large and stable computational loads may provide opportunities for greater integration with electricity systems Data centers can function as anchor customers for long-term renewable power purchase agreements (PPAs), potential participants in flexible demand programs, and focal points for co-located energy storage or hybrid renewable generation42,43. The extent to which these opportunities materialize will depend on the interaction between electricity market design, infrastructure investment, and regulatory frameworks.\nAt the policy level, these findings highlight the potential value of anticipatory frameworks that explicitly incorporate AI-related demand into national energy modeling and long-term planning. Conventional forecasts that treat digital infrastructure as exogenous commercial load risk underestimating the spatial concentration of new demand and its implications for annual energy balancing. Integrating firm-level digital-load projections into planning processes could enable regulators and utilities to better anticipate load growth in high-concentration regions. In emerging markets, policy design may need to balance the benefits of digital-infrastructure development against the risk of exacerbating existing demand-supply imbalances.\nImportantly, the firm-level granularity of this analysis offers a policy-relevant tool that complements aggregate demand forecasts. Because AI-related electricity demand is highly concentrated among a small number of identifiable corporate actors, regulators and utilities can use firm-level demand trajectories to engage with specific operators rather than with a generic “technology sector.” This granularity may support more targeted policy interventions—such as renewable procurement requirements, locational guidance, or mandatory load disclosure—calibrated to the expansion strategies and geographical footprints of individual firms.\nSeveral limitations of the current framework warrant explicit acknowledgment. First, the LLM-based inference relies on corporate sustainability reports, press releases, and annual filings, which may contain optimistic bias regarding efficiency trajectories, renewable integration commitments, or future investment plans. To the extent that such documents overstate actual or planned improvements, the sentiment-based siting probabilities may be biased toward high-profile or low-carbon locations, potentially understating deployment in less prominently disclosed regions. Moreover, regions where AI infrastructure projects are less frequently disclosed (such as Africa or Australia), less systematically reported, or still at an early stage of development may be underrepresented within the inferred deployment database. Second, the efficiency-gain assumption of 1−3% annually—calibrated to historical hyperscale patterns—reflects a central scenario. Further improvements in chip efficiency, including more energy-efficient accelerators or advances in cooling architectures, could reduce energy intensity below projected levels. Third, the EDPI is constructed from aggregate annual energy quantities—available electricity supply (generation plus positive net imports) and projected data-center demand in the numerator—and does not capture intra-regional transmission network constraints or interconnection queue delays. In systems such as the PJM interconnection (a regional transmission organization coordinating wholesale electricity markets and grid operations across parts of the eastern United States, including Virginia), grid integration bottlenecks and interconnection queues can be binding constraints on the ability to absorb new large loads, even when aggregate available supply appears sufficient. The EDPI may therefore underestimate localized electricity-system pressures in transmission-constrained areas, and future extensions of this framework should incorporate network-level indicators. Fourth, the firm-level analysis covers six technology companies that together account for ~70−75% of global hyperscale and cloud-linked data-center electricity demand. The remaining share comprises a heterogeneous set of operators—including smaller cloud-service providers, colocation facility operators, telecommunications companies with data-center divisions, and regional or national operators—that are excluded primarily because their infrastructure footprints and electricity consumption are less consistently disclosed in public filings and their siting patterns are not readily identifiable by the LLM-based inference approach applied here. Excluding this segment may lead to an underestimate of total AI-related electricity demand. Fifth, available electricity projections are based on CAGR estimates from 2019–2024 with a small sample and pandemic impact, so they may not fully capture future changes in electricity systems, including renewable deployment, grid expansion, and infrastructure investments. In particular, in regions with relatively small electricity systems, such developments could substantially increase future electricity availability beyond historical trends. Likewise, future national and regional energy policies, market reforms, and unforeseen extreme events may significantly reshape electricity-supply trajectories.\nMore broadly, the co-evolution of digitalization and electrification marks a conceptual turning point in energy-system governance13. AI infrastructure is transforming electricity demand from a passive outcome of economic activity into an active determinant of system configuration. Its social and economic value increasingly depends on the sustainability of its power source and its integration within low-carbon grids. Future research should quantify the feedback loops between AI-compute trajectories, carbon intensity, and investment flows across the energy-digital interface. Several specific directions are particularly promising. First, the co-optimization of AI data-center siting with other emerging high-load technologies—such as hydrogen electrolyzers—could help minimize transmission congestion and renewable curtailment; strategic spatial allocation of such loads has been shown to reduce boundary reinforcement costs44. Second, the interaction between AI data-center loads and the electrification of transport warrants coordinated analysis, as both introduce large and potentially concurrent demands on distribution and transmission infrastructure; holistic planning frameworks that integrate gigawatt-scale AI campuses alongside electric-vehicle charging have been identified as an emerging priority45. Third, the integration of data centers into urban energy systems through waste heat recovery for building heating offers a dual benefit of improved cooling efficiency and urban decarbonization; coordinated control strategies for building thermal flexibility informed by data-center waste heat could reduce local distribution network congestion46. Fourth, the application of advanced risk assessment and stochastic optimization techniques—analogous to those used in commodity supply-chain management47—could better account for uncertainties in energy pricing, grid availability, and AI compute demand, supporting more robust infrastructure planning under deep uncertainty. Understanding and governing the coupling between computational growth and electricity systems will be central to achieving a resilient, low-carbon digital economy.\nMethods\nThis study develops a hybrid retrieval-augmented generation (RAG) + large language model (LLM) forecasting framework to predict the geographic and operational evolution of hyperscale data centers operated by major technology companies. The framework integrates structured document retrieval, language sentiment analysis, and LLM-based contextual reasoning to infer (i) potential expansion locations, (ii) baseline technical parameters, and (iii) multi-year energy trajectories.\nThe workflow (Fig. 6) consists of six main stages: (1) construction of a RAG knowledge base from multi-year corporate and regional sources, (2) sentiment-aware identification of likely future data center locations, (3) retrieval of location-specific contextual information, (4) extraction of baseline operational parameters for the current year, (5) multi-year forecasting of key technical and environmental metrics using an LLM-guided model and physically interpretable energy equations, and (6) construction of the Electricity Demand Pressure Index (EDPI) that compares projected data center demand against the available regional electricity supply capacity.\nKnowledge-base construction\nA multi-source text corpus was constructed for each firm, including press releases, annual earnings reports, environmental sustainability disclosures, infrastructure investment announcements, government policy documents, and major news articles from 2015 to 2025. All texts were preprocessed and embedded using the Hugging Face all-MiniLM-L6-v2 model, then indexed in a FAISS (Facebook AI similarity search) vector database for semantic retrieval. Each entry retained metadata such as publication date, geographic reference, and document type, enabling targeted retrieval (e.g., “AWS Taiwan 2025 expansion energy infrastructure”).\nLocation Identification via RAG + LLM\nTo infer likely future expansion sites between 2025 and 2030, the framework explicitly prompts the LLM to predict which geographic regions a firm is most likely to expand into. This prediction is based on contextual evidence retrieved from the RAG knowledge base, which includes firm sustainability reports, capital expenditure statements, and government or media coverage.\nThe RAG system is first queried using firm-specific and temporal prompts, such as {firm} 2025–2030 data center investment or construction plans. The retrieved documents contain both qualitative statements (e.g., “strong regional demand growth,” “strategic infrastructure partnership”) and quantitative indicators (e.g., “$2B allocated to APAC expansion”).\nEach document is analyzed by the LLM not only for content relevance but also for its sentiment and linguistic tone. Positive sentiment toward investment, expansion, or infrastructure development—particularly in earnings calls or annual reports—acts as an implicit prior that increases the model’s belief in a region’s likelihood of receiving new data center capacity. Conversely, negative sentiment (e.g., mentions of cost control, divestment, or policy uncertainty) decreases the associated probability weight.\nThe LLM then synthesizes these cues into a structured probability distribution over candidate locations, producing outputs of the form: P(expansion at site i ∣evidence) = fLLM(retrieved text embeddings, sentiment scores).\nThis probabilistic representation is used in later stages to weight forecasts of capacity growth and energy consumption at each site.\nBaseline parameter extraction\nFor each confirmed or high-probability site, the system used a RAG-conditioned LLM prompt to extract current-year baseline parameters describing the data center’s IT and facility operations. The baseline operational parameters used in the forecasting framework are summarized in Table 1, including accelerator counts, average power draw, utilization factors, operating hours, and power usage effectiveness (PUE).\nThe model was instructed to reason explicitly about local grid mix, cooling efficiency, and infrastructure maturity, and to avoid copying assumptions between regions. The extracted parameters included: {Ntrain, Ninference, Pavg,train, Pavg,inference, utrain, uinference, Htrain, Hinference, PUE, g}\nAll outputs were formatted as structured JSON (JavaScript Object Notation) objects, accompanied by short notes summarizing assumptions and references. Each variable corresponds to a measurable operational attribute of the local compute fleet.\nForecasting future parameters\nA second RAG + LLM chain projected the time evolution of these parameters over a 5-year horizon. The prompt incorporated both historical operational trends and retrieved regional information, constraining LLM reasoning with empirically observed hyperscale patterns: 10−20% annual accelerator capacity growth, 1−3% annual efficiency gains, and gradual improvements in utilization. These parameters characterize the supply-side evolution of infrastructure, specifically capturing the growth in installed accelerator capacity and incremental improvements in energy efficiency at the facility level. The model generated structured forecasts for each year t ∈ [2025, 2030], each annotated with explanatory notes and relevant document sources.\nEnergy computation\nForecasted operational parameters were post-processed in Python to derive the physical energy demand of each site. The IT load energy (EIT) represents the direct electrical consumption of computing equipment, while the estimated AI data center energy demand (EDC) includes both IT and non-IT overheads such as cooling and power distribution losses.\nHere, N, Pavg, u, and H denote the number of accelerators, average power draw (kW), utilization factor, and annual operating hours, respectively. EIT is thus measured in MWh, and EDC scales this load by the facility’s power usage effectiveness (PUE), capturing site-specific cooling and infrastructure efficiency.\nImplementation\nThe full pipeline was implemented in Python using the LangChain framework for composable RAG pipelines, FAISS for vector retrieval, and the OpenAI GPT-4o-mini model as the reasoning component (temperature = 0.3). All modules were orchestrated via reproducible runnables with strict JSON parsing for downstream numerical analysis.\nSentiment-aware expansion probability modeling\nThe sentiment analysis stage produced, for each candidate region i, an expansion likelihood score Pi defined as:\nwhere Si is the normalized sentiment score for region i (ranging from −1 to +1), and Ri is the retrieval relevance score (cosine similarity in the embedding space). These probabilities informed the sampling or prioritization of forecast targets in subsequent steps. Regions frequently mentioned in corporate reports and described with positive financial tone thus received elevated probabilities of selection.\nThis coupling of semantic retrieval and sentiment tone weighting enables the model to capture implicit strategic intent in corporate communications—an important predictor of near-term data center expansion behavior.\nBy fusing semantic retrieval, sentiment weighting, and physical energy modeling, this hybrid framework bridges qualitative corporate signals and quantitative operational forecasting. The approach captures how firms’ language choices in financial and environmental disclosures correlate with real-world infrastructure trajectories, enabling a more interpretable and evidence-grounded forecast of global data center development.\nScenario design and aggregation\nTo capture uncertainty in both firm-level expansion intensity and technological efficiency, we constructed three forward scenarios s-conservative, neutral, and optimistic-for 2025–2030. Each scenario jointly controls (i) the rate of new AI data-center additions and (ii) the efficiency gains of all operating sites. The annual growth rate of new AI load is set to gnew,s ∈ {0.15, 0.25, 0.35} for scenario s among the three scenarios, while existing stock consumption for each firm grows at gstock = 0.10. Specifically, the three forward scenarios for AI-driven data-center expansion (15, 25, and 35% annual growth) are designed to span conservative, central, and high-growth trajectories of AI workload expansion. These values are informed by recent industry and research evidence on the growth of AI compute demand and infrastructure. AI training workloads are estimated to grow at around 22% annually, while inference workloads may expand at rates approaching 35% over the next five years48. Other estimates indicate that overall AI-related infrastructure demand may grow in the range of 20−25% annually in baseline scenarios49. Based on this range of evidence, the 15% scenario reflects a lower-bound trajectory consistent with moderate expansion in AI deployment, the 25% scenario captures the central tendency of observed infrastructure growth rates, and the 35% scenario represents an upper-tail case aligned with rapid scaling of inference workloads and AI-ready capacity. These scenarios are designed to capture the empirically observed range of AI compute and infrastructure growth, while also reflecting uncertainty in future AI adoption intensity and technological scaling.\nFirm-specific AI shares pAI(f, t) increase gradually over time to reflect the rising share of AI workloads in total compute demand. Firm-specific AI shares pAI(f, t) were defined to represent the evolving proportion of AI-related compute workloads within each operator’s total data-center activity. For each firm f, pAI(f, t) increases over time according to a discrete schedule calibrated to reflect the firm’s relative intensity of AI adoption between 2025 and 2030:\nwhere p1(f), p2(f), and p3(f) correspond to the firm’s baseline (2025), mid-phase (2026–2027), and mature-phase (2028–2030) AI workload shares, respectively. Formally, this can be expressed as:\nThe parameter set {p1(f), p2(f), p3(f)} is firm-specific, capturing heterogeneity in AI adoption trajectories across major hyperscale operators:\nThis stepwise formulation approximates each firm’s increasing allocation of computational resources to AI model training and inference. The heterogeneity in pAI(f, t) reflects differences in investment magnitude, infrastructure build-out, and strategic focus among major operators. Firms with large-scale and early commitments to generative AI—such as Amazon, Google, Meta, and Microsoft—exhibit steeper trajectories, consistent with multibillion-dollar data-center expansions, dedicated AI accelerator deployment (e.g., Trainium, H100, and TPUv5), and vertically integrated AI ecosystems. By contrast, vertically integrated or enterprise-focused operators such as Apple and Oracle show slower growth, consistent with smaller-scale AI infrastructure investment, reliance on consumer-device optimization rather than large-model training, and a later pivot toward cloud-based AI services. Accordingly, the evolution of pAI(f, t) captures each firm’s position along the spectrum of AI-driven digital-infrastructure transformation, serving as a proxy for the relative share of compute capacity devoted to AI workloads within their overall fleet.\nFor each firm f, the existing stock electricity consumption evolves as\nNew-site electricity demand follows compounding growth of AI workloads and the firm-specific AI share schedule:\nThe total firm-level electricity demand is then\nTo represent uncertainty arising from multiple independent siting pathways, we simulate a set of Pt,s plausible firm-region configurations for each year t and scenario s. Global electricity demand in year t and scenario s is the ensemble mean:\nand the range\ndefines the shaded uncertainty bands in the scenario plots, where \\({E}_{t,s,p}^{{{\\rm{glob}}}}\\) denotes the global total electricity demand in year t under scenario s for projection path p ∈ {1, …, Pt,s}. The 2024 observed total serves as the common anchor point for all trajectories.\nRegional allocation of energy\nFirm-level electricity demand is spatially distributed according to two complementary weights: (i) AI siting probabilities derived from the LLM siting model, and (ii) historical stock weights from the existing data-center inventory.\nFor AI-related new sites, firm-year location weights are proportional to modeled AI energy at each location ℓ:\nHere \\({E}_{f,t,\\ell }^{AI,\\,{\\mbox{loc}}\\,}\\) denotes the modeled AI-related electricity demand assigned to firm f at location ℓ in year t, as determined from the LLM-based siting model and corresponding regional AI workload intensity.\nFor legacy and non-AI sites, we use historical weights based on observed facility counts. For legacy and non-AI sites, we use historical weights based on observed facility counts:\nLegacy (historical) data-center counts refer to the number of distinct operating campuses under each firm’s control as of end-2024, including both owned and long-term leased facilities. These are derived from the observed 2015–2024 dataset and serve as proxies for baseline spatial capacity prior to the AI expansion. Since 2024, the global data-center landscape has entered a rapid expansion phase driven by generative-AI workloads, with AI infrastructure investment reaching a record $57 billion globally50. This “AI data-center boom” marks a structural inflection in digital-infrastructure growth.\nLet r(ℓ) map locations to regions. Regional electricity use for firm f and scenario s is:\nSumming over all firms yields regional totals:\nThis approach preserves firm-level energy consistency while ensuring that AI-driven growth concentrates in high-probability siting regions identified by the LLM model, and that legacy loads follow empirically observed footprints.\nCross-validation\nAggregate electricity consumption by the six leading hyperscale operators is projected to rise from ~118 TWh in 2024 to 239–295 TWh by 2030. This projection is broadly consistent with the International Energy Agency’s (IEA) estimate that global data-center electricity demand will reach roughly 945 TWh by 2030 (IEA, 2025). To benchmark the modeled values, we derive an implied 2030 consumption level for the six leading firms based on the IEA’s global forecast. Assuming (i) that hyperscale data centers account for 70% of global data-center electricity use (Synergy Research Group, 2024), and (ii) that the top six operators collectively represent 40% of hyperscale activity, the corresponding implied electricity consumption is:\nThis cross-validation demonstrates that the projected range of 239–295 TWh lies well within the values implied by independent international forecasts. The close correspondence reinforces confidence in the representativeness of the modeled global totals and supports the credibility of the upper-bound scenario in light of established external benchmarks. Overall, the modeled range remains broadly aligned with independent global projections and reflects plausible hyperscale market dynamics.\nElectricity demand pressure index (EDPI)\nTo assess the relative demand burden associated with data-center expansion, we construct an electricity demand pressure index (EDPI), defined as the ratio of annual data-center electricity demand to the regional available electricity supply—comprising in-region generation plus positive net electricity imports. For each region r and year t:\nwhere \\({E}_{r,t}^{{\\mathrm{Avail}}} = {E}_{r,t}^{{\\mathrm{Gen}}} + {\\max} \\left(0, {{E}^{{\\mathrm{NetImport}}}_{r,t}}\\right), {{E}^{{\\mathrm{DC}}}_{r,t}}\\) denotes total annual data-center electricity demand (in TWh), with \\({E}_{r,t}^{Gen}\\) denoting regional electricity generation (TWh), \\({E}_{r,t}^{{\\mathrm{Gen}}}\\) denoting net electricity imports, and \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) denoting total available electricity supply (TWh). Positive net imports are included because imported electricity contributes directly to regional electricity supply. For net-exporting regions, historical electricity exports are treated as adjustable rather than permanently fixed. Accordingly, negative net imports are set to zero in the denominator, reflecting the assumption that part of the electricity historically exported could instead be retained to serve additional local electricity demand associated with AI data-center expansion. This approximation avoids overstating electricity demand pressure in regions that are persistent net electricity exporters while providing a consistent basis for cross-regional comparison.\nFor US states, \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) is derived from EIA state electricity profiles as total retail electricity sales, which by the energy balance identity equals in-state net generation plus net interstate electricity receipts. For countries, \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) is the sum of total electricity generation and positive net electricity imports, both drawn from the Ember Yearly Full Release dataset (2019–2024). The EDPI is interpreted as a demand-share indicator capturing the relative contribution of data-center loads to regional electricity supply.\nRegional available electricity supply is projected using historical trends of available electricity as a baseline approximation. Specifically, available electricity is extrapolated from the period 2019–2024 using a region-specific compound annual growth rate (CAGR), consistent with the five-year window used in the CAGR formula below.\nRegional available electricity baselines\n(\\({E}_{r,2024}^{{\\mathrm{Avail}}}\\)) were obtained from EIA State Electricity Profiles (US states) and the Ember Yearly Full Release dataset (countries), covering 2019–2024, expressed in terawatt-hours (TWh). Available electricity data were extrapolated to 2030 using a compound annual growth rate (CAGR):\nwith \\({{\\mbox{CAGR}}}_{avail,r}={\\left(\\frac{{E}_{r,2024}^{Avail}}{{E}_{r,2019}^{Avail}}\\right)}^{1/5}-1\\).\nRegional electricity available supply is projected using region-specific compound annual growth rates (CAGR) estimated from historical observations during 2019–2024. This approach assumes that recent historical trends provide a reasonable baseline approximation of future electricity-system evolution. The resulting projections should therefore be interpreted as baseline-trend estimates rather than policy-conditioned forecasts.\nData availability\nThe publicly available electricity data used in this study were obtained from the Ember Yearly Electricity Data Explorer (https://ember-energy.org/data/yearly-electricity-data/) and the US Energy Information Administration State Electricity Profiles (https://www.eia.gov/electricity/state/). The International Energy Agency’s Energy and AI report (https://www.iea.org/reports/energy-and-ai) was used for external benchmarking. Additional public corporate and contextual information was obtained from SEC EDGAR filings (https://www.sec.gov/edgar/search/), corporate annual and sustainability reports, press releases, government and grid-operator publications, and media reports. Historical facility-level data for Amazon, Microsoft, Google, Meta, Oracle, and Apple were obtained from S&P Capital IQ (https://www.capitaliq.com/). The processed data and model outputs underlying the data-center siting estimates, electricity-demand projections, electricity demand pressure index calculations, figures, and tables are publicly available at https://github.com/dchen219/ai-data-center-electricity.\nCode availability\nThe custom code used for document processing, LLM-based data-center siting inference, electricity-demand projection, electricity demand pressure index calculation, and figure generation is publicly available at https://github.com/dchen219/ai-data-center-electricity. Analyses were conducted in Python using LangChain, FAISS, the Hugging Face all-MiniLM-L6-v2 embedding model, and OpenAI’s GPT-4o-mini model.\nReferences\n- Stylianou, N. et al. Inside the relentless race for AI capacity. Financial Times. https://ig.ft.com/ai-data-centres/ (2025). \n- Storey, V. C., Yue, W. T., Zhao, J. L. & L., R. Generative artificial intelligence: evolving technology, growing societal impact, and opportunities for information systems research. Inform. Syst. Front. https://doi.org/10.1007/s10796-025-10581-7 (2025). \n- Cho, J. & Moon, J. H. Numerical coupling of energy efficiency and thermal performance for cold plate cooling optimization in high-density compute AI data centers. Energy Build. 348, 116441 (2025). \n- Pilz, K. F., Mahmood, Y. & Heim, L. AI’s Power Requirements under Exponential Growth: Extrapolating ai Data Center Power Demand and Assessing its Potential Impact on U.S. Competitiveness. Tech. Rep. https://www.rand.org/pubs/research_reports/RRA3572-1.html (RAND Corporation, 2025). \n- Oxford Institute for Energy Studies. Artificial Intelligence and its Implications for Electricity Systems. Tech. Rep. 145. https://www.oxfordenergy.org/wpcms/wp-content/uploads/2025/05/OEF-145.pdf (Oxford Institute for Energy Studies, 2025)) \n- Chen, X., Wang, X., Colacelli, A., Lee, M. & Xie, L. Electricity demand and grid impacts of AI data centers: Challenges and prospects. Nexus 3, 100162 https://doi.org/10.1016/j.ynexs.2026.100162 (2026) \n- Stojkovic, J., Zhang, C., Goiri, Íñigo., Torrellas, J. & Choukse, E. DynamoLLM: designing LLM inference clusters for performance and energy efficiency. In 2025 IEEE International Symposium on High Performance Computer Architecture (HPCA) 1348–1362 (2025). \n- Crawford, D., Schallehn, M., Renno, P., Hanbury, P. & Cannarsi, A. How can We meet AI’s Insatiable Demand for Compute Power? https://www.bain.com/insights/how-can-we-meet-ais-insatiable-demand-for-compute-power-technology-report-2025/ (Bain & Company, 2025). \n- Ngata, W. M. et al. The cloud next door: Investigating the environmental and socioeconomic strain of datacenters on local communities. In Proc. 2025 ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies, COMPASS ’25 769–774 (Association for Computing Machinery, 2025). \n- Li, P., Yang, J., Islam, M. A. & Ren, S. Making AI less ‘Thirsty’. Commun. ACM 68, 54–61 (2025). \n- International Energy Agency. Energy and AI. https://www.iea.org/reports/energy-and-ai (2025). \n- Bogmans, C. et al. Power hungry: how AI will drive energy demand. IMF Working Paper 2025/081, International Monetary Fund (2025). \n- Wu, C.-J. et al. Sustainable AI: environmental implications, challenges, and opportunities. In Proc. Machine Learning and Systems 795–813. https://proceedings.mlsys.org/paper_files/paper/2022/hash/462211f67c7d858f663355eff93b745e-Abstract.html (2022). \n- Chen, Z.-M. et al. AI carbon footprint in China sets to double post-2030 carbon peaking. Energy Econ. 150, 108880 (2025). \n- Belkhir, L. & Elmeligi, A. Assessing ICT global emissions footprint: trends to 2040 and recommendations. J. Clean. Prod. 177, 448–463 (2018). \n- Masanet, E., Shehabi, A., Lei, N., Smith, S. & Koomey, J. Recalibrating global data center energy-use estimates. Science 367, 984–986 (2020). \n- Malmodin, J., Lövehagen, N., Bergmark, P. & Lundén, D. ICT sector electricity consumption and greenhouse gas emissions–2020 outcome. Telecommun. Policy 48, 102701 (2024). \n- Debus, C., Piraud, M., Streit, A., Theis, F. & Götz, M. Reporting electricity consumption is essential for sustainable AI. Nat. Mach. Intell. 5, 1176–1178 (2023). \n- Cao, Z. et al. Data center sustainability: revisits and outlooks. IEEE Trans. Sustain. Comput. 9, 236–248 (2024). \n- Xiao, T., Nerini, F. F., Matthews, H. D., Tavoni, M. & You, F. Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA. Nat. Sustain. 8, 1541–1553 (2025). \n- Zhang, Y., Tang, H., Li, H. & Wang, S. Integration and interaction of next-generation AI-focused data centers with smart grids and district energy systems: the state-of-the-art, opportunities and challenges. Renew. Sustain. Energy Rev. 224, 116097 (2025). \n- Han, J. et al. Data-driven distributionally robust optimization of low-carbon data center energy systems considering multi-task response and renewable energy uncertainty. J. Build. Eng. 102, 111937 (2025). \n- Nøland, J. K., Hjelmeland, M. & Korpås, M. Will energy-hungry AI create a baseload power demand boom? IEEE Access 12, 110353–110360 (2024). \n- Luccioni, A. S., Viguier, S. & Ligozat, A.-L. Estimating the carbon footprint of BLOOM, a 176B parameter language model. J. Mach. Learn. Res. 24, 1–15 (2023). \n- Tomlinson, B., Black, R. W., Patterson, D. J. & Torrance, A. W. The carbon emissions of writing and illustrating are lower for AI than for humans. Sci. Rep. 14, 3732 (2024). \n- Ahmed, K. M. U., Bollen, M. H. J. & Alvarez, M. A review of data centers energy consumption and reliability modeling. IEEE Access 9, 152536–152563 (2021). \n- Liu, Y. et al. Energy consumption and emission mitigation prediction based on data center traffic and PUE for global data centers. Glob. Energy Interconnect. 3, 272–282 https://doi.org/10.1016/j.gloei.2020.07.008 (2020). \n- Siddik, M. A. B., Shehabi, A. & Marston, L. The environmental footprint of data centers in the United States. Environ. Res. Lett. https://doi.org/10.1088/1748-9326/abfba1(2021). \n- Cao, Z., Zhou, X., Hu, H., Wang, Z. & Wen, Y. Toward a systematic survey for carbon neutral data centers. IEEE Commun. Surv. Tutor. 24, 895–936 https://doi.org/10.1109/COMST.2022.3161275 (2022). \n- Hyvönen, J. et al. Potential of solar photovoltaics and waste heat utilization in cold climate data centers: case study: Finland and northern Japan. Renew. Sustain. Energy Rev. https://doi.org/10.1016/j.rser.2024.114619 (2024). \n- Shehabi, A. et al. 2024 United States data center energy usage report. Tech. Rep. LBNL-2001637. https://doi.org/10.71468/P1WC7Q (Lawrence Berkeley National Laboratory, 2024). \n- Synergy Research Group. Hyperscale data center count hits 1136; average size increases; US accounts for 54% of total capacity. https://www.srgresearch.com/articles/hyperscale-data-center-count-hits-1136-average-size-increases-us-accounts-for-54-of-total-capacity (2025). \n- Donnellan, D. et al. Uptime Institute global data center survey results 2024. Tech. Rep. https://uptimeinstitute.com/resources/research-and-reports/uptime-institute-global-data-center-survey-results-2024 (Uptime Institute, 2024). \n- Khan, O. Scale your AI transformation with a powerful, secure, and adaptive cloud infrastructure. Microsoft Azure Blog. https://azure.microsoft.com/en-us/blog/scale-your-ai-transformation-with-a-powerful-secure-and-adaptive-cloud-infrastructure/ (2024). \n- Google Sustainability. 2023 Environmental report. https://sustainability.google/reports/google-2023-environmental-report/ (2023) \n- Meta Platforms, Inc. Annual report on Form 10-K for the fiscal year ended december 31, 2024. Tech. Rep. SEC Accession No. 0001326801-25-000017, U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/1326801/000132680125000017/meta-20241231.htm (2025). \n- Sivasubramanian, S. Welcome to a new era of building in the cloud with generative AI on AWS. AWS Machine Learning Blog. https://aws.amazon.com/blogs/machine-learning/welcome-to-a-new-era-of-building-in-the-cloud-with-generative-ai-on-aws/ (2023). \n- Apple Inc. Annual report on Form 10-K for the fiscal year ended september 28, 2024. U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/320193/000032019324000123/aapl-20240928.htm (2024). \n- Oracle Corporation. Oracle offers first zettascale cloud computing cluster. Oracle News press release. https://www.oracle.com/news/announcement/ocw24-oracle-offers-first-zettascale-cloud-computing-cluster-2024-09-11/ (2024). \n- Lambert, S., Caron, E., Lefèvre, L. & Grivel, R. Consolidation of virtual machines to reduce energy consumption of data centers by using ballooning, sharing and swapping mechanisms. Future Gener. Comput. Syst. 174, 107968 https://doi.org/10.1016/j.future.2025.107968 (2026). \n- Bolón-Canedo, V., Morán-Fernández, L., Cancela, B. & Alonso-Betanzos, A. A review of green artificial intelligence: towards a more sustainable future. Neurocomputing 599, 128096 https://doi.org/10.1016/j.neucom.2024.128096 (2024). \n- de Roucy-Rochegonde, L. & Buffard, A. AI, Data Centers and Energy Demand: Reassessing and Exploring the Trends. Tech. Rep. https://www.ifri.org/en/papers/ai-data-centers-and-energy-demand-reassessing-and-exploring-trends-0 (Institut français des relations internationales (Ifri), 2025). \n- Goldman Sachs Research. GS SUSTAIN: Generational Growth: AI, Data Centers and the Coming US Power Demand Surge. Tech. Rep. https://www.goldmansachs.com/pdfs/insights/pages/generational-growth-ai-data-centers-and-the-coming-us-power-surge/report.pdf (Goldman Sachs, 2024). \n- Giannelos, S., Konstantelos, I., Pudjianto, D. & Strbac, G. The impact of electrolyser allocation on Great Britain’s electricity transmission system in 2050. Int. J. Hydrog. Energy 202, 153097 https://doi.org/10.1016/j.ijhydene.2025.153097 (2026). \n- Amann, G. et al. E-mobility Deployment and Impact on Grids: Impact of EV and Charging Infrastructure on European T&D Grids: Innovation needs. Technical Report MJ-09-22-246-EN-N (Publications Office of the European Union, 2022). \n- Dong, Z., Zhang, X., Zhang, L., Giannelos, S. & Strbac, G. Flexibility enhancement of urban energy systems through coordinated space heating aggregation of numerous buildings. Appl. Energy 374, 123971 (2024). \n- Giannelos, S., Konstantelos, I. & Strbac, G. Optimal supply chain design using machine learning, risk assessment and optimisation applied to coal distribution. EURO J. Decis. Process. 13, 100062 (2025). \n- Arora, C., Sorel, M. & Sachdeva, P. The next big shifts in AI workloads and hyperscaler strategies. McKinsey & Company. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-next-big-shifts-in-ai-workloads-and-hyperscaler-strategies (2025) \n- Barth, A., Arora, C., Shenai, G., Noffsinger, J. & Sachdeva, P. The data center balance: how US states can navigate the opportunities and challenges. McKinsey & Company. https://www.mckinsey.com/industries/public-sector/our-insights/the-data-center-balance-how-us-states-can-navigate-the-opportunities-and-challenges (2025). \n- Saavedra, R. & Seaward, S. 2025 Data center marketplace: balancing unprecedented opportunity with strategic risk. U.S. Research Report, Colliers. https://www.colliers.com/en/research/nrep-usdc-data-center-marketplace-2025 (2025). \nAcknowledgements\nNot applicable.\nFunding\nAuthors declare no relevant funding.\nAuthor information\nAuthors and Affiliations\nContributions\nD.C. and Y.C. conceived the study. D.C. designed the analytical framework, collected and processed the data, implemented the LLM/RAG-based analysis, conducted the electricity-demand projections and EDPI analysis, prepared the figures and tables, and drafted the manuscript. Z.Z. contributed to data collection, data validation, and interpretation of results. J.Q. and L.C. contributed to methodological development, results visualization, and validation of the analysis. A.K. and Y.C. supervised the study, provided conceptual guidance, and contributed to the interpretation and revision of the manuscript. All authors reviewed, edited, and approved the final manuscript.\nCorresponding author\nEthics declarations\nCompeting interests\nThe authors declare no competing interests.\nPeer review\nPeer review information\nCommunications Sustainability thanks Rahman Khorramfar and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Nandita Basu. A peer review file is available.\nAdditional information\nPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.\nRights and permissions\nOpen Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.\nAbout this article\nCite this article\nChen, D., Zhou, Z., Cai, Y. et al. Artificial intelligence data centers could reach one percent of global electricity demand by 2030. Commun. Sustain. 1, 147 (2026). https://doi.org/10.1038/s44458-026-00152-5\n- Received: \n- Accepted: \n- Published: \n- Version of record: \n- DOI: https://doi.org/10.1038/s44458-026-00152-5","reading_time_min":40,"extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 62509 characters.","diagnostics_url":"/api/diagnose?url=https%3A//www.nature.com/articles/s44458-026-00152-5","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 62509 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":62509,"summary_length":368,"usable_text_length":62509,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":62509,"summary_length":368}}},"quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 62509 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":62509,"summary_length":368,"usable_text_length":62509,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":62509,"summary_length":368}},"actions":{"read":"/item/88444","export_markdown":"/api/items/88444/export?format=markdown","export_json":"/api/items/88444/export?format=json","diagnose":"/api/diagnose?url=https%3A//www.nature.com/articles/s44458-026-00152-5"}},"digest":{"id":88444,"title":"Artificial intelligence data centers could reach one percent of global electricity demand by 2030 - Nature","url":"https://www.nature.com/articles/s44458-026-00152-5","source":"Nature","topic":"ai","published_at":"2026-09-21T11:04:25+00:00","excerpt":"Abstract The rapid growth of generative artificial intelligence is increasing global electricity demand and placing new pressure on power systems. Here we show that electricity use by data centers built for artificial intelligence could rise from about 118 terawatt-hours in 2024…","quality_bucket":"high","quality_reason":"High confidence: full text extraction produced 62509 characters.","reading_time_min":40,"cluster_id":null},"card":{"display_title":"Artificial intelligence data centers could reach one percent of global electricity demand by 2030 - Nature","subtitle":"Nature · 2026-09-21","summary":"Abstract The rapid growth of generative artificial intelligence is increasing global electricity demand and placing new pressure on power systems. Here we show that electricity use by data centers built for artificial…","badges":["quality:high"],"links":{"read":"/item/88444","original":"https://www.nature.com/articles/s44458-026-00152-5","diagnose":"/api/diagnose?url=https%3A//www.nature.com/articles/s44458-026-00152-5"},"quality_warning":null},"export":{"title":"Artificial intelligence data centers could reach one percent of global electricity demand by 2030 - Nature","url":"https://www.nature.com/articles/s44458-026-00152-5","summary":"Abstract\nThe rapid growth of generative artificial intelligence is increasing global electricity demand and placing new pressure on power systems. Here we show that electricity use by data centers built for artificial intelligence could rise from about 118 terawatt-hours in 2024 to between 239 and 295 terawatt-hours by 2030, or about 1% of global electricity demand.","source":"Nature","date":"2026-09-21T11:04:25+00:00","content":"Abstract\nThe rapid growth of generative artificial intelligence is increasing global electricity demand and placing new pressure on power systems. Here we show that electricity use by data centers built for artificial intelligence could rise from about 118 terawatt-hours in 2024 to between 239 and 295 terawatt-hours by 2030, or about 1% of global electricity demand. New computing infrastructure is highly concentrated in North America, Western Europe, and the Asia-Pacific, which together account for more than 90% of projected computing capacity. Some regions, including Oregon, Ireland, and Iowa, face greater pressure from concentrated data-center loads, whereas larger systems such as Texas can absorb new demand more effectively. These results indicate that artificial intelligence infrastructure is becoming a structural part of power-system dynamics. The study combines large language model analysis of corporate, policy, and media sources with scenario-based projections of future electricity demand.\nSimilar content being viewed by others\nIntroduction\nThe rapid emergence of generative artificial intelligence (AI) and large-scale data analytics has driven a sharp expansion of computational demand1,2. As AI models grow exponentially in size and complexity, their training and inference require vast computing power and data throughput, driving record investment in high-performance data centers and digital infrastructure (Figs. S1 and S2)3,4,5,6,7. According to Bain & Company’s Global Technology Report 2025, sustaining the computational requirements of AI expansion could generate nearly US$2 trillion in annual revenues by 2030–equivalent to the combined GDP of the world’s ten largest emerging economies8. This “AI infrastructure boom” is transforming data centers into the industrial backbone of the digital economy and, increasingly, a major source of electricity demand9,10.\nElectricity use has risen sharply in parallel with the digital transition. The International Energy Agency (IEA, 2025) projects that global data-center electricity consumption will more than double—from about 415 TWh in 2024 to roughly 945 TWh by 2030—with the United States and China accounting for almost 80% of the increase11,12. AI-specific facilities rely on GPU-based computation, which enables large-scale parallel processing but consumes up to six times more power than conventional racks, elevating both cooling intensity and peak-load requirements. These facilities are increasingly concentrated in regions with abundant renewable resources, low electricity prices, and favorable climates, yet such clustering also amplifies local grid stress and transmission constraints. As AI campuses scale up from megawatt to gigawatt levels, ensuring a reliable and low-carbon electricity supply has become a strategic challenge for utilities, regulators, and technology developers13,14.\nA growing body of literature has examined the energy footprint of digital infrastructure. Early studies quantified electricity consumption and emissions associated with information and communication technologies15,16,17, while more recent work highlights that the growth of AI workloads—particularly large-scale GPU clusters and generative models—may substantially increase electricity demand and associated environmental impacts18,19,20,21,22.\nTo quantify these impacts, prior research has developed two main classes of modeling approaches. Top-down methods estimate electricity consumption using aggregate indicators such as data traffic, computing capacity, or ICT statistics15,16,17. While these approaches provide useful global benchmarks, they often struggle to capture the heterogeneous and rapidly evolving workloads associated with AI computing. Bottom-up engineering models, in contrast, estimate energy use from facility-level characteristics, including server hardware configurations, cooling systems, and operational efficiency metrics19,23. However, such models are frequently constrained by limited data availability and incomplete knowledge of proprietary hyperscale infrastructure deployments20. Recent studies attempt to bridge these perspectives by linking computational workloads to energy demand through task-based or model-level accounting frameworks20,24,25, complemented by scenario analyses such as the IEA’s Energy and AI report11.\nDespite these advances, existing approaches remain largely focused on aggregate demand estimation or facility-level energy accounting, providing limited insight into how firm-level infrastructure investment, spatial clustering of AI data centers, and regional electricity-system characteristics jointly shape the geography and magnitude of AI-related energy demand.\nThis study addresses this gap by developing an integrated analytical framework that links AI infrastructure deployment, spatial siting patterns, and regional power-system impacts. Unlike conventional approaches that assume stable relationships between computing intensity, hardware efficiency, and electricity use—an assumption that breaks down under rapidly evolving AI workloads26,27,28—our framework employs a large language model (LLM)-based semantic retrieval and inference to dynamically extract firm-level strategic signals from heterogeneous corporate disclosures. These signals are combined with scenario-based electricity-demand projections to assess how AI data-center expansion translates into regional electricity demand and grid pressure. This approach moves beyond aggregate demand projections and facility-level efficiency metrics to capture how firm-level AI deployment strategies and spatial clustering of data centers generate localized electricity-system pressures.\nEmpirically, we focus on six leading technology firms—Amazon, Microsoft, Google, Meta, Oracle, and Apple—which collectively account for ~70−75% of global hyperscale and cloud-linked data-center electricity demand (Table S1). Their global footprint and relatively transparent reporting enable consistent identification of siting patterns and electricity-demand trajectories across regions.\nUnderstanding and managing this emerging compute-energy nexus is critical to ensuring that the rapid expansion of AI infrastructure evolves in tandem with the development of adequate and reliable electricity systems14. In this study, the compute-energy nexus is modeled as a demand-side relationship: we quantify and spatially allocate the electricity consumption generated by AI data-center expansion. Bidirectional interactions—such as demand response, flexible load scheduling, or participation in electricity markets—are recognized as important dimensions of this nexus but fall outside the scope of the current framework, and are identified as directions for future research. Accordingly, this study addresses three central research questions: (i) what spatial factors shape the siting patterns of large-scale AI data centers; (ii) how firm-level infrastructure investments translate into electricity demand trajectories; and (iii) what regional electricity-system pressures arise from the spatial clustering of AI data-center loads.\nAddressing these questions, we find that AI-oriented infrastructure is concentrating in a limited set of regions—chiefly North America, Western Europe, and the Asia-Pacific, which together account for more than 90% of projected compute capacity—and that aggregate electricity consumption by six leading operators is projected to rise from roughly 118 TWh in 2024 to between 239 and 295 TWh by 2030, equivalent to about 1% of projected global electricity demand. This concentration is associated with markedly higher relative electricity-demand pressure in some host regions, such as Oregon, Ireland, and Iowa, than in regions with larger electricity systems, such as Texas. Together, these findings indicate that the geography of AI compute is becoming a structural feature of regional power-system dynamics rather than a peripheral digital load, underscoring the value of anticipatory planning that aligns computational growth with electricity-system development.\nResults\nGlobal clustering and heterogeneity in AI data-center siting\nAI data-center locations are identified using a large language model (LLM)-based inference approach that extracts and classifies site-level information from public disclosures, assigning each candidate site a probability of AI-oriented deployment (see Methods). Reported siting patterns reflect three components: (i) outputs of the LLM/RAG framework, specifically the site-level probabilities and determinant intensity scores derived from semantic retrieval and sentiment analysis; (ii) evidence contained in the underlying source documents—corporate announcements, SEC filings, sustainability reports, and media coverage—from which the LLM extracts signals; and (iii) the authors’ interpretation of these outputs in the context of the broader literature, which synthesizes these signals into regional and firm-level characterizations. Figure 1 illustrates the global spatial evolution and underlying determinants of AI data-center siting, with legacy infrastructure and AI-specific siting layers presented separately in Supplementary Fig. S3.\nLegacy data centers (blue markers) are relatively dispersed, reflecting historical priorities such as latency reduction and proximity to users. In contrast, newly identified AI-specific data centers (color-coded circles) exhibit strong geographic concentration, forming dense clusters in a limited number of regions. These clusters are most pronounced in North America, Europe, and the Asia-Pacific, which together account for over 90% of projected AI compute capacity among leading firms. The dominant corridors align with regions characterized by strong energy availability, mature grid and fiber-optic infrastructure, favorable climatic conditions, and supportive policy environments29,30.\nDistinct regional configurations highlight the heterogeneity of siting determinants. The following characterizations are based on signals extracted by the LLM framework from input documents (corporate filings, sustainability disclosures, and policy sources); the synthesis of these signals into regional interpretations represents the authors’ reading of those extracted patterns. In North America, clusters are concentrated in regions such as Virginia, Texas, Ohio, and North Carolina, which coincide with established hyperscale infrastructure, large electricity markets, and documented policy support for data-center development11,31. These regions have historically attracted cloud infrastructure investment, although underlying conditions such as energy costs, regulatory incentives, and climatic factors vary substantially across locations. In Europe, data-center development is concentrated in the Netherlands, the United Kingdom, and Italy. Renewable-energy availability, regulatory frameworks, and digital infrastructure jointly shape these patterns although their relative importance varies across countries. Nordic regions, in particular, provide favorable conditions due to hydro and wind resources and naturally cool climates30. In the Asia-Pacific region, clusters in Singapore, Taipei, Malaysia, and Japan align with strong digital demand and policy-supported infrastructure expansion. Similarly, emerging sites in the Middle East are associated with state-led investment strategies and land availability.\nFirm-level differentiation further reinforces these spatial dynamics (Fig. 1a, b). To interpret the firm-determinant intensity matrix (Fig. 1b), we classify siting drivers into six dimensions: corporate integration, energy access, policy environment, market demand, infrastructure maturity, and network connectivity. Higher scores indicate that a determinant category appears more frequently and with stronger positive sentiment in retrieved documents associated with a given firm. Among globally scaling firms, Amazon exhibits the broadest geographic diversification (Fig. 1a), expanding across regions with strong market demand and mature infrastructure. Microsoft and Oracle primarily follow policy- and incentive-driven corridors, while Google and Meta anchor their siting strategies in regions with strong energy availability and high network connectivity to reduce latency and operational constraints. By contrast, domestically oriented firms such as Apple adopt a more vertically integrated strategy, concentrating deployments within US regions-particularly those with favorable energy conditions–to optimize operational efficiency and alignment with their broader manufacturing and service ecosystems.\nOverall, these results reveal a structural transition from historically dispersed siting toward geographically concentrated AI infrastructure, with clustering patterns shaped jointly by regional characteristics and firm-level strategies. Energy availability, policy alignment, and network connectivity emerge as dominant determinants of siting decisions, driving the formation of high-density infrastructure corridors.\nElectricity consumption projection: rapid scaling of AI data centers elevates system-level energy demand\nTo capture uncertainty in expansion pace and AI workload adoption, we evaluate three forward scenarios for 2025–2030: a conservative scenario (15% annual growth), a neutral scenario (25%), and an optimistic scenario (35%). These growth rates span the range of recent estimates of data-center and AI-related electricity demand growth reported in the literature and industry forecasts11,31. In all scenarios, baseline data-center electricity consumption grows at 10% annually, while the share of AI-intensive workloads increases over time, reflecting the rising importance of generative-AI computation. Electricity demand is estimated by combining firm-level data-center expansion scenarios with projected AI workload intensity and baseline energy-use trends (see Methods). Figure 2 presents the projected electricity consumption of AI data centers under the three growth scenarios from 2025 to 2030. Figure 2a–f illustrates firm-level trajectories for Amazon, Microsoft, Google, Meta, Oracle, and Apple across conservative, neutral, and optimistic assumptions. The energy demand forecasts for all firms show consistent upward trends, reflecting the intensification of AI workloads and the proliferation of large-model training clusters. Aggregate electricity use among hyperscale operators is projected to rise by ~40−70% over the period, though the pace and scale vary considerably by firm. Amazon, Microsoft, and Google are forecasted to maintain the highest absolute demand, consistent with their extensive cloud-service portfolios and broad geographic reach. Meta follows comparable but more moderate forecast paths, supported by continuous improvements in model efficiency and renewable procurement strategies32,33,34,35,36,37. Apple and Oracle exhibit lower projected energy demand, aligning with their smaller operational footprints and vertically integrated architectures. Across all firms, the optimistic scenario accelerates notably after 2027, coinciding with large-scale deployment of generative-AI infrastructure and training workloads38,39.\nFigure 3 aggregates these firm-level forecasts to depict the global evolution of electricity demand associated with leading AI operators from 2018 to 2030. The solid gray line represents historical estimates derived from industrial disclosures and utility data, while the colored dashed lines correspond to scenario-based projections. Total electricity consumption by these leading firms’ data centers is expected to rise from ~118 TWh in 2024 to between 239 TWh (conservative) and 295 TWh (optimistic) by 2030, implying a compound annual growth rate (CAGR) of roughly 13−17%. This increase is equivalent to adding the electricity demand of a medium-sized national market every two years, or comparable to the annual electricity use of about 27 million US households. The projected magnitude is quantitatively consistent with the International Energy Agency’s (IEA) forecast that total global data center electricity demand will reach around 945 TWh by 2030 (~3% of global power consumption)11, compared with about 1.5% in 2024 (see Methods). These projections collectively illustrate the accelerating energy footprint of AI infrastructure within the broader global power system.\nRegional electricity demand pressure implications of AI data-center clustering\nFigure 4 characterizes the regional distribution of data-center electricity demand and the relative concentration of AI-related loads within regional electricity systems in 2030. Figure 4 displays projected electricity demand by firm and region, revealing a highly uneven spatial distribution of AI-related loads. Fewer than ten regions account for nearly two-thirds of total projected demand, indicating a strong concentration of infrastructure in a limited number of states and countries. Oregon, Virginia, Iowa, Ohio and Ireland emerge as dominant hubs, each exceeding 15 TWh of annual demand under the neutral scenario. These regions are associated with favorable conditions—such as established hyperscale clusters, access to energy resources, and supportive policy environments—that are consistent with large-scale AI deployment. Second-tier regions, including the Texas, Netherlands, and Washington, exhibit moderate but rapidly increasing demand, often driven by firm-specific expansion. Scenario ranges indicate that uncertainty in deployment pace and firm strategy could alter local electricity demand by up to 30%, underscoring the challenges of forecasting and planning for digital-infrastructure growth.\nTo characterize the relative concentration of data-center electricity demand across regions, we construct an electricity demand pressure index (EDPI), defined as the ratio of projected data-center electricity demand to regional total available electricity supply (see Methods). Figure 5a presents the cross-regional evolution of EDPI between 2025 and 2030 under the neutral scenario, revealing a pronounced and spatially uneven increase in demand concentration. A small number of regions—including Oregon, Ireland and Iowa—display markedly elevated EDPI levels in both years, indicating a persistently high concentration of data-center electricity demand relative to available supply. Virginia, Nebraska, and Ohio occupy the upper-middle range of the distribution, reflecting growing demand concentration. The absolute change in EDPI between 2025 and 2030 (ΔEDPI, expressed in percentage points) further reveals a self-reinforcing spatial pattern: regions with already elevated demand concentration tend to experience the largest absolute increases. In contrast, regions with larger electricity systems—such as Texas—exhibit comparatively small ΔEDPI values, reflecting greater available-supply adequacy to absorb incremental AI-related demand (Fig. 5a–c).\nOverall, these results reveal a pronounced spatial asymmetry in the distribution of data-center electricity demand, with a limited number of regions accounting for a disproportionate share of demand relative to their available-electricity adequacy.\nDiscussion\nThis study links AI data-center siting patterns, electricity-demand projections, and regional electricity demand concentration. By combining large language model (LLM)-based infrastructure identification with scenario-based electricity-demand projections, we assess how the spatial expansion of AI computing infrastructure may influence electricity demand and the relative burden on regional electricity systems. A key methodological contribution of this framework is the fusion of qualitative corporate intelligence—extracted by LLMs from financial disclosures, sustainability reports, and strategic announcements—with quantitative electricity demand projection. This combination enables inference of firms’ implicit strategic intent from language patterns in corporate communications, providing a more interpretable and forward-looking basis for demand forecasting than statistical extrapolation from historical consumption trends alone11,20.\nOur results indicate that AI-driven data-center expansion is emerging as an increasingly influential component of global electricity systems7,40. Across firms and regions, the analysis reveals a consistent sequence linking spatial siting patterns, rapidly increasing electricity demand, and rising regional power stress. This pattern suggests that the digital infrastructure supporting AI development is becoming closely intertwined with electricity-system dynamics rather than remaining an exogenous source of demand. As AI-related computing capacity expands and concentrates geographically, it may influence regional load growth, infrastructure investment priorities, and decarbonization pathways41.\nThe clustering of AI data centers in a limited number of favorable regions—primarily in North America, Western Europe, and parts of the Asia-Pacific—creates pronounced spatial asymmetries in load distribution. Regions such as Oregon, Ireland and Iowa—which exhibit elevated EDPI values—face a high relative concentration of data-center electricity demand within their electricity systems, alongside rapid AI infrastructure expansion. In contrast, regions with larger electricity systems, such as Texas, show lower demand concentration relative to total available electricity supply. These patterns highlight how regional electricity-system size, cross-boundary trade flows, and the spatial footprint of AI infrastructure jointly determine the relative exposure of local electricity systems to AI-driven demand growth. The projected growth in electricity demand from AI data centers therefore presents both challenges and opportunities for electricity systems. On one hand, concentrated digital loads could intensify transmission congestion, complicate renewable integration, and increase electricity price volatility if system expansion does not keep pace with demand. On the other hand, large and stable computational loads may provide opportunities for greater integration with electricity systems Data centers can function as anchor customers for long-term renewable power purchase agreements (PPAs), potential participants in flexible demand programs, and focal points for co-located energy storage or hybrid renewable generation42,43. The extent to which these opportunities materialize will depend on the interaction between electricity market design, infrastructure investment, and regulatory frameworks.\nAt the policy level, these findings highlight the potential value of anticipatory frameworks that explicitly incorporate AI-related demand into national energy modeling and long-term planning. Conventional forecasts that treat digital infrastructure as exogenous commercial load risk underestimating the spatial concentration of new demand and its implications for annual energy balancing. Integrating firm-level digital-load projections into planning processes could enable regulators and utilities to better anticipate load growth in high-concentration regions. In emerging markets, policy design may need to balance the benefits of digital-infrastructure development against the risk of exacerbating existing demand-supply imbalances.\nImportantly, the firm-level granularity of this analysis offers a policy-relevant tool that complements aggregate demand forecasts. Because AI-related electricity demand is highly concentrated among a small number of identifiable corporate actors, regulators and utilities can use firm-level demand trajectories to engage with specific operators rather than with a generic “technology sector.” This granularity may support more targeted policy interventions—such as renewable procurement requirements, locational guidance, or mandatory load disclosure—calibrated to the expansion strategies and geographical footprints of individual firms.\nSeveral limitations of the current framework warrant explicit acknowledgment. First, the LLM-based inference relies on corporate sustainability reports, press releases, and annual filings, which may contain optimistic bias regarding efficiency trajectories, renewable integration commitments, or future investment plans. To the extent that such documents overstate actual or planned improvements, the sentiment-based siting probabilities may be biased toward high-profile or low-carbon locations, potentially understating deployment in less prominently disclosed regions. Moreover, regions where AI infrastructure projects are less frequently disclosed (such as Africa or Australia), less systematically reported, or still at an early stage of development may be underrepresented within the inferred deployment database. Second, the efficiency-gain assumption of 1−3% annually—calibrated to historical hyperscale patterns—reflects a central scenario. Further improvements in chip efficiency, including more energy-efficient accelerators or advances in cooling architectures, could reduce energy intensity below projected levels. Third, the EDPI is constructed from aggregate annual energy quantities—available electricity supply (generation plus positive net imports) and projected data-center demand in the numerator—and does not capture intra-regional transmission network constraints or interconnection queue delays. In systems such as the PJM interconnection (a regional transmission organization coordinating wholesale electricity markets and grid operations across parts of the eastern United States, including Virginia), grid integration bottlenecks and interconnection queues can be binding constraints on the ability to absorb new large loads, even when aggregate available supply appears sufficient. The EDPI may therefore underestimate localized electricity-system pressures in transmission-constrained areas, and future extensions of this framework should incorporate network-level indicators. Fourth, the firm-level analysis covers six technology companies that together account for ~70−75% of global hyperscale and cloud-linked data-center electricity demand. The remaining share comprises a heterogeneous set of operators—including smaller cloud-service providers, colocation facility operators, telecommunications companies with data-center divisions, and regional or national operators—that are excluded primarily because their infrastructure footprints and electricity consumption are less consistently disclosed in public filings and their siting patterns are not readily identifiable by the LLM-based inference approach applied here. Excluding this segment may lead to an underestimate of total AI-related electricity demand. Fifth, available electricity projections are based on CAGR estimates from 2019–2024 with a small sample and pandemic impact, so they may not fully capture future changes in electricity systems, including renewable deployment, grid expansion, and infrastructure investments. In particular, in regions with relatively small electricity systems, such developments could substantially increase future electricity availability beyond historical trends. Likewise, future national and regional energy policies, market reforms, and unforeseen extreme events may significantly reshape electricity-supply trajectories.\nMore broadly, the co-evolution of digitalization and electrification marks a conceptual turning point in energy-system governance13. AI infrastructure is transforming electricity demand from a passive outcome of economic activity into an active determinant of system configuration. Its social and economic value increasingly depends on the sustainability of its power source and its integration within low-carbon grids. Future research should quantify the feedback loops between AI-compute trajectories, carbon intensity, and investment flows across the energy-digital interface. Several specific directions are particularly promising. First, the co-optimization of AI data-center siting with other emerging high-load technologies—such as hydrogen electrolyzers—could help minimize transmission congestion and renewable curtailment; strategic spatial allocation of such loads has been shown to reduce boundary reinforcement costs44. Second, the interaction between AI data-center loads and the electrification of transport warrants coordinated analysis, as both introduce large and potentially concurrent demands on distribution and transmission infrastructure; holistic planning frameworks that integrate gigawatt-scale AI campuses alongside electric-vehicle charging have been identified as an emerging priority45. Third, the integration of data centers into urban energy systems through waste heat recovery for building heating offers a dual benefit of improved cooling efficiency and urban decarbonization; coordinated control strategies for building thermal flexibility informed by data-center waste heat could reduce local distribution network congestion46. Fourth, the application of advanced risk assessment and stochastic optimization techniques—analogous to those used in commodity supply-chain management47—could better account for uncertainties in energy pricing, grid availability, and AI compute demand, supporting more robust infrastructure planning under deep uncertainty. Understanding and governing the coupling between computational growth and electricity systems will be central to achieving a resilient, low-carbon digital economy.\nMethods\nThis study develops a hybrid retrieval-augmented generation (RAG) + large language model (LLM) forecasting framework to predict the geographic and operational evolution of hyperscale data centers operated by major technology companies. The framework integrates structured document retrieval, language sentiment analysis, and LLM-based contextual reasoning to infer (i) potential expansion locations, (ii) baseline technical parameters, and (iii) multi-year energy trajectories.\nThe workflow (Fig. 6) consists of six main stages: (1) construction of a RAG knowledge base from multi-year corporate and regional sources, (2) sentiment-aware identification of likely future data center locations, (3) retrieval of location-specific contextual information, (4) extraction of baseline operational parameters for the current year, (5) multi-year forecasting of key technical and environmental metrics using an LLM-guided model and physically interpretable energy equations, and (6) construction of the Electricity Demand Pressure Index (EDPI) that compares projected data center demand against the available regional electricity supply capacity.\nKnowledge-base construction\nA multi-source text corpus was constructed for each firm, including press releases, annual earnings reports, environmental sustainability disclosures, infrastructure investment announcements, government policy documents, and major news articles from 2015 to 2025. All texts were preprocessed and embedded using the Hugging Face all-MiniLM-L6-v2 model, then indexed in a FAISS (Facebook AI similarity search) vector database for semantic retrieval. Each entry retained metadata such as publication date, geographic reference, and document type, enabling targeted retrieval (e.g., “AWS Taiwan 2025 expansion energy infrastructure”).\nLocation Identification via RAG + LLM\nTo infer likely future expansion sites between 2025 and 2030, the framework explicitly prompts the LLM to predict which geographic regions a firm is most likely to expand into. This prediction is based on contextual evidence retrieved from the RAG knowledge base, which includes firm sustainability reports, capital expenditure statements, and government or media coverage.\nThe RAG system is first queried using firm-specific and temporal prompts, such as {firm} 2025–2030 data center investment or construction plans. The retrieved documents contain both qualitative statements (e.g., “strong regional demand growth,” “strategic infrastructure partnership”) and quantitative indicators (e.g., “$2B allocated to APAC expansion”).\nEach document is analyzed by the LLM not only for content relevance but also for its sentiment and linguistic tone. Positive sentiment toward investment, expansion, or infrastructure development—particularly in earnings calls or annual reports—acts as an implicit prior that increases the model’s belief in a region’s likelihood of receiving new data center capacity. Conversely, negative sentiment (e.g., mentions of cost control, divestment, or policy uncertainty) decreases the associated probability weight.\nThe LLM then synthesizes these cues into a structured probability distribution over candidate locations, producing outputs of the form: P(expansion at site i ∣evidence) = fLLM(retrieved text embeddings, sentiment scores).\nThis probabilistic representation is used in later stages to weight forecasts of capacity growth and energy consumption at each site.\nBaseline parameter extraction\nFor each confirmed or high-probability site, the system used a RAG-conditioned LLM prompt to extract current-year baseline parameters describing the data center’s IT and facility operations. The baseline operational parameters used in the forecasting framework are summarized in Table 1, including accelerator counts, average power draw, utilization factors, operating hours, and power usage effectiveness (PUE).\nThe model was instructed to reason explicitly about local grid mix, cooling efficiency, and infrastructure maturity, and to avoid copying assumptions between regions. The extracted parameters included: {Ntrain, Ninference, Pavg,train, Pavg,inference, utrain, uinference, Htrain, Hinference, PUE, g}\nAll outputs were formatted as structured JSON (JavaScript Object Notation) objects, accompanied by short notes summarizing assumptions and references. Each variable corresponds to a measurable operational attribute of the local compute fleet.\nForecasting future parameters\nA second RAG + LLM chain projected the time evolution of these parameters over a 5-year horizon. The prompt incorporated both historical operational trends and retrieved regional information, constraining LLM reasoning with empirically observed hyperscale patterns: 10−20% annual accelerator capacity growth, 1−3% annual efficiency gains, and gradual improvements in utilization. These parameters characterize the supply-side evolution of infrastructure, specifically capturing the growth in installed accelerator capacity and incremental improvements in energy efficiency at the facility level. The model generated structured forecasts for each year t ∈ [2025, 2030], each annotated with explanatory notes and relevant document sources.\nEnergy computation\nForecasted operational parameters were post-processed in Python to derive the physical energy demand of each site. The IT load energy (EIT) represents the direct electrical consumption of computing equipment, while the estimated AI data center energy demand (EDC) includes both IT and non-IT overheads such as cooling and power distribution losses.\nHere, N, Pavg, u, and H denote the number of accelerators, average power draw (kW), utilization factor, and annual operating hours, respectively. EIT is thus measured in MWh, and EDC scales this load by the facility’s power usage effectiveness (PUE), capturing site-specific cooling and infrastructure efficiency.\nImplementation\nThe full pipeline was implemented in Python using the LangChain framework for composable RAG pipelines, FAISS for vector retrieval, and the OpenAI GPT-4o-mini model as the reasoning component (temperature = 0.3). All modules were orchestrated via reproducible runnables with strict JSON parsing for downstream numerical analysis.\nSentiment-aware expansion probability modeling\nThe sentiment analysis stage produced, for each candidate region i, an expansion likelihood score Pi defined as:\nwhere Si is the normalized sentiment score for region i (ranging from −1 to +1), and Ri is the retrieval relevance score (cosine similarity in the embedding space). These probabilities informed the sampling or prioritization of forecast targets in subsequent steps. Regions frequently mentioned in corporate reports and described with positive financial tone thus received elevated probabilities of selection.\nThis coupling of semantic retrieval and sentiment tone weighting enables the model to capture implicit strategic intent in corporate communications—an important predictor of near-term data center expansion behavior.\nBy fusing semantic retrieval, sentiment weighting, and physical energy modeling, this hybrid framework bridges qualitative corporate signals and quantitative operational forecasting. The approach captures how firms’ language choices in financial and environmental disclosures correlate with real-world infrastructure trajectories, enabling a more interpretable and evidence-grounded forecast of global data center development.\nScenario design and aggregation\nTo capture uncertainty in both firm-level expansion intensity and technological efficiency, we constructed three forward scenarios s-conservative, neutral, and optimistic-for 2025–2030. Each scenario jointly controls (i) the rate of new AI data-center additions and (ii) the efficiency gains of all operating sites. The annual growth rate of new AI load is set to gnew,s ∈ {0.15, 0.25, 0.35} for scenario s among the three scenarios, while existing stock consumption for each firm grows at gstock = 0.10. Specifically, the three forward scenarios for AI-driven data-center expansion (15, 25, and 35% annual growth) are designed to span conservative, central, and high-growth trajectories of AI workload expansion. These values are informed by recent industry and research evidence on the growth of AI compute demand and infrastructure. AI training workloads are estimated to grow at around 22% annually, while inference workloads may expand at rates approaching 35% over the next five years48. Other estimates indicate that overall AI-related infrastructure demand may grow in the range of 20−25% annually in baseline scenarios49. Based on this range of evidence, the 15% scenario reflects a lower-bound trajectory consistent with moderate expansion in AI deployment, the 25% scenario captures the central tendency of observed infrastructure growth rates, and the 35% scenario represents an upper-tail case aligned with rapid scaling of inference workloads and AI-ready capacity. These scenarios are designed to capture the empirically observed range of AI compute and infrastructure growth, while also reflecting uncertainty in future AI adoption intensity and technological scaling.\nFirm-specific AI shares pAI(f, t) increase gradually over time to reflect the rising share of AI workloads in total compute demand. Firm-specific AI shares pAI(f, t) were defined to represent the evolving proportion of AI-related compute workloads within each operator’s total data-center activity. For each firm f, pAI(f, t) increases over time according to a discrete schedule calibrated to reflect the firm’s relative intensity of AI adoption between 2025 and 2030:\nwhere p1(f), p2(f), and p3(f) correspond to the firm’s baseline (2025), mid-phase (2026–2027), and mature-phase (2028–2030) AI workload shares, respectively. Formally, this can be expressed as:\nThe parameter set {p1(f), p2(f), p3(f)} is firm-specific, capturing heterogeneity in AI adoption trajectories across major hyperscale operators:\nThis stepwise formulation approximates each firm’s increasing allocation of computational resources to AI model training and inference. The heterogeneity in pAI(f, t) reflects differences in investment magnitude, infrastructure build-out, and strategic focus among major operators. Firms with large-scale and early commitments to generative AI—such as Amazon, Google, Meta, and Microsoft—exhibit steeper trajectories, consistent with multibillion-dollar data-center expansions, dedicated AI accelerator deployment (e.g., Trainium, H100, and TPUv5), and vertically integrated AI ecosystems. By contrast, vertically integrated or enterprise-focused operators such as Apple and Oracle show slower growth, consistent with smaller-scale AI infrastructure investment, reliance on consumer-device optimization rather than large-model training, and a later pivot toward cloud-based AI services. Accordingly, the evolution of pAI(f, t) captures each firm’s position along the spectrum of AI-driven digital-infrastructure transformation, serving as a proxy for the relative share of compute capacity devoted to AI workloads within their overall fleet.\nFor each firm f, the existing stock electricity consumption evolves as\nNew-site electricity demand follows compounding growth of AI workloads and the firm-specific AI share schedule:\nThe total firm-level electricity demand is then\nTo represent uncertainty arising from multiple independent siting pathways, we simulate a set of Pt,s plausible firm-region configurations for each year t and scenario s. Global electricity demand in year t and scenario s is the ensemble mean:\nand the range\ndefines the shaded uncertainty bands in the scenario plots, where \\({E}_{t,s,p}^{{{\\rm{glob}}}}\\) denotes the global total electricity demand in year t under scenario s for projection path p ∈ {1, …, Pt,s}. The 2024 observed total serves as the common anchor point for all trajectories.\nRegional allocation of energy\nFirm-level electricity demand is spatially distributed according to two complementary weights: (i) AI siting probabilities derived from the LLM siting model, and (ii) historical stock weights from the existing data-center inventory.\nFor AI-related new sites, firm-year location weights are proportional to modeled AI energy at each location ℓ:\nHere \\({E}_{f,t,\\ell }^{AI,\\,{\\mbox{loc}}\\,}\\) denotes the modeled AI-related electricity demand assigned to firm f at location ℓ in year t, as determined from the LLM-based siting model and corresponding regional AI workload intensity.\nFor legacy and non-AI sites, we use historical weights based on observed facility counts. For legacy and non-AI sites, we use historical weights based on observed facility counts:\nLegacy (historical) data-center counts refer to the number of distinct operating campuses under each firm’s control as of end-2024, including both owned and long-term leased facilities. These are derived from the observed 2015–2024 dataset and serve as proxies for baseline spatial capacity prior to the AI expansion. Since 2024, the global data-center landscape has entered a rapid expansion phase driven by generative-AI workloads, with AI infrastructure investment reaching a record $57 billion globally50. This “AI data-center boom” marks a structural inflection in digital-infrastructure growth.\nLet r(ℓ) map locations to regions. Regional electricity use for firm f and scenario s is:\nSumming over all firms yields regional totals:\nThis approach preserves firm-level energy consistency while ensuring that AI-driven growth concentrates in high-probability siting regions identified by the LLM model, and that legacy loads follow empirically observed footprints.\nCross-validation\nAggregate electricity consumption by the six leading hyperscale operators is projected to rise from ~118 TWh in 2024 to 239–295 TWh by 2030. This projection is broadly consistent with the International Energy Agency’s (IEA) estimate that global data-center electricity demand will reach roughly 945 TWh by 2030 (IEA, 2025). To benchmark the modeled values, we derive an implied 2030 consumption level for the six leading firms based on the IEA’s global forecast. Assuming (i) that hyperscale data centers account for 70% of global data-center electricity use (Synergy Research Group, 2024), and (ii) that the top six operators collectively represent 40% of hyperscale activity, the corresponding implied electricity consumption is:\nThis cross-validation demonstrates that the projected range of 239–295 TWh lies well within the values implied by independent international forecasts. The close correspondence reinforces confidence in the representativeness of the modeled global totals and supports the credibility of the upper-bound scenario in light of established external benchmarks. Overall, the modeled range remains broadly aligned with independent global projections and reflects plausible hyperscale market dynamics.\nElectricity demand pressure index (EDPI)\nTo assess the relative demand burden associated with data-center expansion, we construct an electricity demand pressure index (EDPI), defined as the ratio of annual data-center electricity demand to the regional available electricity supply—comprising in-region generation plus positive net electricity imports. For each region r and year t:\nwhere \\({E}_{r,t}^{{\\mathrm{Avail}}} = {E}_{r,t}^{{\\mathrm{Gen}}} + {\\max} \\left(0, {{E}^{{\\mathrm{NetImport}}}_{r,t}}\\right), {{E}^{{\\mathrm{DC}}}_{r,t}}\\) denotes total annual data-center electricity demand (in TWh), with \\({E}_{r,t}^{Gen}\\) denoting regional electricity generation (TWh), \\({E}_{r,t}^{{\\mathrm{Gen}}}\\) denoting net electricity imports, and \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) denoting total available electricity supply (TWh). Positive net imports are included because imported electricity contributes directly to regional electricity supply. For net-exporting regions, historical electricity exports are treated as adjustable rather than permanently fixed. Accordingly, negative net imports are set to zero in the denominator, reflecting the assumption that part of the electricity historically exported could instead be retained to serve additional local electricity demand associated with AI data-center expansion. This approximation avoids overstating electricity demand pressure in regions that are persistent net electricity exporters while providing a consistent basis for cross-regional comparison.\nFor US states, \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) is derived from EIA state electricity profiles as total retail electricity sales, which by the energy balance identity equals in-state net generation plus net interstate electricity receipts. For countries, \\({E}_{r,t}^{{\\mathrm{Avail}}}\\) is the sum of total electricity generation and positive net electricity imports, both drawn from the Ember Yearly Full Release dataset (2019–2024). The EDPI is interpreted as a demand-share indicator capturing the relative contribution of data-center loads to regional electricity supply.\nRegional available electricity supply is projected using historical trends of available electricity as a baseline approximation. Specifically, available electricity is extrapolated from the period 2019–2024 using a region-specific compound annual growth rate (CAGR), consistent with the five-year window used in the CAGR formula below.\nRegional available electricity baselines\n(\\({E}_{r,2024}^{{\\mathrm{Avail}}}\\)) were obtained from EIA State Electricity Profiles (US states) and the Ember Yearly Full Release dataset (countries), covering 2019–2024, expressed in terawatt-hours (TWh). Available electricity data were extrapolated to 2030 using a compound annual growth rate (CAGR):\nwith \\({{\\mbox{CAGR}}}_{avail,r}={\\left(\\frac{{E}_{r,2024}^{Avail}}{{E}_{r,2019}^{Avail}}\\right)}^{1/5}-1\\).\nRegional electricity available supply is projected using region-specific compound annual growth rates (CAGR) estimated from historical observations during 2019–2024. This approach assumes that recent historical trends provide a reasonable baseline approximation of future electricity-system evolution. The resulting projections should therefore be interpreted as baseline-trend estimates rather than policy-conditioned forecasts.\nData availability\nThe publicly available electricity data used in this study were obtained from the Ember Yearly Electricity Data Explorer (https://ember-energy.org/data/yearly-electricity-data/) and the US Energy Information Administration State Electricity Profiles (https://www.eia.gov/electricity/state/). The International Energy Agency’s Energy and AI report (https://www.iea.org/reports/energy-and-ai) was used for external benchmarking. Additional public corporate and contextual information was obtained from SEC EDGAR filings (https://www.sec.gov/edgar/search/), corporate annual and sustainability reports, press releases, government and grid-operator publications, and media reports. Historical facility-level data for Amazon, Microsoft, Google, Meta, Oracle, and Apple were obtained from S&P Capital IQ (https://www.capitaliq.com/). The processed data and model outputs underlying the data-center siting estimates, electricity-demand projections, electricity demand pressure index calculations, figures, and tables are publicly available at https://github.com/dchen219/ai-data-center-electricity.\nCode availability\nThe custom code used for document processing, LLM-based data-center siting inference, electricity-demand projection, electricity demand pressure index calculation, and figure generation is publicly available at https://github.com/dchen219/ai-data-center-electricity. Analyses were conducted in Python using LangChain, FAISS, the Hugging Face all-MiniLM-L6-v2 embedding model, and OpenAI’s GPT-4o-mini model.\nReferences\n- Stylianou, N. et al. Inside the relentless race for AI capacity. Financial Times. https://ig.ft.com/ai-data-centres/ (2025). \n- Storey, V. C., Yue, W. T., Zhao, J. L. & L., R. Generative artificial intelligence: evolving technology, growing societal impact, and opportunities for information systems research. Inform. Syst. Front. https://doi.org/10.1007/s10796-025-10581-7 (2025). \n- Cho, J. & Moon, J. H. Numerical coupling of energy efficiency and thermal performance for cold plate cooling optimization in high-density compute AI data centers. Energy Build. 348, 116441 (2025). \n- Pilz, K. F., Mahmood, Y. & Heim, L. AI’s Power Requirements under Exponential Growth: Extrapolating ai Data Center Power Demand and Assessing its Potential Impact on U.S. Competitiveness. Tech. Rep. https://www.rand.org/pubs/research_reports/RRA3572-1.html (RAND Corporation, 2025). \n- Oxford Institute for Energy Studies. Artificial Intelligence and its Implications for Electricity Systems. Tech. Rep. 145. https://www.oxfordenergy.org/wpcms/wp-content/uploads/2025/05/OEF-145.pdf (Oxford Institute for Energy Studies, 2025)) \n- Chen, X., Wang, X., Colacelli, A., Lee, M. & Xie, L. Electricity demand and grid impacts of AI data centers: Challenges and prospects. Nexus 3, 100162 https://doi.org/10.1016/j.ynexs.2026.100162 (2026) \n- Stojkovic, J., Zhang, C., Goiri, Íñigo., Torrellas, J. & Choukse, E. DynamoLLM: designing LLM inference clusters for performance and energy efficiency. In 2025 IEEE International Symposium on High Performance Computer Architecture (HPCA) 1348–1362 (2025). \n- Crawford, D., Schallehn, M., Renno, P., Hanbury, P. & Cannarsi, A. How can We meet AI’s Insatiable Demand for Compute Power? https://www.bain.com/insights/how-can-we-meet-ais-insatiable-demand-for-compute-power-technology-report-2025/ (Bain & Company, 2025). \n- Ngata, W. M. et al. The cloud next door: Investigating the environmental and socioeconomic strain of datacenters on local communities. In Proc. 2025 ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies, COMPASS ’25 769–774 (Association for Computing Machinery, 2025). \n- Li, P., Yang, J., Islam, M. A. & Ren, S. Making AI less ‘Thirsty’. Commun. ACM 68, 54–61 (2025). \n- International Energy Agency. Energy and AI. https://www.iea.org/reports/energy-and-ai (2025). \n- Bogmans, C. et al. Power hungry: how AI will drive energy demand. IMF Working Paper 2025/081, International Monetary Fund (2025). \n- Wu, C.-J. et al. Sustainable AI: environmental implications, challenges, and opportunities. In Proc. Machine Learning and Systems 795–813. https://proceedings.mlsys.org/paper_files/paper/2022/hash/462211f67c7d858f663355eff93b745e-Abstract.html (2022). \n- Chen, Z.-M. et al. AI carbon footprint in China sets to double post-2030 carbon peaking. Energy Econ. 150, 108880 (2025). \n- Belkhir, L. & Elmeligi, A. Assessing ICT global emissions footprint: trends to 2040 and recommendations. J. Clean. Prod. 177, 448–463 (2018). \n- Masanet, E., Shehabi, A., Lei, N., Smith, S. & Koomey, J. Recalibrating global data center energy-use estimates. Science 367, 984–986 (2020). \n- Malmodin, J., Lövehagen, N., Bergmark, P. & Lundén, D. ICT sector electricity consumption and greenhouse gas emissions–2020 outcome. Telecommun. Policy 48, 102701 (2024). \n- Debus, C., Piraud, M., Streit, A., Theis, F. & Götz, M. Reporting electricity consumption is essential for sustainable AI. Nat. Mach. Intell. 5, 1176–1178 (2023). \n- Cao, Z. et al. Data center sustainability: revisits and outlooks. IEEE Trans. Sustain. Comput. 9, 236–248 (2024). \n- Xiao, T., Nerini, F. F., Matthews, H. D., Tavoni, M. & You, F. Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA. Nat. Sustain. 8, 1541–1553 (2025). \n- Zhang, Y., Tang, H., Li, H. & Wang, S. Integration and interaction of next-generation AI-focused data centers with smart grids and district energy systems: the state-of-the-art, opportunities and challenges. Renew. Sustain. Energy Rev. 224, 116097 (2025). \n- Han, J. et al. Data-driven distributionally robust optimization of low-carbon data center energy systems considering multi-task response and renewable energy uncertainty. J. Build. Eng. 102, 111937 (2025). \n- Nøland, J. K., Hjelmeland, M. & Korpås, M. Will energy-hungry AI create a baseload power demand boom? IEEE Access 12, 110353–110360 (2024). \n- Luccioni, A. S., Viguier, S. & Ligozat, A.-L. Estimating the carbon footprint of BLOOM, a 176B parameter language model. J. Mach. Learn. Res. 24, 1–15 (2023). \n- Tomlinson, B., Black, R. W., Patterson, D. J. & Torrance, A. W. The carbon emissions of writing and illustrating are lower for AI than for humans. Sci. Rep. 14, 3732 (2024). \n- Ahmed, K. M. U., Bollen, M. H. J. & Alvarez, M. A review of data centers energy consumption and reliability modeling. IEEE Access 9, 152536–152563 (2021). \n- Liu, Y. et al. Energy consumption and emission mitigation prediction based on data center traffic and PUE for global data centers. Glob. Energy Interconnect. 3, 272–282 https://doi.org/10.1016/j.gloei.2020.07.008 (2020). \n- Siddik, M. A. B., Shehabi, A. & Marston, L. The environmental footprint of data centers in the United States. Environ. Res. Lett. https://doi.org/10.1088/1748-9326/abfba1(2021). \n- Cao, Z., Zhou, X., Hu, H., Wang, Z. & Wen, Y. Toward a systematic survey for carbon neutral data centers. IEEE Commun. Surv. Tutor. 24, 895–936 https://doi.org/10.1109/COMST.2022.3161275 (2022). \n- Hyvönen, J. et al. Potential of solar photovoltaics and waste heat utilization in cold climate data centers: case study: Finland and northern Japan. Renew. Sustain. Energy Rev. https://doi.org/10.1016/j.rser.2024.114619 (2024). \n- Shehabi, A. et al. 2024 United States data center energy usage report. Tech. Rep. LBNL-2001637. https://doi.org/10.71468/P1WC7Q (Lawrence Berkeley National Laboratory, 2024). \n- Synergy Research Group. Hyperscale data center count hits 1136; average size increases; US accounts for 54% of total capacity. https://www.srgresearch.com/articles/hyperscale-data-center-count-hits-1136-average-size-increases-us-accounts-for-54-of-total-capacity (2025). \n- Donnellan, D. et al. Uptime Institute global data center survey results 2024. Tech. Rep. https://uptimeinstitute.com/resources/research-and-reports/uptime-institute-global-data-center-survey-results-2024 (Uptime Institute, 2024). \n- Khan, O. Scale your AI transformation with a powerful, secure, and adaptive cloud infrastructure. Microsoft Azure Blog. https://azure.microsoft.com/en-us/blog/scale-your-ai-transformation-with-a-powerful-secure-and-adaptive-cloud-infrastructure/ (2024). \n- Google Sustainability. 2023 Environmental report. https://sustainability.google/reports/google-2023-environmental-report/ (2023) \n- Meta Platforms, Inc. Annual report on Form 10-K for the fiscal year ended december 31, 2024. Tech. Rep. SEC Accession No. 0001326801-25-000017, U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/1326801/000132680125000017/meta-20241231.htm (2025). \n- Sivasubramanian, S. Welcome to a new era of building in the cloud with generative AI on AWS. AWS Machine Learning Blog. https://aws.amazon.com/blogs/machine-learning/welcome-to-a-new-era-of-building-in-the-cloud-with-generative-ai-on-aws/ (2023). \n- Apple Inc. Annual report on Form 10-K for the fiscal year ended september 28, 2024. U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/320193/000032019324000123/aapl-20240928.htm (2024). \n- Oracle Corporation. Oracle offers first zettascale cloud computing cluster. Oracle News press release. https://www.oracle.com/news/announcement/ocw24-oracle-offers-first-zettascale-cloud-computing-cluster-2024-09-11/ (2024). \n- Lambert, S., Caron, E., Lefèvre, L. & Grivel, R. Consolidation of virtual machines to reduce energy consumption of data centers by using ballooning, sharing and swapping mechanisms. Future Gener. Comput. Syst. 174, 107968 https://doi.org/10.1016/j.future.2025.107968 (2026). \n- Bolón-Canedo, V., Morán-Fernández, L., Cancela, B. & Alonso-Betanzos, A. A review of green artificial intelligence: towards a more sustainable future. Neurocomputing 599, 128096 https://doi.org/10.1016/j.neucom.2024.128096 (2024). \n- de Roucy-Rochegonde, L. & Buffard, A. AI, Data Centers and Energy Demand: Reassessing and Exploring the Trends. Tech. Rep. https://www.ifri.org/en/papers/ai-data-centers-and-energy-demand-reassessing-and-exploring-trends-0 (Institut français des relations internationales (Ifri), 2025). \n- Goldman Sachs Research. GS SUSTAIN: Generational Growth: AI, Data Centers and the Coming US Power Demand Surge. Tech. Rep. https://www.goldmansachs.com/pdfs/insights/pages/generational-growth-ai-data-centers-and-the-coming-us-power-surge/report.pdf (Goldman Sachs, 2024). \n- Giannelos, S., Konstantelos, I., Pudjianto, D. & Strbac, G. The impact of electrolyser allocation on Great Britain’s electricity transmission system in 2050. Int. J. Hydrog. Energy 202, 153097 https://doi.org/10.1016/j.ijhydene.2025.153097 (2026). \n- Amann, G. et al. E-mobility Deployment and Impact on Grids: Impact of EV and Charging Infrastructure on European T&D Grids: Innovation needs. Technical Report MJ-09-22-246-EN-N (Publications Office of the European Union, 2022). \n- Dong, Z., Zhang, X., Zhang, L., Giannelos, S. & Strbac, G. Flexibility enhancement of urban energy systems through coordinated space heating aggregation of numerous buildings. Appl. Energy 374, 123971 (2024). \n- Giannelos, S., Konstantelos, I. & Strbac, G. Optimal supply chain design using machine learning, risk assessment and optimisation applied to coal distribution. EURO J. Decis. Process. 13, 100062 (2025). \n- Arora, C., Sorel, M. & Sachdeva, P. The next big shifts in AI workloads and hyperscaler strategies. McKinsey & Company. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-next-big-shifts-in-ai-workloads-and-hyperscaler-strategies (2025) \n- Barth, A., Arora, C., Shenai, G., Noffsinger, J. & Sachdeva, P. The data center balance: how US states can navigate the opportunities and challenges. McKinsey & Company. https://www.mckinsey.com/industries/public-sector/our-insights/the-data-center-balance-how-us-states-can-navigate-the-opportunities-and-challenges (2025). \n- Saavedra, R. & Seaward, S. 2025 Data center marketplace: balancing unprecedented opportunity with strategic risk. U.S. Research Report, Colliers. https://www.colliers.com/en/research/nrep-usdc-data-center-marketplace-2025 (2025). \nAcknowledgements\nNot applicable.\nFunding\nAuthors declare no relevant funding.\nAuthor information\nAuthors and Affiliations\nContributions\nD.C. and Y.C. conceived the study. D.C. designed the analytical framework, collected and processed the data, implemented the LLM/RAG-based analysis, conducted the electricity-demand projections and EDPI analysis, prepared the figures and tables, and drafted the manuscript. Z.Z. contributed to data collection, data validation, and interpretation of results. J.Q. and L.C. contributed to methodological development, results visualization, and validation of the analysis. A.K. and Y.C. supervised the study, provided conceptual guidance, and contributed to the interpretation and revision of the manuscript. All authors reviewed, edited, and approved the final manuscript.\nCorresponding author\nEthics declarations\nCompeting interests\nThe authors declare no competing interests.\nPeer review\nPeer review information\nCommunications Sustainability thanks Rahman Khorramfar and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Nandita Basu. A peer review file is available.\nAdditional information\nPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.\nRights and permissions\nOpen Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.\nAbout this article\nCite this article\nChen, D., Zhou, Z., Cai, Y. et al. Artificial intelligence data centers could reach one percent of global electricity demand by 2030. Commun. Sustain. 1, 147 (2026). https://doi.org/10.1038/s44458-026-00152-5\n- Received: \n- Accepted: \n- Published: \n- Version of record: \n- DOI: https://doi.org/10.1038/s44458-026-00152-5","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//www.nature.com/articles/s44458-026-00152-5","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 62509 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 62509 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":62509,"summary_length":368,"usable_text_length":62509,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":62509,"summary_length":368}},"tags":[],"format_contract_version":"news_item_formats.v1"}}}