pith. machine review for the scientific record. sign in

arxiv: 2604.06198 · v1 · submitted 2026-03-13 · 💻 cs.CY · cs.AI

Recognition: 2 theorem links

· Lean Theorem

Concentrated siting of AI data centers drives regional power-system stress under rising global compute demand

Danbo Chen , Zijun Zhou , Yongyang Cai , Jiahong Qin , Ani Katchova , Lei Chen
Authors on Pith no claims yet

Pith reviewed 2026-05-15 12:28 UTC · model grok-4.3

classification 💻 cs.CY cs.AI
keywords AI data centerspower stress indexelectricity consumptiongrid vulnerabilitydata center sitingenergy system modelingcompute demand forecast
0
0 comments X

The pith

Concentrated AI data center siting will drive power stress indices above 0.25 in regions such as Oregon, Virginia, and Ireland by 2030.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper establishes that AI compute growth will concentrate electricity demand from data centers in a handful of regions, creating measurable local vulnerabilities in power grids. It combines LLM analysis of corporate and policy data with energy-system modeling to project that consumption by the six leading firms rises from 118 TWh in 2024 to 239-295 TWh by 2030, or roughly 1 percent of global demand. A sympathetic reader cares because the concentration means some grids absorb the load easily while others face disproportionate stress, turning AI infrastructure into a structural driver of electricity-system planning. The results indicate that siting decisions now will shape grid resilience needs through the end of the decade.

Core claim

The new AI infrastructure is highly concentrated in North America, Western Europe, and the Asia-Pacific, which together account for more than 90 percent of projected compute capacity. Aggregate electricity consumption by the six leading firms rises from roughly 118 TWh in 2024 to between 239 TWh and 295 TWh by 2030. Regions such as Oregon, Virginia, and Ireland experience Power Stress Index values exceeding 0.25, indicating local grid vulnerability, whereas diversified systems such as those in Texas and Japan absorb new loads more effectively. These patterns show AI infrastructure evolving from a marginal digital service into a structural component of power-system dynamics.

What carries the argument

The AI-energy coupling framework, which merges LLM-based extraction of corporate, policy, and media data with quantitative energy-system modeling to compute the Power Stress Index (PSI) that quantifies regional grid vulnerability to added loads.

Load-bearing premise

The LLM-based extraction of corporate, policy, and media data accurately captures future AI compute demand and siting decisions through 2030 without significant bias or omission.

What would settle it

Direct measurement of electricity consumption by the six leading AI firms in 2030 falling well below 239 TWh, or Power Stress Index values in Oregon, Virginia, and Ireland staying below 0.25 despite the projected capacity additions, would falsify the central claim.

read the original abstract

The rapid rise of generative artificial intelligence (AI) is driving unprecedented growth in global computational demand, placing increasing pressure on electricity systems. This study introduces an AI-energy coupling framework that combines large language models (LLMs)-based analysis of corporate, policy, and media data with quantitative energy-system modeling to forecast the electricity footprint of AI-driven data centers from 2025 to 2030. Results show that the new AI infrastructure is highly concentrated in North America, Western Europe, and the Asia-Pacific, which together account for more than 90% of projected compute capacity. Aggregate electricity consumption by the six leading firms is projected to increase from roughly 118 TWh in 2024 to between 239 TWh and 295 TWh by 2030, equivalent to about 1% of global power demand. Regions such as Oregon, Virginia, and Ireland may experience high Power Stress Index (PSI) values exceeding 0.25, indicating local grid vulnerability, whereas diversified systems such as those in Texas and Japan can absorb new loads more effectively. These findings demonstrate that AI infrastructure is evolving from a marginal digital service into a structural component of power-system dynamics, underscoring the need for anticipatory planning that aligns computational growth with renewable expansion and grid resilience.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces an AI-energy coupling framework that combines LLM-based analysis of corporate, policy, and media data with quantitative energy-system modeling to forecast AI data center electricity demand and siting from 2025 to 2030. It projects aggregate consumption by leading firms rising from ~118 TWh in 2024 to 239-295 TWh by 2030 (about 1% of global demand), with >90% of capacity concentrated in North America, Western Europe, and Asia-Pacific. Regional results highlight high Power Stress Index (PSI) values exceeding 0.25 in Oregon, Virginia, and Ireland, indicating grid vulnerability, while diversified systems in Texas and Japan absorb loads more effectively.

Significance. If the projections and PSI thresholds hold after validation, the work would usefully quantify how concentrated AI infrastructure is becoming a structural driver of regional power-system stress, supporting calls for anticipatory grid planning aligned with renewable expansion. The regional differentiation (high-stress vs. resilient grids) could inform policy if the underlying siting forecasts prove robust.

major comments (2)
  1. [Methods] Methods: The LLM-based extraction of 2025-2030 siting decisions and compute demand trajectories lacks any reported validation against 2023-2024 corporate announcements, prompt-engineering details, inter-annotator agreement, or sensitivity tests to temperature/hallucination. These forecasts are the free parameters that directly determine the regional concentration and the PSI values >0.25 reported for Oregon, Virginia, and Ireland; without them the headline claims cannot be evaluated.
  2. [Results] Results: No model equations, uncertainty ranges, or sensitivity tests are provided for the electricity consumption projections or the definition and calculation of the Power Stress Index (PSI). The abstract states specific numerical thresholds and regional comparisons without visible quantitative support or historical validation of the energy-system component, leaving the central claims without load-bearing technical grounding.
minor comments (2)
  1. [Abstract] Abstract: The bounds 239-295 TWh are presented without stating the assumptions or scenarios that produce the range; adding a brief clause would improve clarity.
  2. [Introduction] The introduction of the PSI as a new metric would benefit from a short comparison to existing grid-stress indices in the literature to aid reader interpretation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important areas for improving transparency and technical grounding. We address each major point below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Methods] Methods: The LLM-based extraction of 2025-2030 siting decisions and compute demand trajectories lacks any reported validation against 2023-2024 corporate announcements, prompt-engineering details, inter-annotator agreement, or sensitivity tests to temperature/hallucination. These forecasts are the free parameters that directly determine the regional concentration and the PSI values >0.25 reported for Oregon, Virginia, and Ireland; without them the headline claims cannot be evaluated.

    Authors: We agree that the LLM methodology requires fuller documentation to support evaluation of the forecasts. In the revised manuscript we will expand the Methods section with: (i) the exact prompt templates and chain-of-thought instructions used for corporate, policy, and media extraction; (ii) a validation comparison of LLM outputs against 2023–2024 public corporate announcements and earnings reports for the six leading firms; (iii) inter-annotator agreement metrics obtained from repeated LLM runs with varied seeds; and (iv) sensitivity results across temperature settings (0.0–1.0) and hallucination-mitigation strategies. These additions will directly address the referee’s concern that the siting and demand projections lack reported robustness checks. revision: yes

  2. Referee: [Results] Results: No model equations, uncertainty ranges, or sensitivity tests are provided for the electricity consumption projections or the definition and calculation of the Power Stress Index (PSI). The abstract states specific numerical thresholds and regional comparisons without visible quantitative support or historical validation of the energy-system component, leaving the central claims without load-bearing technical grounding.

    Authors: We accept that the quantitative backbone of the results must be made explicit. The revised manuscript will add: (i) the full mathematical formulation of the electricity-consumption model (including efficiency, utilization, and growth-rate parameters) and the PSI definition (ratio of incremental load to available headroom); (ii) uncertainty ranges for the 2030 projections under low-, central-, and high-demand scenarios; (iii) sensitivity tests on key parameters such as PUE improvement rates and grid capacity expansion assumptions; and (iv) a historical-validation subsection comparing 2024 model outputs to reported electricity consumption figures from the six firms. These changes will supply the load-bearing technical support for the reported thresholds and regional comparisons. revision: yes

Circularity Check

0 steps flagged

No circularity: forecasts derive from external LLM extraction plus independent energy modeling

full rationale

The paper's central results (regional PSI values, 2030 electricity footprints, concentration in Oregon/Virginia/Ireland) are produced by first extracting siting and demand signals from corporate/policy/media texts via LLM and then running those quantities through a separate quantitative energy-system model. No equations, fitted parameters, or self-citations are shown that define the output PSI thresholds in terms of the same extracted data or that rename a fitted quantity as a prediction. The derivation chain therefore remains open to external benchmarks (actual 2023-2024 announcements, grid capacity data) and does not collapse by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 1 invented entities

The central claims rest on the accuracy of LLM extraction of future plans and the applicability of standard energy models to the resulting load additions; no independent evidence for these steps is supplied in the abstract.

free parameters (2)
  • AI compute demand growth trajectory
    Determines the 239-295 TWh range by 2030 and is derived from LLM-processed corporate and policy data.
  • Regional siting distribution
    Concentrates >90% of capacity in three macro-regions and directly drives the PSI values.
axioms (2)
  • domain assumption LLM analysis of corporate, policy, and media sources yields unbiased and complete forecasts of data-center locations and capacity additions through 2030
    Invoked to generate the input data for the energy-system models.
  • domain assumption Standard energy-system models can translate added load into a reliable Power Stress Index without additional grid-specific constraints
    Used to produce the PSI thresholds and regional comparisons.
invented entities (1)
  • Power Stress Index (PSI) no independent evidence
    purpose: Quantifies local grid vulnerability from concentrated AI load additions
    New metric introduced to flag regions exceeding 0.25; no external validation or falsifiable prediction supplied.

pith-pipeline@v0.9.0 · 5541 in / 1559 out tokens · 50754 ms · 2026-05-15T12:28:17.439812+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages

  1. [1]

    Financial Times (2025) https://doi.org/https://www.ft

    Stylianou, N., Learner, S., Bradshaw, T., Uddin, R., Bott, I., Nevitt, C., Clark, D., Joiner, S.: Inside the relentless race for ai capacity. Financial Times (2025) https://doi.org/https://www.ft. com/content/

    work page 2025

  2. [2]

    Information Systems Frontiers (2025) https://doi.org/10.1007/s10796-025-10581-7

    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. Information Systems Frontiers (2025) https://doi.org/10.1007/s10796-025-10581-7

    work page doi:10.1007/s10796-025-10581-7 2025

  3. [3]

    Energy and Buildings348, 116441 (2025) https://doi.org/10.1016/j.enbuild.2025.116441

    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 and Buildings348, 116441 (2025) https://doi.org/10.1016/j.enbuild.2025.116441

    work page doi:10.1016/j.enbuild.2025.116441 2025

  4. [4]

    Research Report RRA-3572-1

    Pilz, K.F., Mahmood, Y., Heim, L.: Ai’s power requirements under exponential growth: Extrapolat- ingaidatacenterpowerdemandandassessingitspotentialimpactonu.s.competitiveness.Technical report, RAND Corporation, Santa Monica, CA (January 28 2025). Research Report RRA-3572-1. https://www.rand.org/pubs/research_reports/RRA3572-1.html

    work page 2025

  5. [5]

    Oxford Energy Forum (145) (2025)

    OxfordInstituteforEnergyStudies:Artificialintelligenceanditsimplicationsforelectricitysystems. Oxford Energy Forum (145) (2025). ISSN 2046-1338

    work page 2025

  6. [6]

    Electricity demand and grid impacts of ai data centers: Challenges and prospects,

    Chen, X., Wang, X., Colacelli, A., Lee, M., Xie, L.: Electricity demand and grid impacts of ai data centers: Challenges and prospects. arXiv preprint arXiv:2509.07218 (2025) arXiv:2509.07218 [eess.SY]. License: CC BY-NC-ND 4.0

    work page arXiv 2025

  7. [7]

    In: 2025 IEEE International Symposium on High Performance Computer Architecture (HPCA), pp

    Stojkovic, J., Zhang, C., Goiri, I., Torrellas, J., Choukse, E.: Dynamollm: Designing llm infer- ence clusters for performance and energy efficiency. In: 2025 IEEE International Symposium on High Performance Computer Architecture (HPCA), pp. 1348–1362 (2025). https://doi.org/10.1109/ HPCA61900.2025.00102 11

    work page arXiv 2025

  8. [8]

    In: Proceedings of the 2025 ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies

    Ngata, W.M., Bashir, N., Westerlaken, M., Liote, L., Chandio, Y., Olivetti, E.: The cloud next door: Investigating the environmental and socioeconomic strain of datacenters on local communities. In: Proceedings of the 2025 ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies. COMPASS ’25, pp. 769–774. Association for Computing Machinery, Ne...

    work page doi:10.1145/3715335.3736324 2025

  9. [9]

    Li, P., Yang, J., Islam, M.A., Ren, S.: Making ai less ’thirsty’. Commun. ACM68(7), 54–61 (2025) https://doi.org/10.1145/3724499

    work page doi:10.1145/3724499 2025

  10. [10]

    Licence: CC BY 4.0

    International Energy Agency: Energy and AI. Licence: CC BY 4.0. https://www.iea.org/reports/ energy-and-ai

    work page

  11. [11]

    In: Marculescu, D., Chi, Y., Wu, C

    Wu, C.-J., Raghavendra, R., Gupta, U., Acun, B., Ardalani, N., Maeng, K., Chang, G., Aga, F., Huang, J., Bai, C., Gschwind, M., Gupta, A., Ott, M., Melnikov, A., Candido, S., Brooks, D., Chauhan, G., Lee, B., Lee, H.-H., Akyildiz, B., Balandat, M., Spisak, J., Jain, R., Rabbat, M., Hazelwood, K.: Sustainable ai: Environmental implications, challenges and ...

    work page 2022

  12. [12]

    Energy Economics150, 108880 (2025) https://doi.org/10.1016/ j.eneco.2025.108880

    Chen, Z.-M., Xiong, Q., Duan, J., Ma, J., Chen, Z., Guo, S.: Ai carbon footprint in china sets to double post-2030 carbon peaking. Energy Economics150, 108880 (2025) https://doi.org/10.1016/ j.eneco.2025.108880

    work page arXiv 2030

  13. [13]

    IEEE Access9, 152536–152563 (2021) https://doi.org/10.1109/ACCESS.2021

    Ahmed, K.M.U., Bollen, M.H.J., Alvarez, M.: A review of data centers energy consumption and reliability modeling. IEEE Access9, 152536–152563 (2021) https://doi.org/10.1109/ACCESS.2021. 3125092

    work page doi:10.1109/access.2021 2021

  14. [14]

    Global Energy Interconnection 3(3), 272–282 (2020) https://doi.org/10.1016/j.gloei.2020.07.008

    Liu,Y.,Wei,X.,Xiao,J.,Liu,Z.,Xu,Y.,Tian,Y.:Energyconsumptionandemissionmitigationpre- diction based on data center traffic and pue for global data centers. Global Energy Interconnection 3(3), 272–282 (2020) https://doi.org/10.1016/j.gloei.2020.07.008

    work page doi:10.1016/j.gloei.2020.07.008 2020

  15. [15]

    Environmental Research Letters16(6), 064017 (2021) https://doi.org/10.1088/1748-9326/ abfba1

    Siddik, M.A.B., Shehabi, A., Marston, L.: The environmental footprint of data centers in the united states. Environmental Research Letters16(6), 064017 (2021) https://doi.org/10.1088/1748-9326/ abfba1

    work page doi:10.1088/1748-9326/ 2021

  16. [16]

    In: 2025 8th International Conference on Computer Informa- tion Science and Application Technology (CISAT), pp

    Zhou, Z., Zhao, C., Li, X., Zhang, H., Chang, R.: Diverse stacking ensemble for attributing llm outputs via relational reasoning. In: 2025 8th International Conference on Computer Informa- tion Science and Application Technology (CISAT), pp. 1089–1092 (2025). https://doi.org/10.1109/ 12 CISAT66811.2025.11181963

    work page arXiv 2025

  17. [17]

    International Journal on Recent and Innovation Trends in Computing and Communication 12(2), 508–18 (2024)

    Mavani, C., Mistry, H.K., Patel, R., Goswami, A.: Artificial intelligence (ai) based data center net- working. International Journal on Recent and Innovation Trends in Computing and Communication 12(2), 508–18 (2024)

    work page 2024

  18. [18]

    IEEE Communications Surveys & Tutorials24(2), 895–936 (2022) https://doi.org/ 10.1109/COMST.2022.3161275

    Cao, Z., Zhou, X., Hu, H., Wang, Z., Wen, Y.: Toward a systematic survey for carbon neutral data centers. IEEE Communications Surveys & Tutorials24(2), 895–936 (2022) https://doi.org/ 10.1109/COMST.2022.3161275

    work page doi:10.1109/comst.2022.3161275 2022

  19. [19]

    case study: Finland and northern japan

    Hyvönen, J., Mori, T., Saunavaara, J., Hiltunen, P., Pärssinen, M., Syri, S.: Potential of solar photovoltaics and waste heat utilization in cold climate data centers. case study: Finland and northern japan. Renewable and Sustainable Energy Reviews201, 114619 (2024) https://doi.org/ 10.1016/j.rser.2024.114619

    work page doi:10.1016/j.rser.2024.114619 2024

  20. [20]

    Industry Report (2024)

    Synergy Research Group: Hyperscale Data Center Capacity: 2024 Market Share Update. Industry Report (2024). https://www.srgresearch.com

    work page 2024

  21. [21]

    Uptime Institute (2024)

    Uptime Institute: Global Data Center Survey 2024. Uptime Institute (2024). https:// uptimeinstitute.com

    work page 2024

  22. [22]

    Azure Blog (2024)

    Microsoft Azure: Azure AI Infrastructure Updates: H100 Clusters and Liquid-Cooled Data Centers. Azure Blog (2024). https://azure.microsoft.com/en-us/blog

    work page 2024

  23. [23]

    Google Sustainability (2023)

    Google: 24/7 Carbon-Free Energy Annual Report. Google Sustainability (2023). https:// sustainability.google

    work page 2023

  24. [24]

    https://investor.fb.com

    Meta Platforms, Inc.: Form 10-K Annual Report (2024). https://investor.fb.com

    work page 2024

  25. [25]

    AWS Blog (2023)

    Amazon Web Services: AWS re:Invent 2023 — Key Announcements: Trainium2, Inferentia2 and AI Infrastructure Updates. AWS Blog (2023). https://aws.amazon.com/blogs/aws

    work page 2023

  26. [26]

    https://investor.apple.com

    Apple Inc.: Form 10-K Annual Report (2024). https://investor.apple.com

    work page 2024

  27. [27]

    Press Release (2024)

    Oracle Corporation: Oracle and NVIDIA Announce Expanded Partnership for AI Infrastructure. Press Release (2024). https://www.oracle.com/news

    work page 2024

  28. [28]

    https: //www.iea.org/reports/electricity-security-2021

    International Energy Agency: Electricity Security 2021: Power Systems in Transition (2021). https: //www.iea.org/reports/electricity-security-2021

    work page 2021

  29. [29]

    Applied Energy300, 117316 (2021) https://doi

    Doering,K.,Sendelbach,L.,Steinschneider,S.,LindsayAnderson,C.:Theeffectsofwindgeneration 13 and other market determinants on price spikes. Applied Energy300, 117316 (2021) https://doi. org/10.1016/j.apenergy.2021.117316

    work page doi:10.1016/j.apenergy.2021.117316 2021

  30. [30]

    Neurocomputing599, 128096 (2024) https://doi.org/10.1016/j.neucom.2024.128096

    Bolón-Canedo, V., Morán-Fernández, L., Cancela, B., Alonso-Betanzos, A.: A review of green artificial intelligence: Towards a more sustainable future. Neurocomputing599, 128096 (2024) https://doi.org/10.1016/j.neucom.2024.128096

    work page doi:10.1016/j.neucom.2024.128096 2024

  31. [31]

    Future Generation Computer Systems174, 107968 (2026) https://doi.org/10.1016/j.future.2025.107968

    Lambert, S., Caron, E., Lefevre, L., Grivel, R.: Consolidation of virtual machines to reduce energy consumption of data centers by using ballooning, sharing and swapping mechanisms. Future Generation Computer Systems174, 107968 (2026) https://doi.org/10.1016/j.future.2025.107968

    work page doi:10.1016/j.future.2025.107968 2026

  32. [32]

    IFRI (2025)

    De Roucy-Rochegonde, L., Buffard, A.: AI, Data Centers and Energy Demand. IFRI (2025)

    work page 2025

  33. [33]

    Davenport, C., Singer, B., Mehta, N., Lee, B., Mackay, J., Modak, A., Corbett, B., Miller, J., Hari, T., Ritchie, J., et al.: Ai, data centers and the coming us power demand surge. Goldman Sachs26 (2024) Methods This study develops a hybrid Retrieval-Augmented Generation (RAG) + Large Language Model (LLM) forecasting framework to predict the geographic an...

    work page 2024

  34. [34]

    AWS Taiwan 2025 expansion energy infrastructure

    Knowledge Base Construction A 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 2015–2025. All texts were preprocessed and embedded using the Hugging Faceall-M...

    work page 2015

  35. [35]

    strong regional demand growth,

    Location Identification via RAG + LLM To infer likely future expansion sites between 2025 and 2030, the framework explicitly prompts the LLM topredict 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 expe...

    work page 2025

  36. [36]

    The model was instructed to reason explicitly about local grid mix, cooling efficiency, and infrastructure maturity, and to avoid copying assumptions between regions

    Baseline Parameter Extraction For each confirmed or high-probability site, the system used a RAG-conditioned LLM prompt to extract current-year baseline parametersdescribing the data center’s IT and facility operations. The model was instructed to reason explicitly about local grid mix, cooling efficiency, and infrastructure maturity, and to avoid copying...

    work page

  37. [37]

    Forecasting Future Parameters A second RAG + LLM chain projected the time evolution of these parameters over a five-year horizon. The prompt incorporated both historical operational trends and retrieved regional information, con- straining LLM reasoning with empirically observed hyperscale patterns: 10–20% annual capacity growth, 1–3% annual efficiency ga...

    work page 2025

  38. [38]

    Energy Computation Forecasted operational parameters were post-processed in Python to derive the physical energy demand of each site. TheIT load energy(E IT) represents the direct electrical consumption of computing equipment, while the estimated AIdata center energy demand(E DC) includes both IT and non-IT overheads such as cooling and power distribution...

    work page

  39. [39]

    All modules were orchestrated via reproducible runnables with strict JSON parsing for downstream numerical analysis

    Implementation The full pipeline was implemented in Python using theLangChainframework for composable RAG pipelines,FAISSfor vector retrieval, and theOpenAI GPT-4o-minimodel as the reasoning component (temperature = 0.3). All modules were orchestrated via reproducible runnables with strict JSON parsing for downstream numerical analysis

    work page

  40. [40]

    These probabilities informed the sampling or 18 prioritization of forecast targets in subsequent steps

    Sentiment-Aware Expansion Probability Modeling The sentiment analysis stage produced, for each candidate regioni, anexpansion likelihood score Pi defined as: Pi = Si ·R iP j Sj ·R j (3) whereS i is the normalized sentiment score for regioni(ranging from−1to+1), andR i is the retrieval relevance score (cosine similarity in the embedding space). These proba...

    work page

  41. [41]

    Each scenario jointly con- trols (i) the rate of new AI data-center additions and (ii) the efficiency gains of all operating sites

    Scenario design and aggregation To capture uncertainty in both firm-level expansion intensity and technological efficiency, we constructed three forward scenarioss-conservative,neutral, andoptimistic-for 2025–2030. Each scenario jointly con- trols (i) the rate of new AI data-center additions and (ii) the efficiency gains of all operating sites. The annual...

    work page 2025

  42. [42]

    Regional allocation of energy Firm-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. For AI-related new sites, firm–year location weights are proportional to modelled AI energy at...

    work page

  43. [43]

    Cross-validation Aggregate electricity consumption by the six leading hyperscale operators is projected to rise from approximately 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 ...

    work page 2024

  44. [44]

    AI data-center boom

    Regional Power Stress Index (PSI) To assess the potential strain imposed by data center expansion on regional electricity systems, we constructed aPower Stress Index (PSI)that compares projected hyperscale data center demand against the available regional electricity supply capacity. 1Legacy (historical) data-center countsrefer to the number of distinct o...

    work page 2024

  45. [45]

    AI Compute Fabric

    for regional AI infrastructure; focus on enterprise-grade AI work- loads and secure environments. Apple ~20 data centers (U.S., Europe) Mid-scale, vertically integrated <2% (internal capacity) Emphasis on on-device/private AI; investment in internal “AI Compute Fabric” for LLM training and edge inference; limited public-cloud expo- sure. Note:Market share...

    work page 2025