Powering the Future of AI: Navigating the Trade-offs for Europe's Energy Transition and Net-Zero Goals

Gbemi Oluleye; Mohammad Hemmati; Vassilis M. Charitopoulos

arxiv: 2606.09617 · v1 · pith:OSHR6F2Dnew · submitted 2026-06-08 · 🧮 math.OC · cs.AI· cs.CY· cs.SY· eess.SY

Powering the Future of AI: Navigating the Trade-offs for Europe's Energy Transition and Net-Zero Goals

Mohammad Hemmati , Gbemi Oluleye , Vassilis M. Charitopoulos This is my paper

Pith reviewed 2026-06-27 15:29 UTC · model grok-4.3

classification 🧮 math.OC cs.AIcs.CYcs.SYeess.SY

keywords AI data centerselectricity demandemissions overshootnet-zero targetspower system optimizationEurope energy transitioncapacity expansionsystem flexibility

0 comments

The pith

AI data centers could add 73-723 TWh of demand in Europe by 2050 and risk 67-181 MtCO2 extra emissions from 2030 onward.

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

The paper applies a spatially explicit optimisation model to Europe under 21 AI growth scenarios to measure the added electricity demand, generation capacity, emissions, and operational changes from hyperscale data centers. It calculates that AI expansion drives between 73 and 723 TWh of extra demand by 2050, which could produce cumulative emissions overshoots of 67 to 181 MtCO2 between 2030 and 2050. The analysis shows that net-zero targets remain reachable by 2050 in the modeled pathways, yet intermediate-year emission risks rise sharply unless policies adjust to the speed of digital growth. Location choices for new AI infrastructure shift after 2030 toward sites with firm power and flexibility rather than simply abundant renewables.

Core claim

Using a spatially explicit optimisation model of Europe across 21 AI growth scenarios, the study shows AI could drive 73-723 TWh of extra demand by 2050, risking cumulative emissions overshoots of 67-181 MtCO2 between 2030 and 2050. After 2030 the geography of AI infrastructure will be shaped more by firm power and system flexibility than by the mere abundance of clean energy. In moderate scenarios AI requires an additional 200 hours of firm generation, which increases LCOE by 35 EUR/MWh in key hubs. Existing infrastructure would require 70 GW additional capacity under pessimistic scenarios and up to 226 GW under managed growth pathways. Workload dynamics strongly shape energy dispatch, syst

What carries the argument

Spatially explicit optimisation model run across 21 AI growth scenarios that quantifies additional demand, capacity, emissions, and operational impacts of data centers.

If this is right

AI requires 70 GW additional capacity in pessimistic cases and up to 226 GW under managed growth pathways.
After 2030, AI infrastructure locations prioritize firm power and flexibility over abundant clean energy.
In moderate scenarios, AI adds 200 hours of firm generation and raises LCOE by 35 EUR/MWh in key hubs.
Improved data-center efficiency markedly lowers required capacity additions and reduces system peaks.
Workload dynamics of data centers directly influence energy dispatch, flexibility needs, and emission outcomes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Grid planners may need to incorporate AI demand forecasts into capacity auctions and transmission planning well before 2030.
Similar spatially explicit models could be applied to other high-AI regions to compare flexibility and emission trade-offs.
If workload scheduling improves beyond current assumptions, the upper bounds on both capacity and emissions could decrease.
Market designs that reward flexibility may become essential to accommodate AI-driven peaks without new fossil capacity.

Load-bearing premise

The 21 AI growth scenarios together with the spatially explicit optimisation model accurately represent future workload dynamics, system flexibility requirements, and emission factors without major unmodeled technological or policy disruptions.

What would settle it

Tracking whether actual European electricity demand and emissions attributable to AI data centers between 2030 and 2050 fall inside or outside the modeled ranges of 73-723 TWh and 67-181 MtCO2 would confirm or refute the central projections.

read the original abstract

The rapid expansion of AI globally has led to the proliferation of energy-intensive hyperscale data centres (DCs), making them as a structurally challenging component in power system planning and operation. Using a spatially explicit optimisation model of Europe across 21 AI growth scenarios, we systematically quantify additional demand, capacity requirements, emissions, and operational impacts of DCs. Results indicate that AI could drive 73-723 TWh of extra demand by 2050, risking cumulative emissions overshoots of 67-181 MtCO2 between 2030 and 2050. Our analysis indicates that after 2030, the geography of AI infrastructure will be shaped more by firm power and system flexibility than by the mere abundance of clean energy. In moderate scenarios, AI requires an additional of 200 hours of firm generation, which increases LCOE by 35 EUR/MWh in key hubs. We show that even under the pessimistic scenarios, existing infrastructure would require 70 GW additional capacity, while under managed growth pathways, this expansion could reach 226 GW. We further find DCs workload dynamics strongly shape energy dispatch, system flexibility, and emissions, while improved efficiency significantly reduces capacity needs, and system peaks. While our findings suggest that net-zero targets for 2050 may be achieved, critical emission risks may appear in the intermediate years, and the EU may compromise its carbon-neutral goals unless policies adapt to this accelerating digital transformation.

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.

Desk Editor's Note private letter to a colleague

The paper runs an existing European power-system model on 21 AI data-center scenarios and reports demand and capacity ranges, but supplies no model details or checks so the numbers stay hard to evaluate.

read the letter

The paper takes a spatially explicit optimization model already used for European power systems and applies it to 21 AI growth scenarios. It produces ranges for added electricity demand (73-723 TWh by 2050), capacity additions (70-226 GW), and cumulative emission overshoots (67-181 MtCO2). It also notes that workload patterns affect dispatch and that efficiency gains cut capacity needs, while firm power and flexibility matter more than raw renewable availability after 2030.

That application to AI loads is the main extension. The spatial treatment and the split between moderate and pessimistic pathways give a usable set of planning numbers for people tracking data-center siting.

The clear limitation is that the text gives no model equations, data sources, historical validation, or sensitivity results. Without those, it is impossible to judge whether the 200-hour firm-generation figure or the 35 EUR/MWh LCOE increase follows from the optimization or from the chosen inputs. The intermediate-year emission risks rest on the same unshown assumptions.

This is for energy-systems researchers and EU policy staff who need order-of-magnitude estimates on digital load growth. A reader already comfortable with capacity-expansion models will extract the scenario comparisons quickly.

It deserves a serious referee if the full methods section supplies the missing formulation, data provenance, and checks and those hold up under review. Otherwise the quantitative claims remain too opaque to rely on.

Referee Report

1 major / 0 minor

Summary. The paper uses a spatially explicit optimization model of the European power system, run across 21 AI growth scenarios, to quantify the impacts of hyperscale data centers. It reports that AI could add 73-723 TWh of electricity demand by 2050, require 70-226 GW of additional capacity, and produce cumulative emission overshoots of 67-181 MtCO2 from 2030-2050, while arguing that net-zero targets remain achievable with policy adaptation but that intermediate-year risks arise from the need for firm power and flexibility after 2030.

Significance. If the model formulation, data, and validation prove robust, the multi-scenario quantification of demand, capacity, and emission trade-offs could inform EU energy planning at the intersection of digital infrastructure and decarbonization. The emphasis on workload dynamics, efficiency gains, and geography shifting toward firm power rather than renewable abundance is a potentially useful framing.

major comments (1)

[Abstract] Abstract: the central numerical claims (73-723 TWh demand, 67-181 MtCO2 overshoots, 70-226 GW capacity, 200 h firm generation, 35 EUR/MWh LCOE increase) are presented as outputs of a spatially explicit optimisation model, yet the manuscript supplies no model equations, objective function, constraints on flexibility or firm capacity, data sources for emission factors or load profiles, scenario construction details, or any validation/sensitivity analysis; without these the reported ranges cannot be assessed for internal consistency or sensitivity to the free parameters (AI growth rates and efficiency improvements).

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. We agree that greater transparency on the model is needed to allow assessment of the results and will revise the manuscript to address this.

read point-by-point responses

Referee: [Abstract] Abstract: the central numerical claims (73-723 TWh demand, 67-181 MtCO2 overshoots, 70-226 GW capacity, 200 h firm generation, 35 EUR/MWh LCOE increase) are presented as outputs of a spatially explicit optimisation model, yet the manuscript supplies no model equations, objective function, constraints on flexibility or firm capacity, data sources for emission factors or load profiles, scenario construction details, or any validation/sensitivity analysis; without these the reported ranges cannot be assessed for internal consistency or sensitivity to the free parameters (AI growth rates and efficiency improvements).

Authors: We agree the abstract omits these elements due to length limits and that the manuscript would benefit from more explicit presentation of the model. The full paper contains a Methods section on the spatially explicit optimization model, but we will revise to add the objective function (cost minimization), key constraints on firm capacity and flexibility, data sources (e.g., emission factors, load profiles), scenario construction details for the 21 cases, and expanded validation/sensitivity results directly in the main text or a dedicated table. We will also insert a sentence in the abstract referencing the model formulation. These changes will be incorporated in the revised version. revision: yes

Circularity Check

0 steps flagged

No circularity detectable; no derivation chain or equations provided

full rationale

The available manuscript text consists solely of the abstract and headline results from a spatially explicit optimisation model run on 21 AI growth scenarios. No model formulation, equations, parameter sources, fitting procedures, scenario construction details, or self-citations appear in the text. Per the hard rules, circularity can only be claimed when a specific reduction is exhibited by quoting the paper's own equations or self-citation chain; none exist here to inspect. The derivation is therefore not analyzable for self-definition, fitted-input predictions, or imported uniqueness, resulting in a score of 0.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on scenario definitions, efficiency assumptions, and the validity of the optimization model; these are not independently verified in the provided abstract.

free parameters (2)

AI growth rates across 21 scenarios
Different growth paths are defined but their derivation or fitting basis is not stated.
Efficiency improvement parameters
Improved efficiency is stated to reduce capacity needs, implying fitted or assumed rates.

axioms (1)

domain assumption The spatially explicit optimisation model correctly captures DC workload dynamics, system flexibility, and dispatch impacts.
All operational and emission results depend on this representation.

pith-pipeline@v0.9.1-grok · 5810 in / 1415 out tokens · 36799 ms · 2026-06-27T15:29:44.939550+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

33 extracted references

[1]

1 Powering the Future of AI: Navigating the Trade-offs for Europe’s Energy Transition and Net-Zero Goals Mohammad Hemmati 1, Gbemi Oluleye 2, and Vassilis M. Charitopoulos 1 1 Department of Chemical Engineering, Sargent Centre for Process Systems Engineering, University College London (UCL), Torrington Place, WC1E 7JE, London, United Kingdom 2 Centre for ...

2050
[2]

In moderate scenarios, AI requires an additional of 200 hours of firm generation, which increases LCOE by €35/MWh in key hubs

Our analysis indicates that after 2030, the geography of AI infrastructure will be shaped more by firm power and system flexibility than by the mere abundance of clean energy. In moderate scenarios, AI requires an additional of 200 hours of firm generation, which increases LCOE by €35/MWh in key hubs. We show that even under the pessimistic scenarios, exi...

2030
[3]

In Europe, this sector currently consumes roughly 70 TWh of electricity, but the trajectory is steepening15

The global digital landscape is undergoing a massive structural shift, with data centres (DCs) now accounting for approximately 1.5% of global electricity demand14. In Europe, this sector currently consumes roughly 70 TWh of electricity, but the trajectory is steepening15. While the International Energy Agency (IEA) forecasts a rise to 115 TWh by 203015,1...

2021
[4]

What if AI growth stalls or collapses after 2030, and what are the structural risks to power grid expansion?

2030
[5]

How will AI growth and efficiency gains reshape European generation patterns, cross-border flows, electricity price and emission pathways? By addressing these questions, this paper provides the first integrated, spatially explicit, and scenario-based analysis of how large-scale AI-driven demand may interact with Europe’s decarbonisation plans, offering ne...

2024
[6]

This close convergence with pan-European benchmarks confirms the model’s robustness as a stable foundation for evaluating the marginal impacts of AI load growth

The results demonstrate a high degree of alignment with official projections, showing a total installed capacity of 4664 GW (3.7% deviation from TYNDP projection) and 103 GW of additional cross-border transmission capacity (<1% deviation from TYNDP projection) by 2050 to reach a total of 233 GW. This close convergence with pan-European benchmarks confirms...

2050
[7]

These scenarios show higher generation from wind and solar, alongside roughly +2% additional nuclear output by 2030, reducing the need for gas, and particularly coal

By contrast, in lower-growth scenarios (e.g., High Eff and Head), clean energy resources absorb the additional load. These scenarios show higher generation from wind and solar, alongside roughly +2% additional nuclear output by 2030, reducing the need for gas, and particularly coal. As a result, emissions decline by up to 12 MtCO2 in some scenarios (blue ...

2030
[8]

Structural evolution of the European power system and emission trajectories under 21 AI-demand scenarios with the US Profile. a, Heatmaps illustrating the temporal deviation in annual electricity generation (TWh) relative to the baseline scenario (2025–2050) across eight key technologies: Wind, Solar, Nuclear, CCGT, CCGT-CCS, H₂-CCGT, Biomass, and Oil & C...

2025
[9]

France remains the most resilient location across all scenarios due to its robust nuclear baseline and the planned deployment of 9 GW of H2CCGT. Similarly, the UK through a diversified portfolio of offshore wind (107 GW), nuclear (18 GW), and a flexible backup fleet of CCGT-CCS (10 GW) and BECCS (2 GW), mountains also relatively stable share in DCs growth...

2035
[10]

These factors transform Poland into a highly attractive destination for DC, capable of matching France’s capacity in lift-off scenarios

This potential is further strengthened in the following decade through a more diversified energy mix, including the deployment of nuclear (17 GW) and H2CCGT (3.7 GW). These factors transform Poland into a highly attractive destination for DC, capable of matching France’s capacity in lift-off scenarios. In contrast, Italy presents a paradox. Despite posses...

2050
[11]

However, in pessimistic scenarios where DC capacity collapses 8 between 2035 and 2045 and then recovers, Italy plays a crucial role in the subsequent recovery. In such cases, the system prefers to leverage existing gas infrastructure to meet additional DC demand rather than investing in new renewable or battery assets, then Italy can host +4 GW in more ag...

2035
[12]

Under this scenario, in both the baseline scenario and the lift-off scenarios, DC capacity could reach up to four times the 2030 capacity in the Nordic

This is particularly evident in the UK profile case, which is more stable and reduces the need for fast-ramp sources, where the DC profile requires a higher level of energy in the early morning hours and can be matched with regions that have high wind availability during that period (Supplementary Material Figure S.12 and S.13). Under this scenario, in bo...

2030
[13]

Pessimistic

a, Spatial mapping of annual AI electricity demand (TWh) across 21 scenarios, categorized by three primary demand trajectories (IEA, ICIS, and McKinsey). Each map represents the regional intensity of AI-related infrastructure under varying growth assumptions, ranging from "Pessimistic" to "Lift-off" scenarios. The colour scale indicates total demand, with...

2035
[14]

Impact of AI demand collapse on system adequacy, technology mix and (LCOE). a. Trade-off between peak demand (GWh) and the share of firm generation (%) in 2050 for ICIS-based pessimistic scenarios (Deflation, Medium, and End) relative to the ICIS Base scenario. Bubble sizes represent the average system LCOE (€/MWh), illustrating a counterintuitive cost in...

2050
[15]

Multidimensional analysis of net electricity trade (TWh/yr) versus cumulative CO2 emissions (Mt CO2, 2030–2050) for 12 key European nations

a. Multidimensional analysis of net electricity trade (TWh/yr) versus cumulative CO2 emissions (Mt CO2, 2030–2050) for 12 key European nations. Yellow bubbles represent the Base Case (without AI), while grey bubbles denote the moderate ICIS Base scenario. The y-axis identifies net exporters (positive) and net importers (negative), while bubble radii corre...

2030
[16]

Arrow thickness represents the magnitude of trade reversal or intensification

Blue shading indicates production increases (e.g., Spain, France), while red shading denotes production displacement (e.g., Sweden). Arrow thickness represents the magnitude of trade reversal or intensification. (a) (b) 15 Influence of 24/7 AI demand on net demand profiles and dispatchable energy requirements This section examines the hourly impacts of DC...

2050
[17]

Varying PUE from 1.4 to 1.1 impacts total cost, emissions, and grid peak demand (Fig

Our Base Case assumes a conservative 1.25 post-2030, but uncertainties in technology or infrastructure upgrades require sensitivity analysis. Varying PUE from 1.4 to 1.1 impacts total cost, emissions, and grid peak demand (Fig. 7). If, contrary to expectations, PUE improves only to 1.4, representing a modest 5% reduction from current levels, total costs i...

2030
[18]

A PUE of 1.4 raises carbon emissions by up to 8% in 2030 and 6% in

2030
[19]

profile reduces emissions by 5%, while the same improvement under the UK scenario cuts only 3%

The effect of efficiency varies with workload volatility: improving PUE to 1.1 under the U.S. profile reduces emissions by 5%, while the same improvement under the UK scenario cuts only 3%. This shows efficiency gains are roughly 1.6 times more effective for volatile AI workloads. The 2050 peak-demand analysis shows that each 0.05 improvement in PUE lower...

2050
[20]

Multi-dimensional sensitivity analysis of AI workload efficiency on grid metrics. The contour plots illustrate the percentage change in total system cost (top), cumulative emissions (middle), and peak demand (bottom) as a function of PUE (x-axis) and temporal progression from 2030 to 2050 (y-axis). The left column represents the UK steady profile, while t...

2030
[21]

However, it substantially increases the likelihood of exceeding Europe’s interim emission targets during the 2030s

AI-driven power demand does not prevent the achievement of long-term net-zero targets. However, it substantially increases the likelihood of exceeding Europe’s interim emission targets during the 2030s. This deviation reflects a temporary gap between the pace of DC demand expansion and the deployment of firm, low-carbon dispatchable capacity required to m...

2035
[22]

To reduce the computational complexity of the large- 24 scale optimisation problem, the planning horizon to 2050 is divided into six multi-year periods with five-year strategic time steps. At the annual level, electricity demand and the availability of onshore wind, offshore wind, and solar resources were derived for the 33 countries under study, resultin...

2050
[23]

This scenario is characterised by the rapid expansion of renewable energy sources, widespread electrification across transport, industry, and heating sectors, as well as the large-scale deployment of low-carbon technologies such as green hydrogen and energy storage. Given that the primary objective of this research is to assess DC power demand expansion u...

2050
[24]

Accordingly, the proposed model does not impose minimum build-rate constraints for renewable technologies, and endogenous development of CCGT capacity is permitted. Moreover, while TYNDP does not explicitly enforce must-run constraints for thermal units beyond 2030, neglecting such constraints, particularly minimum generation and ramping limits, can lead ...

2030
[25]

In addition, whereas many European power system expansion models include penalties VOLL, the cost of renewable curtailment is often neglected. Given projections indicating that renewable curtailment could exceed 310 TWh by 2040 due to grid congestion 42 (outside the scope of this study), a curtailment penalty of 200 €/MWh is explicitly included in the obj...

2040
[26]

According to the IEA, total DC capacity in Europe is projected to increase from the current 16 GW to 27 GW by 2030, corresponding to a CAGR of 11%. ICIS projects a higher annual growth rate of 15%, resulting in a total capacity of 30.4 GW by 2030, while McKinsey forecasts an even more aggressive growth scenario, reaching 35 GW by 2030 with an annual growt...

2030
[27]

It should be noted that all reported values refer to IT capacity

Overall, the reported capacities show reasonable consistency with the figures presented in the IEA report. It should be noted that all reported values refer to IT capacity. The exact IT capacity for hyperscale and colocation DC for all European countries in 2024 is provide in the Supplementary Material, Table S.1. The credibility of these figures can be v...

2024
[28]

A key distinction between the two lies in their temporal variability: the UK-based DCs exhibit a relatively stable load profile with a load factor consistently exceeding 67%

The historical hourly load profiles for both datasets are presented in the Supplementary Material, Figure S.2. A key distinction between the two lies in their temporal variability: the UK-based DCs exhibit a relatively stable load profile with a load factor consistently exceeding 67%. In contrast, the Texas hyperscale data-centre profiles display more pro...

2024
[29]

29 Reference 1 Bevacqua, E., Schleussner, C.-F

47, and 𝐷𝐶𝑃𝑟𝑜𝑓𝑖𝑙𝑒&,% is the normalised hourly demand profile used as model input in cluster k and hour h. 29 Reference 1 Bevacqua, E., Schleussner, C.-F . & Zscheischler, J. A year above 1.5 C signals that Earth is most probably within the 20-year period that will reach the Paris Agreement limit. Nat. Clim. Change. 15, 262-265 (2025). 2 Levin, T., Bistlin...

2025
[30]

(IEA, Paris, 2025). 15 IEA. Energy and AI. ( IEA, 2025). 16 European Commission. In focus: Data centres – an energy-hungry challenge. WebPage, (2025). 17 Green, A., Tai, H., No_singer, J., Sachdeva, P ., Bhan, A., & Sharma, R. How data centers and the energy sector can sate ai’s hunger for power. McKinsey & Company. (2024). 18 Independent Commodity Intell...

2025
[31]

19 Granskog, A., & Hernandez Diaz, D

WebPage, (2025). 19 Granskog, A., & Hernandez Diaz, D. The role of power in unlocking the European AI revolution. McKinsey & Company. (2024). 20 De Roucy-Rochegonde, L. & Bu_ard, A. AI, Data Centers and Energy Demand: Reassessing and Exploring the Trends. The French Institute of International Relations. (2025). 21 Anderson, J., Sweeney, D., & Canonica, R....

arXiv 2025
[32]

41 Entso-e

WebPage, (2022). 41 Entso-e. European Resource Adequacy Assessment 2025 Edition ENTSO-E’s proposal for ACER’s approval. (2024). 42 EUROSTAT, G. Net Production of Electricity and Derived Heat by Type of Plant and Operator. WebPage. (2024). 43 Elsner, P ., Collaer, J., & Uihlein, A., The Onshore Wind Potential of the EU and Neighbouring Countries - ENSPRESO...

2022
[33]

49 Dutch Data Center Association

(2025). 49 Dutch Data Center Association. State of The Dutch Data Centers. (2025). 50 Belgian Digital Infrastructure Association. Overview of the Belgian data center market. (2025). 51 Tehranchi, K., Barnes, T., Frysztacki, M., & Azevedo, I. Pypsa-USA: A ﬂexible open-source energy system model and optimization tool for the united states. (2025) 52 Polish ...

2025

[1] [1]

1 Powering the Future of AI: Navigating the Trade-offs for Europe’s Energy Transition and Net-Zero Goals Mohammad Hemmati 1, Gbemi Oluleye 2, and Vassilis M. Charitopoulos 1 1 Department of Chemical Engineering, Sargent Centre for Process Systems Engineering, University College London (UCL), Torrington Place, WC1E 7JE, London, United Kingdom 2 Centre for ...

2050

[2] [2]

In moderate scenarios, AI requires an additional of 200 hours of firm generation, which increases LCOE by €35/MWh in key hubs

Our analysis indicates that after 2030, the geography of AI infrastructure will be shaped more by firm power and system flexibility than by the mere abundance of clean energy. In moderate scenarios, AI requires an additional of 200 hours of firm generation, which increases LCOE by €35/MWh in key hubs. We show that even under the pessimistic scenarios, exi...

2030

[3] [3]

In Europe, this sector currently consumes roughly 70 TWh of electricity, but the trajectory is steepening15

The global digital landscape is undergoing a massive structural shift, with data centres (DCs) now accounting for approximately 1.5% of global electricity demand14. In Europe, this sector currently consumes roughly 70 TWh of electricity, but the trajectory is steepening15. While the International Energy Agency (IEA) forecasts a rise to 115 TWh by 203015,1...

2021

[4] [4]

What if AI growth stalls or collapses after 2030, and what are the structural risks to power grid expansion?

2030

[5] [5]

How will AI growth and efficiency gains reshape European generation patterns, cross-border flows, electricity price and emission pathways? By addressing these questions, this paper provides the first integrated, spatially explicit, and scenario-based analysis of how large-scale AI-driven demand may interact with Europe’s decarbonisation plans, offering ne...

2024

[6] [6]

This close convergence with pan-European benchmarks confirms the model’s robustness as a stable foundation for evaluating the marginal impacts of AI load growth

The results demonstrate a high degree of alignment with official projections, showing a total installed capacity of 4664 GW (3.7% deviation from TYNDP projection) and 103 GW of additional cross-border transmission capacity (<1% deviation from TYNDP projection) by 2050 to reach a total of 233 GW. This close convergence with pan-European benchmarks confirms...

2050

[7] [7]

These scenarios show higher generation from wind and solar, alongside roughly +2% additional nuclear output by 2030, reducing the need for gas, and particularly coal

By contrast, in lower-growth scenarios (e.g., High Eff and Head), clean energy resources absorb the additional load. These scenarios show higher generation from wind and solar, alongside roughly +2% additional nuclear output by 2030, reducing the need for gas, and particularly coal. As a result, emissions decline by up to 12 MtCO2 in some scenarios (blue ...

2030

[8] [8]

Structural evolution of the European power system and emission trajectories under 21 AI-demand scenarios with the US Profile. a, Heatmaps illustrating the temporal deviation in annual electricity generation (TWh) relative to the baseline scenario (2025–2050) across eight key technologies: Wind, Solar, Nuclear, CCGT, CCGT-CCS, H₂-CCGT, Biomass, and Oil & C...

2025

[9] [9]

France remains the most resilient location across all scenarios due to its robust nuclear baseline and the planned deployment of 9 GW of H2CCGT. Similarly, the UK through a diversified portfolio of offshore wind (107 GW), nuclear (18 GW), and a flexible backup fleet of CCGT-CCS (10 GW) and BECCS (2 GW), mountains also relatively stable share in DCs growth...

2035

[10] [10]

These factors transform Poland into a highly attractive destination for DC, capable of matching France’s capacity in lift-off scenarios

This potential is further strengthened in the following decade through a more diversified energy mix, including the deployment of nuclear (17 GW) and H2CCGT (3.7 GW). These factors transform Poland into a highly attractive destination for DC, capable of matching France’s capacity in lift-off scenarios. In contrast, Italy presents a paradox. Despite posses...

2050

[11] [11]

However, in pessimistic scenarios where DC capacity collapses 8 between 2035 and 2045 and then recovers, Italy plays a crucial role in the subsequent recovery. In such cases, the system prefers to leverage existing gas infrastructure to meet additional DC demand rather than investing in new renewable or battery assets, then Italy can host +4 GW in more ag...

2035

[12] [12]

Under this scenario, in both the baseline scenario and the lift-off scenarios, DC capacity could reach up to four times the 2030 capacity in the Nordic

This is particularly evident in the UK profile case, which is more stable and reduces the need for fast-ramp sources, where the DC profile requires a higher level of energy in the early morning hours and can be matched with regions that have high wind availability during that period (Supplementary Material Figure S.12 and S.13). Under this scenario, in bo...

2030

[13] [13]

Pessimistic

a, Spatial mapping of annual AI electricity demand (TWh) across 21 scenarios, categorized by three primary demand trajectories (IEA, ICIS, and McKinsey). Each map represents the regional intensity of AI-related infrastructure under varying growth assumptions, ranging from "Pessimistic" to "Lift-off" scenarios. The colour scale indicates total demand, with...

2035

[14] [14]

Impact of AI demand collapse on system adequacy, technology mix and (LCOE). a. Trade-off between peak demand (GWh) and the share of firm generation (%) in 2050 for ICIS-based pessimistic scenarios (Deflation, Medium, and End) relative to the ICIS Base scenario. Bubble sizes represent the average system LCOE (€/MWh), illustrating a counterintuitive cost in...

2050

[15] [15]

Multidimensional analysis of net electricity trade (TWh/yr) versus cumulative CO2 emissions (Mt CO2, 2030–2050) for 12 key European nations

a. Multidimensional analysis of net electricity trade (TWh/yr) versus cumulative CO2 emissions (Mt CO2, 2030–2050) for 12 key European nations. Yellow bubbles represent the Base Case (without AI), while grey bubbles denote the moderate ICIS Base scenario. The y-axis identifies net exporters (positive) and net importers (negative), while bubble radii corre...

2030

[16] [16]

Arrow thickness represents the magnitude of trade reversal or intensification

Blue shading indicates production increases (e.g., Spain, France), while red shading denotes production displacement (e.g., Sweden). Arrow thickness represents the magnitude of trade reversal or intensification. (a) (b) 15 Influence of 24/7 AI demand on net demand profiles and dispatchable energy requirements This section examines the hourly impacts of DC...

2050

[17] [17]

Varying PUE from 1.4 to 1.1 impacts total cost, emissions, and grid peak demand (Fig

Our Base Case assumes a conservative 1.25 post-2030, but uncertainties in technology or infrastructure upgrades require sensitivity analysis. Varying PUE from 1.4 to 1.1 impacts total cost, emissions, and grid peak demand (Fig. 7). If, contrary to expectations, PUE improves only to 1.4, representing a modest 5% reduction from current levels, total costs i...

2030

[18] [18]

A PUE of 1.4 raises carbon emissions by up to 8% in 2030 and 6% in

2030

[19] [19]

profile reduces emissions by 5%, while the same improvement under the UK scenario cuts only 3%

The effect of efficiency varies with workload volatility: improving PUE to 1.1 under the U.S. profile reduces emissions by 5%, while the same improvement under the UK scenario cuts only 3%. This shows efficiency gains are roughly 1.6 times more effective for volatile AI workloads. The 2050 peak-demand analysis shows that each 0.05 improvement in PUE lower...

2050

[20] [20]

Multi-dimensional sensitivity analysis of AI workload efficiency on grid metrics. The contour plots illustrate the percentage change in total system cost (top), cumulative emissions (middle), and peak demand (bottom) as a function of PUE (x-axis) and temporal progression from 2030 to 2050 (y-axis). The left column represents the UK steady profile, while t...

2030

[21] [21]

However, it substantially increases the likelihood of exceeding Europe’s interim emission targets during the 2030s

AI-driven power demand does not prevent the achievement of long-term net-zero targets. However, it substantially increases the likelihood of exceeding Europe’s interim emission targets during the 2030s. This deviation reflects a temporary gap between the pace of DC demand expansion and the deployment of firm, low-carbon dispatchable capacity required to m...

2035

[22] [22]

To reduce the computational complexity of the large- 24 scale optimisation problem, the planning horizon to 2050 is divided into six multi-year periods with five-year strategic time steps. At the annual level, electricity demand and the availability of onshore wind, offshore wind, and solar resources were derived for the 33 countries under study, resultin...

2050

[23] [23]

This scenario is characterised by the rapid expansion of renewable energy sources, widespread electrification across transport, industry, and heating sectors, as well as the large-scale deployment of low-carbon technologies such as green hydrogen and energy storage. Given that the primary objective of this research is to assess DC power demand expansion u...

2050

[24] [24]

Accordingly, the proposed model does not impose minimum build-rate constraints for renewable technologies, and endogenous development of CCGT capacity is permitted. Moreover, while TYNDP does not explicitly enforce must-run constraints for thermal units beyond 2030, neglecting such constraints, particularly minimum generation and ramping limits, can lead ...

2030

[25] [25]

In addition, whereas many European power system expansion models include penalties VOLL, the cost of renewable curtailment is often neglected. Given projections indicating that renewable curtailment could exceed 310 TWh by 2040 due to grid congestion 42 (outside the scope of this study), a curtailment penalty of 200 €/MWh is explicitly included in the obj...

2040

[26] [26]

According to the IEA, total DC capacity in Europe is projected to increase from the current 16 GW to 27 GW by 2030, corresponding to a CAGR of 11%. ICIS projects a higher annual growth rate of 15%, resulting in a total capacity of 30.4 GW by 2030, while McKinsey forecasts an even more aggressive growth scenario, reaching 35 GW by 2030 with an annual growt...

2030

[27] [27]

It should be noted that all reported values refer to IT capacity

Overall, the reported capacities show reasonable consistency with the figures presented in the IEA report. It should be noted that all reported values refer to IT capacity. The exact IT capacity for hyperscale and colocation DC for all European countries in 2024 is provide in the Supplementary Material, Table S.1. The credibility of these figures can be v...

2024

[28] [28]

A key distinction between the two lies in their temporal variability: the UK-based DCs exhibit a relatively stable load profile with a load factor consistently exceeding 67%

The historical hourly load profiles for both datasets are presented in the Supplementary Material, Figure S.2. A key distinction between the two lies in their temporal variability: the UK-based DCs exhibit a relatively stable load profile with a load factor consistently exceeding 67%. In contrast, the Texas hyperscale data-centre profiles display more pro...

2024

[29] [29]

29 Reference 1 Bevacqua, E., Schleussner, C.-F

47, and 𝐷𝐶𝑃𝑟𝑜𝑓𝑖𝑙𝑒&,% is the normalised hourly demand profile used as model input in cluster k and hour h. 29 Reference 1 Bevacqua, E., Schleussner, C.-F . & Zscheischler, J. A year above 1.5 C signals that Earth is most probably within the 20-year period that will reach the Paris Agreement limit. Nat. Clim. Change. 15, 262-265 (2025). 2 Levin, T., Bistlin...

2025

[30] [30]

(IEA, Paris, 2025). 15 IEA. Energy and AI. ( IEA, 2025). 16 European Commission. In focus: Data centres – an energy-hungry challenge. WebPage, (2025). 17 Green, A., Tai, H., No_singer, J., Sachdeva, P ., Bhan, A., & Sharma, R. How data centers and the energy sector can sate ai’s hunger for power. McKinsey & Company. (2024). 18 Independent Commodity Intell...

2025

[31] [31]

19 Granskog, A., & Hernandez Diaz, D

WebPage, (2025). 19 Granskog, A., & Hernandez Diaz, D. The role of power in unlocking the European AI revolution. McKinsey & Company. (2024). 20 De Roucy-Rochegonde, L. & Bu_ard, A. AI, Data Centers and Energy Demand: Reassessing and Exploring the Trends. The French Institute of International Relations. (2025). 21 Anderson, J., Sweeney, D., & Canonica, R....

arXiv 2025

[32] [32]

41 Entso-e

WebPage, (2022). 41 Entso-e. European Resource Adequacy Assessment 2025 Edition ENTSO-E’s proposal for ACER’s approval. (2024). 42 EUROSTAT, G. Net Production of Electricity and Derived Heat by Type of Plant and Operator. WebPage. (2024). 43 Elsner, P ., Collaer, J., & Uihlein, A., The Onshore Wind Potential of the EU and Neighbouring Countries - ENSPRESO...

2022

[33] [33]

49 Dutch Data Center Association

(2025). 49 Dutch Data Center Association. State of The Dutch Data Centers. (2025). 50 Belgian Digital Infrastructure Association. Overview of the Belgian data center market. (2025). 51 Tehranchi, K., Barnes, T., Frysztacki, M., & Azevedo, I. Pypsa-USA: A ﬂexible open-source energy system model and optimization tool for the united states. (2025) 52 Polish ...

2025