Rapid AI expansion could add as much as 723 TWh of electricity demand by 2050 and force up to 226 GW of extra generation, driving higher local power costs and intermediate-year emissions unless efficiency and grid flexibility are scaled up; even moderate growth raises hourly firm-generation needs and pushes LCOE up by ~€35/MWh in key hubs.

Main Finding

Large-scale AI/data-centre (DC) growth can materially reshape Europe’s power system: it may add 73–723 TWh/yr of electricity demand by 2050, drive large increases in gas-fired and other firm capacity, create meaningful short‑to‑medium‑term carbon‑budget overshoots (≈67–181 MtCO2, 2030–2050) and alter cross‑border trade and locational attractiveness for DCs. Where DCs locate after ~2030 will be driven more by access to firm, flexible power than by the mere abundance of renewables. Efficiency, workload timing and managed policy can substantially reduce capacity needs and emissions; unmanaged growth risks stranded assets, higher LCOE in some hubs, and interim emission spikes even if net‑zero by 2050 remains feasible.

Key Points

• Demand and emissions

  • Scenario range: 73–723 TWh of additional DC demand in Europe by 2050 (21 scenarios).
  • Cumulative emission overshoots of ~67–181 MtCO2 between 2030–2050 under different trajectories.
  • Moderate ICIS scenario example: ~280 TWh of AI demand by 2050 → ~+77 MtCO2 (2030–2050), largely from increased CCGT use (+300 TWh) and some coal increases in Central Europe.

• Capacity and technology mix

  • Europe may need +70 GW to +226 GW of additional capacity depending on scenario (pessimistic vs managed growth).
  • Near term (to 2030): wind is the preferred expansion (additional 2–50 GW by 2030; ≈+80 TWh generation).
  • Firm/flexible assets (CCGT, CCGT-CCS, H2‑CCGT, nuclear) become increasingly necessary post‑2035 to meet reserves and peaks driven by DC loads (examples: +20–60 GW CCGT by 2050; some scenarios add H2‑CCGT, 2–12 GW nuclear).
  • Renewables and batteries respond non‑linearly: increased in moderate/high-efficiency paths but often reduced later as firm capacity is retained; many scenarios show 80–160 TWh less wind generation than baseline by 2050.

• System operation and economics

  • DC workload temporal shape matters (UK vs US profiles used): load timing affects energy dispatch, flexibility needs, and where DCs are most compatible.
  • Moderate scenarios require ~200 extra hours of firm generation, increasing LCOE in key hubs (abstract cites ~€35/MWh in hubs).
  • Counterintuitively, if AI expansion stalls (collapse scenarios), the system can end up with higher average LCOE (€2–4/MWh increase observed) because pre‑committed firm capacity stays online and renewables get spilled/curtailed — illustrating stranded‑asset and over‑capacity risk.

• Geography and siting

  • Post‑2030 siting driven by access to flexible, low‑carbon firm power, not only renewable resource abundance.
  • Germany, Netherlands, Ireland may face "deadlock" (gas dependence, carbon penalties) and see DC growth plateau.
  • Spain/Portugal, France, Poland and the UK emerge as relatively favourable destinations in many scenarios (due to solar/wind + flexibility or nuclear baseload).

Data & Methods

• Model

  • Spatially explicit mixed‑integer linear programming (MILP) model covering 33 European countries.
  • Joint long‑term capacity expansion and hourly dispatch optimization (hourly resolution).
  • Myopic rolling horizon (5‑year foresight) used for collapse‑scenario analysis to mimic realistic investment decisions.

• Scenarios and inputs

  • 21 AI demand scenarios built from IEA, ICIS, McKinsey projections, and stylized pessimistic/optimistic variants (Base, Lift‑off, Headwinds, High‑eff, Deflation, Medium, End).
  • Two empirically derived hourly DC load profiles (UK and US) to capture workload timing and variability.
  • Baseline (“Without AI”) validated against ENTSO‑E TYNDP 2024 (model ≈3.7% capacity deviation; transmission ≈<1%).
  • Technology and cost assumptions, existing development plans, and climate policy targets incorporated (details in Supplementary Material of original paper).

• Key outputs reported

  • Annual generation deviations by technology (TWh), installed capacity changes (GW), hourly dispatch & reserves, cross‑border flows, LCOE changes, renewable curtailment, and cumulative emissions.

• Limitations (model caveats)

  • Focused on the power sector (no economy‑wide feedbacks, limited demand‑side behavioral modeling outside DCs).
  • Scenario uncertainty large (AI trajectory, efficiency improvements, policy responses).
  • Results sensitive to assumed DC load shapes, techno‑economic inputs, and carbon policy enforcement.

Implications for AI Economics

• Investment & siting decisions

  • Power availability, flexibility and access to low‑carbon firm capacity will become primary determinants of DC siting economics after ~2030, not only renewable resource quality.
  • Developers face trade‑offs: cheaper/noisy grid energy vs carbon exposure and potential future retrofits or curtailment risk. Power purchase agreements (PPAs) and long‑term firm capacity contracts (including low‑carbon firm capacity premiums) will gain importance.

• Cost risk & stranded assets

  • Rapid, uncertain DC growth creates risk of over‑investment in firm capacity (or renewables+storage) that can raise average system LCOE if demand collapses — raising financing risk and required returns for both utilities and DC operators.
  • Policymakers and investors need staged/contingent investment signals and decommissioning or repurposing options to reduce stranded‑asset costs.

• Electricity price formation & externalities

  • Additional AI demand can raise local LCOE (hub impacts up to tens of €/MWh in some scenarios) and change cross‑border price differentials — affecting other industries and consumers.
  • Carbon budget impacts imply external costs not currently internalized by many DC operators; carbon pricing or dedicated low‑carbon power obligations for hyperscalers may be necessary to align incentives.

• Value of efficiency and flexibility

  • Model shows efficiency improvements and workload flexibility (temporal shifting, demand response) substantially reduce system capacity needs, peaks and emissions — indicating high economic value for investments in model efficiency, scheduling algorithms, and flexible contracts.
  • Markets valuing flexibility (fast ramps, demand shifting) and long‑duration low‑carbon firm products will increase in importance for AI firms.

• Policy and market design

  • Integrate AI/DC demand explicitly into national and EU power planning (TYNDP and similar models) to avoid mis‑sizing the system.
  • Support low‑carbon firm resources (CCS, H₂‑fired plants, nuclear, long‑duration storage) and create procurement mechanisms to match DC needs without short‑term carbon backsliding.
  • Encourage locational guidance, transmission upgrades, and market products for capacity/reserves that reflect DC operational profiles.

Overall, the paper highlights that AI’s electricity footprint is not a marginal concern: it can shift generation mixes, infrastructure investment paths, prices, and carbon trajectories. For AI economics, that means accounting for power system constraints, carbon externalities, location‑specific firm‑power premiums, and the value of demand‑side flexibility in commercial and policy decisions.