The Commonplace
Home Papers Evidence Explore Trends Syntheses Digests References Docs 🎲 Workforce Futures
← Papers
Direction, evidence grade, and study type are AI-generated labels (gpt-5-mini), not human-verified. Syntheses are LLM-written. "Tensions" are machine-detected candidates, not confirmed contradictions. A research-acceleration tool, not peer review. How this is built →

Generative-AI data-centers could drive Northern Virginia's power reserve margin below adequacy within 3–5 years, with data-centers consuming over a third of regional electricity by 2031 in realistic scenarios; regulator-triggered responses can restore supply but only after a costly multi-year undershoot because demand adjusts far faster than new supply comes online.

Opening the black box of data center electricity consumption forecasting: An application of system dynamics modeling to Northern Virginia
Rémi Paccou, Fons Wijnhoven · August 08, 2026 · Sustainable Futures
openalex theoretical low evidence 7/10 relevance Summary only summary available; pdf_status=paywall DOI Source PDF

Structured author observations

Linked only from stored provider relations; the raw author line above is never matched by name.

OpenAlex

Latest observation:

  1. Rémi Paccou provider ID
  2. Fons Wijnhoven provider ID

Semantic Scholar

Latest observation:

  1. Rémi Paccou provider ID
  2. Fons Wijnhoven provider ID
A system dynamics model calibrated to Northern Virginia finds that fast-growing GenAI data-center demand can push regional reserve margins below adequacy thresholds within 3–5 years and that regulatory feedbacks restore adequacy only after a multi-year undersupply overshoot due to asymmetric adjustment speeds.

Citation observations

Cumulative provider counts captured on specific dates; providers are never combined.

Policymakers in data center-intensive regions face surging electricity demand driven by artificial intelligence (AI), especially Generative AI (GenAI), which can create tensions between data center expansion and regional grid capacity. Institutional data center electricity consumption forecasts from JLARC, EPRI, and the IEA project continued demand growth but do not model the feedback dynamics — time delays, regulatory responses, rebound effects — that determine whether grid adequacy thresholds are breached or maintained. This article develops a system dynamics model that makes these feedback structures explicit, disaggregating GenAI and non-GenAI data center demand and coupling them with three NERC-aligned electricity supply trajectories. We apply the model to Northern Virginia, the world’s largest data center market, where data centers already consume approximately 25 % of regional electricity. Under four different baseline scenarios, the anticipated reserve margin (ARM) falls below the reference margin level (RML) of 17.7 % within a 3–5 year window, with the abundance-realistic scenario breaching this threshold by 2028–2029 and data center share of total electricity consumption reaching 36–40 % by 2031. When ARM-triggered regulatory feedback is introduced, the model shows that grid adequacy can be restored, but only after a four-year overshoot during which the ARM remains below the RML — a delay driven by the asymmetry between fast demand adjustment (0–2 years) and slow supply expansion (3–10 years). These findings indicate that static forecasts systematically underestimate the timing challenge. The model is designed to be portable to other DC-intensive regions by substituting locally valid parameter values.

Summary

Main Finding

A system dynamics model that explicitly represents feedbacks (time delays, regulatory responses, rebound effects) shows that static institutional forecasts understate how quickly data-center electricity demand driven by Generative AI can push regional grids below adequacy thresholds. Applied to Northern Virginia, the model projects the anticipated reserve margin (ARM) will fall below the reference margin level (RML = 17.7%) within 3–5 years across four baseline scenarios; in the abundance-realistic case the breach occurs by 2028–2029 and data centers consume 36–40% of regional electricity by 2031. Introducing ARM-triggered regulatory feedback can restore adequacy, but only after a roughly four-year undersupply overshoot because demand adjusts quickly (0–2 years) while supply expands slowly (3–10 years).

Key Points

  • Institutional forecasts (JLARC, EPRI, IEA) project continued data center electricity growth but do not model feedback dynamics that govern whether/when grid adequacy is breached.
  • The paper builds a system dynamics model disaggregating GenAI and non-GenAI data-center demand and coupling those demands to three NERC-aligned electricity supply trajectories.
  • In Northern Virginia (current data-center share ~25% of electricity), under four baseline scenarios ARM drops below the 17.7% RML within 3–5 years.
  • The abundance-realistic scenario breaches RML by 2028–2029; by 2031 data centers may account for ~36–40% of regional electricity.
  • Regulatory responses tied to ARM can recover adequacy but after a multi-year overshoot because of asymmetric adjustment speeds: demand reduces rapidly (0–2 years) while new supply and transmission take 3–10 years to come online.
  • Conclusion: static, single-path forecasts systematically underestimate the timing and severity of adequacy shortfalls when fast-growing GenAI demand is present.

Data & Methods

  • Modeling approach: system dynamics model that makes feedback loops explicit (time lags, regulatory triggers, rebound effects).
  • Demand side: disaggregated into GenAI-driven and non-GenAI data-center loads, with scenario-based growth assumptions for GenAI uptake/intensity.
  • Supply side: coupled to three NERC-aligned electricity supply trajectories (scenario envelopes reflecting conservative → abundant supply outcomes).
  • Scenarios: four baseline demand-supply combinations (including an “abundance-realistic” case highlighted in results); ARM computed and compared to RML = 17.7%.
  • Key parameters highlighted in results: current DC share ≈ 25% of regional electricity; demand adjustment latency (0–2 years); supply expansion latency (3–10 years).
  • Sensitivity: results driven by interaction of growth rates, supply trajectories, and policy-trigger timing (model reported to be portable to other regions via parameter substitution).

Implications for AI Economics

  • Forecasting and planning: Economic assessments of GenAI deployment must move beyond static demand forecasts to dynamic models that capture feedbacks and timing; otherwise policymakers will underestimate short-term adequacy risks.
  • Policy design: Regulatory triggers tied to reserve margins can work but must be anticipatory—delays in supply-side responses mean corrective policies need early deployment or faster supply-side channels (permitting, transmission builds, peaking capacity, storage).
  • Market signals & incentives: Time-varying pricing, demand-response programs, and siting incentives (locating high-load AI facilities where capacity exists) become economically important to internalize grid constraints and avoid costly overshoots.
  • Investment and externalities: Rapid GenAI-induced demand growth can shift electricity allocation and prices, creating externalities for other consumers and for regional economic development; these should be internalized in cost–benefit and location decisions for AI infrastructure.
  • Transferability: The model is portable — regions with different grid characteristics should run analogous dynamic scenarios to identify timing risks and optimal mixes of supply- and demand-side interventions.

Assessment

Paper Typetheoretical Evidence Strengthlow — Findings derive from a simulation model and scenario assumptions rather than empirical causal estimation or out-of-sample validation; results are informative about possible dynamics but highly sensitive to growth, latency, and policy-parameter choices. Methods Rigormedium — The study explicitly models feedbacks, disaggregates GenAI vs non-GenAI demand, ties supply to NERC-aligned trajectories, and reports sensitivity to key parameters; however, it lacks empirical validation, formal uncertainty quantification (e.g., probabilistic forecasts), and may omit market-price and cross-regional interactions that would affect real-world outcomes. SampleA system dynamics simulation calibrated to Northern Virginia regional electricity data (current data-center share ≈ 25% of electricity). Demand is disaggregated into GenAI-driven and non-GenAI data-center loads with four scenario-based GenAI growth/intensity paths (including an 'abundance-realistic' case); supply is represented by three NERC-aligned electricity supply trajectories (conservative → abundant). Key modeled parameters include demand adjustment latency (0–2 years), supply expansion latency (3–10 years), and a reference maximum load (RML = 17.7%) used to compute anticipated reserve margin (ARM). Themesadoption governance IdentificationNo empirical causal identification; uses a system dynamics simulation that encodes mechanistic feedbacks (time lags, regulatory triggers, rebound effects) and scenario-based parameter values to project how GenAI-driven data-center demand interacts with prescribed supply trajectories. GeneralizabilityResults depend on scenario assumptions (GenAI growth, intensity) and specific regional parameters; other regions with different mix/dispatch, interconnection capacity, or demand profiles may behave differently., Model lacks empirical validation against historical episodes of rapid demand growth, so timing estimates are uncertain., Does not appear to fully endogenize electricity prices, cross-regional transfers, or firm-level siting/market responses, limiting applicability to real-world market outcomes., Policy and permitting processes, which vary across jurisdictions, will alter supply latency and regulatory effectiveness, reducing direct transferability without re-parameterization., Technological change in GenAI model efficiency, hardware, and cooling could materially change demand trajectories and thus limit long-run generalizability.

Claims (6)

ClaimDirectionOutcomeConfidence & EvidenceDetails
Across four baseline demand-supply scenarios for Northern Virginia, the anticipated reserve margin falls below the reference margin level of 17.7% within 3–5 years. Fiscal And Macroeconomic negative Regional electricity-grid reserve adequacy
Reading fidelity high
Study strength medium
ARM falls below RML within 3–5 years
0.12
In the abundance-realistic scenario, the reserve-margin breach occurs by 2028–2029. Fiscal And Macroeconomic negative Timing of regional electricity-grid adequacy shortfall
Reading fidelity high
Study strength medium
breach by 2028–2029
0.12
Under the abundance-realistic scenario, data centers may consume approximately 36–40% of Northern Virginia's regional electricity by 2031. Adoption Rate positive Data-center share of regional electricity consumption
Reading fidelity high
Study strength medium
36–40% of regional electricity by 2031
0.12
ARM-triggered regulatory feedback can restore grid adequacy, but only after an approximately four-year undersupply overshoot. Governance And Regulation mixed Grid adequacy following regulatory intervention
Reading fidelity high
Study strength medium
roughly four-year undersupply overshoot
0.12
The delayed recovery following regulatory intervention is driven by asymmetric adjustment speeds: demand can adjust within 0–2 years, whereas supply expansion and transmission take 3–10 years. Task Allocation negative Speed of electricity-demand adjustment relative to supply expansion
Reading fidelity high
Study strength medium
demand adjustment latency 0–2 years; supply expansion latency 3–10 years
0.12
Static, single-path institutional forecasts systematically underestimate the timing and severity of electricity-grid adequacy shortfalls when rapidly growing GenAI demand is present. Fiscal And Macroeconomic negative Accuracy of forecasts of grid adequacy shortfalls
Reading fidelity high
Study strength medium
not reported
0.12

Notes