Skip to content
Studies · Study 1: PJM Grid → Study 2: Spatial Siting

Beyond 19 GW: Where Data Centers Actually Matter

Study 1 asked "how much?" and got 19 GW. But 70% of PJM's data center load concentrates in Dominion Virginia. When you ask "where?" — same model, same data — the answer drops to 1–2 GW. The system-wide answer overstates by 10x. Factor in uncertainty and it gets worse: only 62% of Monte Carlo draws show reliable capacity even at today's load.

19 GW
System-Wide
1–2 GW
Northern Virginia
62%
Baseline MC Reliable
200
Stochastic Draws
Fidelity Escalation

Three Levels of Resolution

Each step changed the answer. The same dispatch engine, three different questions.

19 GW
System-wide
1–2 GW
Zonal
62%
Stochastic
Fidelity Lesson

What the Grid Revealed

Three lessons about resolution, uncertainty, and compound returns.

The System Answer Was Wrong

Single-node said 19 GW. Zonal said 1–2 GW. The right fidelity depends on the question, not the system.

Q1: Spatial concentration

MC Changed It Again

Deterministic zonal: 6 hours at baseline. 200-draw MC: only 62% reliable. Stochastic fidelity matters as much as spatial fidelity.

Q1: Monte Carlo overlay

One Model, Compound Returns

Same dispatch engine as Study 1. Same MC framework. New question, fundamentally different answer. Each study reveals the next question.

All questions
Sensitivity Analysis

What Drives the Answer?

Three inputs tested. Two dominate. One is negligible.

Transfer Capacity (CETL)

#1 driver. HVDC reduces mean hours 7x (544→78). But only 46% of stochastic scenarios reliable with HVDC.

Sensitivity rank #1

Fleet Calibration

#2 driver. CT 4–6 GW range shifts absolute hours 3x. But 0% reliability under all calibrations at 70/30.

Sensitivity rank #2

Demand Noise: Negligible

±1% to ±7% changes mean hours by <10%. Spatial dynamics dominate demand uncertainty.

Negligible driver
Validation Context

How Does This Compare to Reality?

Sanity check: the model says 3 GW DC needs +3 GW transfer costing $4.8B. Dominion is building exactly that. This confirms consistency with Dominion's own capacity adequacy assessment — not an independent prediction.

Limitations

What the Model Doesn't Capture

Zonal Dispatch Only

No intra-zone transmission congestion modeled. The zonal DCOPF assumes load and generation mix freely within Dominion. Real grid has internal bottlenecks that matter at high concentrations — actual breaking points may be lower than modeled.

All questions

Zone Fractions Estimated

The 70/30 DOM/non-DOM split is calibrated to historical load data. Counterfactual fractions (what if load were redistributed?) are approximated, not modeled from first principles. Redistribution scenarios are directional, not precise.

Q1, Q3

ERCOT Excursion Unvalidated

Texas zone fractions are illustrative estimates calibrated to public load data, not to ERCOT's dispatch model. The ERCOT result (concentrate in DFW) is directionally informative — it shows market structure changes the answer — but the specific numbers are not independently validated.

Excursion

MC Tail Risk Uncertainty

200 draws are adequate for central estimates and p50–p75 reliability metrics. Exceedance probabilities beyond p90 carry ±3–5% sampling uncertainty. The finding that “62% of scenarios reliable at baseline” is robust; exact tail counts are not.

Q1 stochastic
Recommendations

What the Analysis Supports

Five takeaways from the grid analysis. Each follows directly from the model results.

  • 01

    Cap concentration at 10–20% of zonal load

    Q3's threshold analysis shows DOM can absorb 10–20% of its load as data centers before reliability degrades. Today's 70% share wasn't planned — it accumulated. Applying that fraction to a growing load base creates a reliability problem that no economical transmission investment can solve after the fact.

  • 02

    Plan transmission before load growth, not after

    The $4.8B HVDC investment required to support 3 GW in Dominion is a grid-planning decision with a decade-long lead time. By the time any individual developer identifies the reliability bottleneck, the planning window has already closed. Transmission investment must be coordinated ahead of load commitments — not triggered by them.

  • 03

    Require stochastic reliability in capacity adequacy assessments

    Deterministic: 6 hours unserved per year at baseline — within conventional adequacy standards. Stochastic: only 62% of scenarios reliable at baseline. These two numbers come from the same model. Using only the deterministic result passes current standards while hiding the material tail risk that stochastic analysis reveals.

  • 04

    Don’t apply PJM findings to other markets

    ERCOT's market structure — different topology, no capacity market, higher renewable penetration, no firm import rights — produces the opposite optimal siting recommendation: concentrate in DFW rather than distribute. There is no universal answer. Site-specific modeling is required for each market.

  • 05

    Treat the methodology as the transferable asset

    The compound return from sequencing spatial and stochastic resolution — same dispatch engine, new question, fundamentally different answer — is the pattern to replicate. Any grid market with concentrated load growth faces this same analytical sequence. The specific numbers are PJM-specific; the method is not.

Data Sources

Reproducible with Public Data

All inputs are publicly available. Demand data is real PJM historical load; no synthetic or assumed demand profiles.

  • Primary Demand Data PJM DADS (Day-Ahead Demand and Supply), 2014–2022. 87,672 hours of real DOM zone load used for dispatch model calibration and stochastic weather sampling.
  • Reliability Standard NERC 2023 Long-Term Reliability Assessment. LOLE and EUE adequacy standards used to define the reliability threshold (1 day/10 years target).
  • Validation Reference Dominion Energy Virginia 2023 Integrated Resource Plan (IRP). The $4.8B HVDC cost figure and 3 GW transfer capacity design are consistent with Dominion’s own published capacity adequacy analysis.
  • Load Growth Context Virginia State Corporation Commission (SCC), Annual Report on Data Center Power Demand in Northern Virginia, 2023. Contextualizes the 70% Dominion zone fraction as a reflection of current data center geography.
  • ERCOT Excursion ERCOT Nodal Protocol Load Zone Definitions and 2022 Load Forecast. Used to construct illustrative zone fractions for the cross-market excursion. Not independently validated.
  • Market Design Context PJM State of the Market Report 2023 (Monitoring Analytics). Capacity auction results and data center load growth projections used for scenario framing.

3 questions, 1 excursion, 200 stochastic draws, 9 weather years, 87,672 hours real DOM demand, 3 sensitivity analyses, PJM + ERCOT.