Skip to content
Studies · Case Study

Can PJM Handle Virginia's Data Centers?

PJM serves 67 million people across 13 states. Capacity market prices went from $29 to $333/MW-day in three years. The latest auction had a 6,623 MW shortfall — the first in PJM's history. Data centers account for 94% of projected load growth. We asked five questions in sequence — each answer built on the last.

The simulation ran once. Every downstream question built on what it produced.

153 GW
Peak Demand
30 GW
DC Projected by 2030
848 TWh
Annual Generation (2024)
$333
/MW-day Capacity Price
The Analysis

Five Models, Five Different Answers

Five analyses of the same grid. The only thing that changes is how much reality you include. The “safe” data center capacity drops from unlimited to zero. The investment recommendation goes from nothing to impossible.

Level Method Answer Investment Recommendation
Screening Annual energy balance No limit None needed
Deterministic dispatch 8,760-hour simulation, 2024 profiles 19 GW +12 GW gas ($13B)
9-year Monte Carlo Same dispatch, 9 weather years 14–24 GW range +20 GW gas ($22B)
200-draw stochastic MC Weather + forced outages 5 GW at P90 +20 GW gas achieves only 42% reliability
Stress scenario Dec 2022 profiles + confirmed outage rates Grid fails at 0 GW DC No feasible investment prevents failure at Elliott severity

Each level uses the same underlying dispatch engine. The only changes are inputs: temporal resolution, weather sample size, outage modeling, and scenario selection.

Fidelity Lesson

What Changed at Each Step

Each escalation doesn’t just refine the number — it changes the nature of the answer. The question itself shifts.

Screening → Deterministic Dispatch
“No problem” → “19 GW”

Annual energy balance hides temporal mismatch. There’s enough energy in aggregate — just not at 6 PM on August 14th. Hourly dispatch reveals the gap.

Deterministic Dispatch → 9-Year Monte Carlo
19 GW → 14–24 GW range

Weather year selection swings the answer by 10 GW. A single year gives a single number. Nine years reveal the range — and show that planning to the best year is gambling.

9-Year MC → 200-Draw Stochastic MC
“100% reliable with +20 GW gas” → “42% reliable”

Forced outages are correlated with stress. When temperatures spike, generators trip — exactly when you need them most. The 9-year MC assumed all capacity was always available. It wasn’t.

Stochastic MC → Stress Scenario
“Needs investment” → “Already failing”

The grid’s vulnerability is structural. Winter Storm Elliott profiles with confirmed outage rates show the grid fails with zero data center load. The problem isn’t future demand — it’s current fragility.

Step What Changed Dollar Impact
Screening → Dispatch Revealed the problem exists $0 → $13B+
Dispatch → 9-Year MC Increased investment requirement by 67% +$9B
9-Year → Stochastic MC Showed +20 GW gas still inadequate Investment exceeds gas alone
Stochastic → Stress Reframes the problem entirely From “build more” to “build differently”
The Cost of Low Fidelity
$9 billion. That’s the cost of trusting the deterministic answer instead of running the Monte Carlo. 200 draws. 10 minutes of compute. Changed a $13B investment recommendation by 67%.
Sensitivity Analysis

What Drives the Answer?

Across four investigations, three inputs dominate. One widely-discussed solution has zero effect.

Coal Retirement Timing Critical

When coal plants close determines whether the grid has enough dispatchable backup to survive peak stress events.

558 hrs of reliability impact
Gas Forced Outage Rate High

Correlated with temperature extremes. The generators most needed are the ones most likely to fail.

480 hrs of reliability impact
Demand Response High

Flexible data center load and grid-wide DR programs can absorb peaks that would otherwise cause blackouts. 1 GW = $122M/year in avoided capacity costs.

109 hrs of reliability impact → See Investigation 05
Wind & Solar at Current Levels Moderate

Adds energy but not reliable capacity. Helps in average conditions, absent during stress events.

Moderate reliability effect
Battery Storage Zero

At current deployment levels, battery storage contributes effectively zero to multi-day reliability events. The duration problem — hours of storage vs. days of stress — makes batteries irrelevant to the capacity question.

0 hrs of reliability impact at current scale

The parameters that don’t matter are as important as the ones that do. Battery storage is the most-discussed solution in energy policy. In this analysis, it contributes zero hours of reliability improvement. Coal retirement timing — rarely discussed in the data center debate — swings reliability by 558 hours. The sensitivity analysis tells you where to look — and where to stop looking.

Validation

Seven Findings, Seven Checks

Each model prediction compared against publicly available real-world data.

Model Finding Grid breaks at ~19 GW of data center load
Real-World Data 2027/28 BRA: 6,623 MW shortfall at only 5 GW DC load
Validated
Model may be optimistic — real shortfall appeared at far lower DC load than predicted
Model Finding Planned buildout makes it worse — retirements outpace additions
Real-World Data PJM Energy Transition Report: 40 GW retirements, only 5% queue completion rate
Validated
Model Finding Gas EFOR: 32% during extreme cold
Real-World Data Actual Winter Storm Elliott: 37% (FERC/NERC confirmed)
Conservative
Model understates real gas failure rates by 5 percentage points
Model Finding Coal EFOR: 15% during extreme cold
Real-World Data Actual Winter Storm Elliott: ~16%
Accurate
Model Finding 30 GW data center load by 2030
Real-World Data PJM 2025 Load Forecast: 32 GW, 94% driven by data centers
Confirmed
Model Finding Coal retirements reduce firm capacity faster than replacements arrive
Real-World Data 12–25 GW at risk by 2028–2030 per multiple independent analyses
Validated
Model Finding Monte Carlo variability: 10 GW range in effective capacity
Real-World Data Reserve margin swung 5.5 percentage points in three consecutive auctions
Consistent

Elliott Confirmation

Winter Storm Elliott (December 2022) was the real-world stress test for our outage rate assumptions. Our model used 32% gas forced outage rate and 15% coal forced outage rate during extreme cold events.

FERC and NERC's joint inquiry confirmed the actual numbers: 37% for gas and ~16% for coal. Both were worse than what our model assumed.

The model is not pessimistic — it’s optimistic. Real gas outage rates are higher (37% vs. 32%), the capacity shortfall appeared earlier (at 5 GW, not 19 GW), and queue completion rates (5%) are far below what the planned fleet scenarios assume. Every assumption we made was generous to the grid. Reality is worse. The conclusions from this study should be read as lower bounds on the actual risk.

Capacity Auction Trajectory

PJM's Base Residual Auction results tell the story in dollars and megawatts.

Delivery Year Clearing Price Reserve Margin Target Shortfall
2024/25 $28.92/MW-day 20.4% ~17–20% None
2025/26 $269.92/MW-day 18.5% 17.8% None (barely)
2026/27 $329.17/MW-day 18.9% 19.1% 309 MW
2027/28 $333.44/MW-day 14.9% 20.0% 6,623 MW
The Numbers
Prices increased 10x in three years. Data centers responsible for 63% of the price increase ($9.3B in capacity costs). Source: PJM Independent Market Monitor.
Limitations

What This Doesn't Model

Every model has a boundary. These are ours.

No Intra-Zone Transmission

The model treats each PJM zone as a single node. Transmission constraints within zones — which can strand capacity or strand load — are not captured. Local reliability problems may be worse than the system-level numbers suggest.

Queue Completion at Historical Rates

New generation is modeled at the current 5% interconnection queue completion rate. If permitting or interconnection reform accelerates, the capacity shortfall timeline shifts. The model is a baseline, not a forecast.

No Market Behavior Response

Generators and data center operators are treated as price-takers. In reality, the 10x price increase in capacity markets will trigger investment and behavioral responses that the model doesn’t capture. This is a structural stress analysis, not an equilibrium model.

No Compound Multi-Week Events

Weather scenarios are drawn from historical records. A compound event — two consecutive polar vortex weeks, or a summer heatwave following a wet spring that fills reservoirs and limits hydro — is not represented. Tail risk is likely understated.

Storage and DR at Current Deployment

Battery storage and demand response are modeled at 2024 deployment levels. Future buildout scenarios are not included. The “storage = zero” finding is specific to the current fleet, not a statement about storage at scale.

Scheduled Retirements as Given

Coal and gas retirements follow announced schedules. Regulatory delays, market signals, or policy changes could accelerate or defer retirements in ways the model doesn’t anticipate.

Recommendations

What the Analysis Supports

Five conclusions grounded in validated, bounded analysis. Each is traceable to a specific finding.

  • 01

    Gate coal retirements on firm replacement capacity

    Coal retirement timing is the #1 sensitivity driver (558 hours of reliability impact). Retiring before firm replacement is in service is the single largest avoidable risk in the current grid transition. This is not an argument against retiring coal — it is an argument for sequencing.

  • 02

    Require and verify gas winterization

    FERC mandated winterization after Elliott. Real gas EFOR during Elliott was 37% — 5 points above our model assumption, and above what FERC’s own post-Elliott standards require. Mandate compliance monitoring with actual performance data, not self-certification.

  • 03

    Enable data center interruptibility as a grid resource

    Interruptibility is at least 5x cheaper than gas in every weather year tested. 18% DC curtailment for 1.2% of operating hours is an operational constraint that hyperscale operators can manage. The policy and market structures to formalize this don’t yet exist at scale.

  • 04

    Don’t count battery storage toward near-term reliability targets

    200 GWh of storage had zero reliability impact in this analysis. Storage requires adequate renewables to charge from; at current grid composition, the correlation problem (generation shortfalls happen when storage is already depleted) makes batteries irrelevant to the multi-day reliability question. Plan storage as energy arbitrage, not reliability insurance.

  • 05

    Plan to the stochastic distribution, not the deterministic answer

    The deterministic answer (19 GW) underestimates the investment requirement by $9B. Planning authorities, interconnection policy, and capacity market design should all use stochastic analysis as the baseline standard. The cost of the Monte Carlo is 10 minutes of compute. The cost of skipping it is $9 billion.

References

Sources

All data comes from publicly available regulatory filings, market reports, and independent analyses.

  • 1PJM Interconnection — Base Residual Auction Results (2024/25 through 2027/28 delivery years)
  • 2FERC/NERC — Joint Inquiry into Winter Storm Elliott (2023), generator forced outage analysis
  • 3PJM Interconnection — Energy Transition in PJM: Frameworks for Analysis (2023), retirement and queue completion data
  • 4PJM Independent Market Monitor (Monitoring Analytics) — State of the Market Report, capacity cost attribution
  • 5Grid Strategies — The Era of Flat Power Demand is Over (2023), data center load growth projections
  • 6Institute for Energy Economics and Financial Analysis (IEEFA) — PJM capacity market analysis
  • 7PJM Interconnection — 2025 Load Forecast Report, data center demand projections (32 GW, 94% DC-driven)
  • 8PJM Interconnection — Effective Load Carrying Capability (ELCC) and Reserve Requirement Study

5 questions, 5 fidelity levels, 200 stochastic draws, 9 weather years, 26 data files, 13 Python modules, validated against PJM auction data and FERC/NERC reports.