Skip to main content
PJM Data Center Study → Question 1

At What Load Does the Grid Break?

PJM has 160 GW of firm generation serving 153 GW of peak demand. Every GW of data center load is flat baseload — present at peak. The arithmetic says 19 GW. The Monte Carlo says 14–24 GW. The stochastic MC says 5 GW at P90 confidence.

Monte Carlo — 9 Weather Years

Tipping Point by Weather Year

Tipping Point (GW of DC Load) by Weather Year
Finding
The tipping point ranges from 14 GW (worst year) to 24 GW (best year). A planner using 2024 alone gets 19 GW. The 10 GW range means a single-year analysis can be off by 50%.

Each bar represents the maximum data center load PJM can absorb before exceeding the reliability standard (LOLE: 0.1 events/year) for that weather year. The dashed line marks the 9-year mean of 20 GW. The shaded band shows the full range: 14–24 GW.

Model: hourly merit-order dispatch, 9 weather years (2016–2024), demand-normalized to 2024. Tipping point = first GW where unserved energy exceeds 9 hours/year.

Stochastic Monte Carlo — 200 Draws

With Forced Outages, the Answer Changes

% Reliable Draws vs. Data Center Load (GW)
Finding
200 stochastic draws with sampled forced outages: the grid is only 93.5% reliable at 0 GW DC. At P90 confidence, it breaks below 5 GW — not 19 GW.

The deterministic model treats every generator as available at rated capacity. The stochastic model samples forced outage rates (7–10%) across 200 draws per load level. The gap between the two answers — 19 GW vs. 5 GW — is entirely explained by whether you include the possibility that generators fail when you need them most.

Model: 200 stochastic draws per DC load level, 7–10% forced outage rates, ±3% demand noise. Reliable = <9 unserved hours.

Investment Analysis

What Keeps It Reliable?

If the grid breaks, what does it take to fix it? We tested gas-only and balanced (gas + renewables + storage) investment packages across all 9 weather years at 30 GW of data center load.

Gas-only: +20 GW CCGT achieves 100% reliability across all 9 weather years at 30 GW DC load. Mean unserved energy: 0.4 hours. Worst year (2020): 4 hours.

Balanced packages fail in bad years. A mixed portfolio (+10 GW gas, +10 GW wind, +20 GW solar, +40 GWh storage) is only 67% reliable — 6 of 9 weather years. Renewables underperform in the years when the grid needs them most.

Finding
The Monte Carlo added 67% to the investment requirement. The single-year analysis said +12 GW of gas. The 9-year MC says +20 GW. The cost of trusting the deterministic answer: $9B in missing capacity.
Storage Analysis

Battery Storage: Zero Impact

200 GWh of battery storage had zero impact on reliability. The battery depletes in early demand hours and has no surplus renewable energy to recharge. Storage has zero reliability value without enough renewables to create surplus energy to charge it.

PJM currently has 19 GW of wind and solar capacity. At that penetration level, renewables produce just enough to serve existing demand — there is no surplus to charge batteries during off-peak hours. The storage sits empty when the grid needs it during peak demand. This finding is consistent across all 9 weather years and all DC load levels tested.

Fidelity Check

What Doesn't Need Simulation?

The tipping point is arithmetic: peak demand minus firm capacity. PJM has ~153 GW peak and ~160 GW firm — the gap is ~7 GW of reserves. Every GW of flat data center baseload eats directly into that margin. A napkin calculation gives you 19 GW, which matches the deterministic model exactly.

Finding
The simulation validated what a napkin calculation tells you about the tipping point. It earned its keep on the investment mix question — where renewables and storage interact with demand temporally and weather variability determines whether a balanced portfolio survives.

The rule: don't simulate what you can calculate. The deterministic tipping point doesn't need a model. The stochastic reliability curve does. The investment portfolio question — where gas, renewables, and storage interact across weather years — is where the simulation actually matters.