What Keeps the Lights On?
For 15 GW of data center load at 70% renewables, we swept 180 combinations of gas capacity (40–80 GW), battery storage (50–1,000 GWh), and battery duration (4–24 hours).
84 of 180 combinations meet the reliability standard in 2024. But 2024 was the best weather year. Monte Carlo across all ten years confirms: 60 GW gas + 300 GWh storage has 6 blackout hours in 2024 but 49 in the worst year. Only 70 GW gas + 50 GWh storage stays under the LOLE threshold in every weather year (max 8 hours).
One weather year isn't a plan. The 2024 sweep said 60 GW + 300 GWh is reliable. But 2023's weather breaks that config. Run ten weather years and you need 70 GW of gas instead of 60. That's a $12B difference, and the only thing that changed is looking at more than one year of history.
Forced outages push it higher still. Stochastic Monte Carlo (200 draws, 7–10% forced outage rates, ±3% demand noise) shows the 70 GW + 50 GWh config has a P90 of 270 blackout hours. To stay reliable at P90 you need 85 GW — 15 GW more gas than any deterministic model suggested, roughly $18B of additional investment.
Model: hourly dispatch, 180 parameter combos (single year) + 9 key combos × 10 weather years (2015–2024), demand-normalized to 2024 + stochastic MC (200 draws).
Q2's hourly dispatch model assumes plants are optimally committed before dispatch runs — that is, Security-Constrained Unit Commitment (SCUC) is solved first, then economic dispatch follows. Q6 confirms it: heuristic unit commitment (starting plants based on rules of thumb rather than solving SCUC) generates 266× more blackout hours than optimal SCUC at the same installed capacity. Q2 characterizes the investment need under optimal operations; Q6 explains why getting operations right matters as much as getting capacity right. See Q6: Unit Commitment →
Where Do Batteries Help — and Where Don't They?
Battery storage is the most-discussed grid solution. But how much do you actually need? We swept storage from 0 to 500 GWh at three gas capacity levels to find where batteries stop helping.
Power trumps energy. 200 GWh at 2-hour duration (100 GW power) gives 8 blackout hours. The same 200 GWh at 24-hour duration (8.3 GW power) gives 58 hours. Grid deficits are sharp peaks, not long plateaus, so how fast the battery can discharge matters more than how much total energy it holds.
Model: hourly dispatch, storage swept 0–500 GWh at 4h duration. Duration comparison at fixed 200 GWh.
What Does It Cost?
Deterministic Answer
Stochastic Answer (with forced outages)
Cheap and reliable don't coexist easily. 60 GW gas + 300 GWh storage costs $66B for 6 blackout hours (deterministic). 70 GW gas + 50 GWh costs $24B for 2 hours — until you run the stochastic model and get a P90 of 270 hours. Gas is still the dominant lever, but you need 15 GW more of it than any single-year model would tell you.
Costs: NREL ATB 2024. Gas $1.2B/GW CAPEX. Battery: $300/kWh (4h), $230/kWh (8h), $180/kWh (12h).
Each layer of realism added cost. Single-year deterministic: 60 GW. Multi-year weather: 70 GW. Stochastic outages: 85 GW. The grid didn't change between those runs — the only thing that changed is which uncertainties we included. $12–30B of investment rides on whether you bother to model them.