University paper confirms fixing the rules frees enough electricity to cover years of US AI growth using grid connection flexibility. FERC is making it happen. Several other sources confirm this conclusion. Duke figures are based on past experience. I assume if flexible connection is widely adopted the effective peak will be lowered and a higher percentage of time will be required for the same result. I’ve reflect that in my reporting.
Important facts
Claude reminded me to clarify the different numbers. The data center load is measured in gigawatts and is close to 100% these days. Solar and wind are measured at peak gigawatts, rarely achieved. The continuous flow is between 15% & 50% of the peak, depending on things like the amount of battery backup. That’s why my headline number is less than the numbers in the original article.
The authors pulled nine years of hourly load data (2016–2024) from EIA’s Hourly Load Monitor across 22 balancing authorities — seven RTOs/ISOs, eight non-RTO Southeastern utilities, seven non-RTO Western utilities — covering 744 of roughly 777 GW of US summer peak load, or 95% of the total. Aggregate load factor (average demand divided by peak demand) across these systems ranges 43% to 61%, average and median 53%. Winter load factors run lower than summer: average 59% versus 63%.
Using a goal-seek function against nine years of hourly data, the authors calculate how much constant new load each balancing authority could add before pushing total demand past its historical peak, given the new load agrees to curtail itself as needed. At a 0.25% annual curtailment limit — the new load curtailed for about 85 hours a year, averaging 1.7 hours per event — the 22 systems combined can absorb 76 GW. At 0.5% curtailment, 98 GW. At 1.0%, 126 GW. At 5.0%, 215 GW.
Headroom isn’t evenly distributed. At 0.5% curtailment, PJM can take 18 GW, MISO 15 GW, ERCOT 10 GW, SPP 10 GW, and Southern Company 8 GW — the five largest of the 22 systems studied. Across all systems, 88% of curtailment hours still leave the new load with at least half its power; 60% leave at least three-quarters.
The paper ties headroom directly to load factor: systems with higher seasonal load factor — meaning they already run close to peak most hours — have less spare capacity to absorb new load without curtailing it. That relationship gets stronger as the curtailment limit rises, with load factor explaining 48% of the variation in headroom at a 0.5% curtailment limit and 86% of it at 5.0%.
Curtailment scenarios
Rather than forecasting future capacity, the paper models four curtailment-tolerance scenarios — 0.25%, 0.5%, 1.0%, and 5.0% of a new load’s maximum annual potential consumption — chosen because they sit within the range of existing interruptible demand-response programs (PG&E’s and SCE’s Base Interruptible Programs, for instance, cap annual interruption at 2.0% of hours). Average curtailment event length rises with the tolerance: 1.7 hours at the 0.25% limit, 2.1 hours at 0.5%, 2.5 hours at 1.0%, and 4.5 hours at 5.0%. Curtailment hours are also seasonally lopsided — CAISO’s curtailment is 92% concentrated in winter, while AZPS’s is 92% concentrated in summer — tracking each system’s seasonal load-factor gap.
Author and credits
Tyler H. Norris, Tim Profeta, Dalia Patino-Echeverri, and Adam Cowie-Haskell, all affiliated with Duke University’s Nicholas School of the Environment and/or Sanford School of Public Policy. Published by the Nicholas Institute for Energy, Environment & Sustainability at Duke University, 2025 (report number NI R 25-01). Methodology built on EIA-930 hourly demand data with a Python (SciPy) goal-seek solver; full data-cleaning and outlier-handling process documented in the paper’s appendices.
Link
https://nicholasinstitute.duke.edu/publications/rethinking-load-growth
Summary
The paper opens with the interconnection-delay problem — queues stretching seven to ten years in some utilities, transformer lead times up two to five years since 2020 with an 80% cost increase, circuit breaker lead times up 130% year over year as of 2024 — and argues load flexibility is a faster lever than new transmission or generation. It reviews why data centers have historically opted out of demand response (SLA penalties, colocation complexity, weak financial incentives) before arguing AI training workloads break that pattern: they’re delay-tolerant and shiftable across time and geography in a way conventional enterprise workloads aren’t, citing Google’s existing carbon-aware workload shifting as a working example. The headroom analysis itself is the paper’s empirical core, followed by a discussion connecting load factor to headroom and a lengthy limitations section that the authors treat as seriously as the results.
It complicates the load-growth narrative in both directions. Against alarmist framing, it shows meaningful headroom exists without new capacity. Against the opposite instinct — that flexibility solves the problem outright — it’s explicit that this is a first-order, system-wide estimate that ignores transmission-level constraints entirely: a load could stay under the system-wide peak and still overload a local substation. It also ignores generator ramp limits, minimum up/down times, and startup constraints, and it assumes new load is a flat constant draw rather than the variable profile a real data center would have.
Conclusions
The paper lands on load flexibility as a near-term, complementary tool to supply-side investment — not a replacement for building new generation or transmission, and not, in the authors’ own framing, an argument that the US can meet load growth without either. It explicitly declines to model network-level constraints, intertemporal generator operations, or shifts in loss-of-load-expectation timing as renewable penetration changes when system stress actually occurs — all flagged as needed next steps rather than settled questions. The throughline for ratepayer fights: if a load will accept curtailment, the case for billing it as if it needs firm, always-on service gets weaker.
edited and authored by Dave with close collaboration by Claude