When a renewable energy developer pulls a project from the grid interconnection queue, the costs that project was supposed to cover don't vanish. They get redistributed to the remaining projects in the study cluster. Some of those projects, now facing higher costs, withdraw too. Their costs reallocate again. A feedback loop begins. Zahra Heidari at Colorado State University argues that this withdrawal-restudy cascade is not a side effect of slow processing. It is an architectural vulnerability of the U.S. interconnection queue itself, independent of how fast studies are completed. Drawing on contagion models from financial crises and interdependent infrastructure failure, the paper shows that the queue system can amplify a 10% shock into near-total collapse under certain connectivity conditions, and that the phase transition depends on how costs are allocated.
The bottleneck that holds back clean energy deployment
The U.S. interconnection queue is the gateway through which every new generator and storage project must pass to connect to the transmission grid. As of year-end 2025, roughly 8,200 projects totaling 2,061 GW of capacity sit in the queues across seven major ISO/RTOs. That queued capacity exceeds the entire installed U.S. generating base of about 1,280 GW, a figure the country took decades to build. Solar alone accounts for 773 GW of queued generation and 749 GW of storage, wind contributes 220 GW, and natural gas surged 86% year-over-year to 253 GW, partly driven by hyperscale AI data center demand. Zero-carbon capacity makes up about 77% of the queued generation portfolio.
But the queue doesn't deliver. Only 13% of the capacity queued between 2000 and 2020 has reached commercial operation. By project count, the completion rate is about 19%. Another 76% was withdrawn, and the remaining 10% is still waiting. The median time from interconnection request to commercial operation has more than doubled, from under two years for projects that came online between 2000 and 2007 to 5.2 years for those reaching operation in 2024-2025. PJM's capacity auction costs jumped from $2.2 billion to $14.7 billion between the 2024/2025 and 2025/2026 delivery years, driven in part by generator retirements and a congested queue that limits new resource entry.
FERC Order 2023 restructured the process from serial, first-come-first-served to cluster-based, first-ready-first-served, with stricter financial readiness requirements and penalties for missed deadlines. These reforms address processing speed. They do not address the structural feedback loop that causes withdrawals to cascade.
Four mechanisms that make the queue fragile
The paper identifies four coupled failure propagation mechanisms. The first is withdrawal-restudy coupling: when a project withdraws, the network-flow assumptions underlying interdependent studies no longer hold, and the transmission provider must restudy the remaining projects. The scope of restudy is determined by electrical interdependence, not administrative boundaries. A withdrawal in one cluster can trigger restudies in another through shared transmission constraints.
The second is cost concentration risk. Restudies frequently reveal that upgrade costs previously shared among many projects now redistribute among fewer remaining projects. That reallocation can turn previously viable projects into uneconomic ones. The cost differential tells the story: withdrawn projects average $373/kW in assessed interconnection costs, while completed projects average $73/kW, a factor of roughly five.
The third is temporal clustering. Withdrawals don't happen uniformly over time. They concentrate around study milestone dates, restudy completion announcements, and financial commitment deadlines. When developers receive new cost information, many decide simultaneously. This concentration creates shocks that exceed the absorption capacity of affected projects, amplifying cascade propagation.
The fourth is the absence of circuit-breakers. Financial markets halt trading during cascading sell-offs. No equivalent mechanism exists in the interconnection queue. When a cascade begins, it runs until the system reaches a new equilibrium, potentially after withdrawing a substantial fraction of affected projects.
Grid operators have described these dynamics in their own regulatory filings. MISO stated in its 2021 queue reform filing that withdrawals "change the underlying assumptions in the interconnection studies conducted for lower-queued projects. The resulting restudies, which often have a cascading effect, impair the ability to administer the queue in a timely fashion." SPP attributed multi-year delays to customers "withdrawing projects from the queue late in the game", noting that "such withdrawals can shift the responsibility for upgrades to lower queued customers, requiring cascading restudies."
Empirical evidence: withdrawals are not random
Using LBNL Queued Up 2026 data covering 38,201 project records across seven ISO/RTOs, Heidari tests whether withdrawals behave like independent random events or like correlated failures. The analysis covers 16,206 withdrawals, with withdrawal dates available for 86.3% of them, and spans 2010 through 2025.
The dispersion index measures whether monthly withdrawal counts are more variable than a Poisson process would predict. A value of 1 means random. The actual values are staggering: CAISO at 102.4, ISO-NE at 68.1, MISO at 67.5, PJM at 62.9, SPP at 34.3, ERCOT at 6.8, NYISO at 5.0. All reject the random null hypothesis at p < 0.001. Even after removing the three largest months per region, every dispersion index remains well above unity.
The paper identifies 39 monthly withdrawal bursts, where capacity withdrawn exceeds the regional mean by more than two standard deviations. The largest hit 67.4 times the regional mean: CAISO in December 2024, when 420 out of 426 records carry the date 2 December and 303 belong to a single Cluster 15 identifier block totaling 132.1 GW. PJM's December 2024 burst reached 18.1x the mean with 43.4 GW across 485 projects. MISO's December 2025 burst reached 17.1x the mean with 42.0 GW, distributed across four burst months during the implementation of reformed queue procedures. The distribution of monthly withdrawal ratios is heavy-tailed, deviating substantially from exponential in the upper tail.
A 1,000-permutation test checks whether withdrawal timing concentrates within technology categories. The z-statistics range from 2.58 to 4.78 across all seven regions, with p <= 0.005 everywhere. Solar projects tend to withdraw near other solar projects. Wind projects withdraw near wind projects. The effect is universal across ISO/RTOs.
Co-withdrawal within cohorts (projects entering the queue in the same region, same technology, within a six-month window) is significant in four of seven regions under the entry-year-stratified null, with excess co-withdrawal of 0.5 to 1.8 percentage points. The effect sizes are modest but real, confirming that correlated exits occur beyond what entry-timing trends alone would explain.
The 2025 withdrawal wave stands out: 2,417 project withdrawals totaling about 398 GW, roughly 3.5 times higher by project count and 3.8 times higher by capacity than the 2010-2024 annual average. MISO alone saw 171.7 GW of withdrawals, with 146.4 GW concentrated in four burst months. Across all seven regions, December 2025 withdrawals reached 74.0 GW. This magnitude corresponds to the model's calibrated scenario for a 24.9% shock, which yields 38.5% total system failure.
Contagion modeling: when does a queue become a cascade?
To move beyond statistical correlation, Heidari builds a computational contagion model adapting threshold cascade theory from Watts (2002) and interdependent-network percolation from Buldyrev et al. (2010). The queue study cluster is represented as an Erdos-Renyi random graph with 500 nodes. Each node is a project with a base cost load drawn from U(0.3, 0.6) and a viability threshold from U(0.55, 0.95). Initially, about 1% of nodes are in failure.
An initial shock removes a fraction of nodes. Each failed node's cost load is redistributed to its active neighbors, with each neighbor receiving a fraction alpha = 0.08 of the failed node's load. This is not divided among neighbors. Each active neighbor absorbs the full transfer. The total load introduced per failure is k times alpha times the load, where k is the average network degree. When k times alpha exceeds 1.0, each failure introduces more load than it removes, creating a self-sustaining regime.
At low connectivity (k = 3, k times alpha = 0.24), a 10% shock produces 11.7% total failure, net amplification of 1.06x. At moderate connectivity (k = 10, k times alpha = 0.80), a 10% shock yields 15.3% total failure, 1.38x amplification. The system is stressed but stable.
At high connectivity (k = 20, k times alpha = 1.60), something qualitatively different happens. Even the zero-shock baseline is unstable, bimodally distributed between negligible failure and near-total collapse with a mean of 10.8% and standard deviation of 18.7%. A 5% shock produces 95.8% failure. A 10% shock produces 97.7% failure, an amplification of 8.69x. The system has crossed a phase-transition boundary from progressive failure to categorical fragility. Any perturbation, however small, can trigger systemic collapse.
The phase transition is a boundary condition of the specific cost allocation rule. Under pro-rata redistribution, where costs are divided proportionally among neighbors rather than transferred undivided, the same k = 20 network produces only 11.1% total failure and 1.00x amplification. The phase transition disappears. Systemic instability depends on whether individual projects must absorb undiluted reallocated costs.
Circuit-breakers: partial mitigation through cost caps
The model tests circuit-breaker interventions that cap the maximum cost transfer per neighbor at a value beta. At k = 10 with a 10% shock, a cap of beta = 0.04 reduces total failure from 15.3% to 14.6% (5% reduction). Beta = 0.03 reduces it to 13.4% (13% reduction). Beta = 0.02 reduces it to 12.3% (20% reduction). The mean per-neighbor transfer drops by 44% under the strictest cap. The relationship between cost-transfer reduction and cascade-size reduction is sublinear: a 44% reduction in transfers yields only a 20% reduction in failures. Circuit-breakers help, but they don't eliminate the vulnerability.
Under the strictest cap, net amplification decreases from 1.38x to 1.08x. The system remains in the amplification regime but approaches the boundary of stability.
The AI data center dimension
The paper draws attention to the intersection of queue fragility and AI infrastructure demand. Hyperscale data centers require 100 MW to 1 GW of reliable power per site, and the national development pipeline exceeds 100 GW. Each facility enters the interconnection queue through collocated or contracted generation. Between 2024 and 2025, more than a gigawatt of data center load disconnected from the grid within seconds. NERC escalated from a Level 2 Alert on large loads in September 2025 to a Level 3 Essential Action Alert in May 2026, only the third Level 3 Alert in NERC's 58-year history. NERC initiated "Project 2026-02" and introduced a mandatory registration category for large computational loads.
In June 2026, FERC issued show cause orders under Section 206 of the Federal Power Act to all six jurisdictional ISO/RTOs, requiring a 30-day informational report on resource adequacy for large loads and a 60-day filing defending or proposing tariff reforms. Five operators requested suspension of the 60-day deadline; FERC capped extensions at 90 days. The regulatory machinery is moving, but the structural analysis in this paper suggests that accelerating study throughput alone won't prevent cascading failure. Large-load priority lanes may even redirect analytical resources away from standard queue processing, extending timelines for non-priority projects and increasing their attrition risk.
What resilience engineering would actually require
The paper frames the queue as a complex adaptive system that satisfies all five defining criteria: heterogeneous agents with different objectives and risk tolerances, nonlinear interactions where a single withdrawal can trigger disproportionate cost reallocations, adaptation as developers modify their strategies in response to observed dynamics, emergence as the 76% withdrawal rate cannot be explained by individual rationality alone, and path dependence as the queue's trajectory is shaped by its history of prior studies and allocations.
The resilience engineering prescription is not a single reform but a shift in perspective. Four recommendations: transmission providers should model and report dispersion and cascade-amplification metrics, analogous to Dodd-Frank stress tests for financial institutions. Queue operators should publish standardized cascading risk metrics. NERC or FERC should establish an interconnection resilience standards body. And research support should flow through DOE grid modernization, NSF, or ARPA-E to develop the formal models the system currently lacks.
The paper also notes that the vulnerability is not inherent to market-based electricity systems. Alternative architectures, including integrated resource planning, centralized siting, and coordinated transmission-and-generation planning, eliminate the cost-sharing interdependencies that create the contagion mechanism. The fragility is a consequence of how most ISO/RTOs currently process interconnection requests, not a fundamental property of competitive electricity markets.
With projected generator retirements totaling 115 GW between 2025 and 2034, resource additions falling short of industry projections for two consecutive cycles, and capacity leaving the queue faster than new capacity can arrive, the structural bottleneck described here is not an academic exercise. It is a system that processes thousands of simultaneous requests from intermittent resources, storage systems, and hybrid facilities without any model of how failures propagate through it. The queue was designed for a different era. It now behaves like a complex system susceptible to systemic failure, and the analysis in this paper provides both the framework and the evidence that the architecture itself needs redesign, not just faster processing.