Grid interconnection, not construction, now sets the schedule for data center delivery. A data center building is erected in about a year, while the median successful grid connection completed in 2025 took 5.6 years and most requests never connect at all. The findings below summarize 26 years of United States interconnection queues. The Method tab documents the estimation; the Risk Planner evaluates the fitted model for a configuration you choose.
The demand landscape and the state of the queues
The map below places the 1,477 physical data center facilities of the IM3 atlas, sized by floor area, over the median time that active requests in each state have already been waiting. Proximity to concentrated demand confers no completion advantage once region and technology are held fixed: the level of delivery performance is set by regional market membership, and the decline is present everywhere. Hover any state for its queue statistics.
The demand shock and the delivery decline
The figures below report capacity entering United States queues by technology and the median request-to-operation time among completed projects. Capacity seeking connection nearly tripled after 2020 and its mix turned toward natural gas, the firm-power signature of data center load. Over the same period the median time from request to commercial operation rose from 1.4 years for 2005 completions to 5.6 years for 2025 completions. An exact least squares segmentation of the duration series dates the most recent shift to 2021Q3, before the 2023 demand surge and before FERC Order No. 2023.
Capacity entering U.S. queues, by technology
Gigawatts per entry year · hover for the mix
Median request-to-operation time, completed projects
By completion year · band = interquartile range · horizontal steps = segment means
Cumulative incidence of completion and withdrawal
Every request ends in one of two ways, commercial operation or withdrawal, and either outcome removes the project from the queue, so the chart below estimates the two as competing risks by entry cohort. Five-year completion falls from 26.2% for entrants of 2000 to 2007 to 6.5% for entrants of 2018 to 2020, and the 2021 to 2022 cohort reaches 0.6% by year three. Withdrawal is the modal outcome in every cohort. Curves stop where fewer than 100 projects remain at risk.
Study costs and the withdrawal mechanism
Merging interconnection study costs onto the queue identifies the mechanism behind withdrawal. In the left figure, projects allocated top-quartile network upgrade costs withdraw at 2.9 times the rate of the cheapest quartile, and only 5.7% of them reach operation within three years of the study. The right figure tracks the escalation of the allocated costs themselves. Because the allocation is not observable when the request is filed, abandonment can be diversified with parallel queue positions but it cannot be forecast at entry.
Exit hazards by allocated-cost quartile
vs. cheapest quartile · bars = 95% CI · late-entry Cox, n = 3,247
Median allocated network cost by study year
2024 $ per kW · six reporting territories
Highlights of the findings
Six results organize the contribution of the study. Each links to the exhibit, the model, or the source behind it.
1Entry timing is a quantified schedule risk. Each additional year of later queue entry reduced the completion hazard by 6.7 percent, a single number that summarizes a 25 year deterioration and survives every control, stratification, and imputation test in the paper. It converts a diffuse institutional trend into a figure an owner can put directly into a schedule risk register; the estimation framework describes where it comes from.
2Completion has collapsed across entry cohorts. The five-year probability of reaching commercial operation fell from 26.2 percent for entrants of 2000 to 2007 to 6.5 percent for 2018 to 2020, and the 2021 to 2022 cohort stands at 0.6 percent by year three. Delivery capability contracted exactly as demand tripled, which is why interconnection, not construction, is now the binding constraint for large load programs. The Risk Planner evaluates any specific configuration.
3The slowdown predates the AI demand surge. Exact least squares segmentation places the latest regime shift in completion times at 2021Q3, before the 2023 wave of data center requests and before FERC Order No. 2023. Accounts that attribute current durations to the AI shock therefore mis-date the problem, and evaluations of queue reform need a baseline that was already deteriorating.
4Proximity to demand buys nothing; region buys a lot. Projects sited near the 1,477 facilities of the IM3 atlas complete no faster once region and technology are held fixed, while regional membership moves five-year completion from 0.18 in ERCOT to roughly zero in ISO New England. For siting, market membership, not distance to load, is the delivery variable. The map shows both layers together.
5Withdrawal is priced, and priced late. A top-quartile network cost allocation multiplies the withdrawal hazard by 2.88 and cuts the completion hazard to 0.56, and the allocation is revealed only at the study stage, years after filing. Much of what looks like speculative attrition is a rational response to late-arriving prices, which is why the cost identification supports diversifying abandonment risk with parallel queue positions rather than trying to forecast it.
6The queue mix records the procurement strategy. The 2025 entering class totals 513.6 GW, including 151 GW of natural gas, the largest fossil inflow recorded in the Berkeley Lab queue data. The turn toward firm power for large loads is visible in the composition of the queue itself, before it appears in any capacity statistic.
Method
This page summarizes the estimation pipeline behind the site. The paper gives the full specification; the supplemental materials report every estimate, and a machine readable planning grid accompanies both.
The estimation pipeline, interactive
Data sources
The core dataset is the Lawrence Berkeley National Laboratory Queued Up compilation of interconnection requests, which covers regions responsible for roughly 98 percent of installed United States generating capacity. The analysis joins it with the laboratory's interconnection cost series, the IM3 Open Source Data Center Atlas from Pacific Northwest National Laboratory, the Census Bureau Value of Construction Put in Place series, and EIA Form 861 sales data.
Sample construction
From 38,201 raw queue records, sequential exclusions remove duplicate and non-generation entries, records outside the 2000 to 2025 entry window, and records missing the dates that define the outcome. The analysis sample holds 26,695 requests: 2,871 completed, 14,870 withdrawn, and 8,954 still pending at censoring. Sensitivity analyses re-enter the excluded records under three imputation scenarios and reproduce every cohort ordering.
Competing risks estimation
Completion and withdrawal are competing exits from the queue, so the analysis estimates them jointly: Aalen-Johansen cumulative incidence curves, truncated where fewer than 100 projects remain at risk; cause-specific Cox proportional hazards models; Fine-Gray subdistribution hazards; and log-logistic accelerated failure time models, selected over the Weibull form by information criterion. Concordance is 0.844 for the completion model and 0.633 for withdrawal, which is itself a finding: what makes projects succeed is far more systematic than what makes them quit.
Changepoint segmentation
Regime shifts in the quarterly duration series are located by exact least squares segmentation solved with dynamic programming, which evaluates every admissible break quarter rather than a coarsened search grid. Three breaks are identified, in 2008Q3, 2013Q3, and 2021Q3, and the last of them predates both the 2023 demand surge and FERC Order No. 2023.
Predictive benchmark and temporal validation
The classical models are benchmarked against random survival forests, gradient boosting, and a deep discrete-time competing-risks network in the DeepHit family. Validation is strictly temporal: every model trains on entries through 2017 and is scored on entries from 2018 onward, so no model sees the future it predicts. A ten-seed ensemble of the selected architecture reaches a time-dependent concordance of 0.813 (90 percent interval 0.796 to 0.828) for completion and 0.599 for withdrawal. The Risk Planner on this site evaluates that ensemble refitted on all 26,695 requests.
Cost mechanism identification
Project-level interconnection study costs from six reporting territories are merged onto the queue (3,247 late-entry requests) and enter cause-specific Cox models as allocated cost quartiles. A top-quartile allocation multiplies the withdrawal hazard by 2.88 (95 percent interval 2.50 to 3.32) and the completion hazard by 0.56 (0.43 to 0.73), identifying late-revealed network upgrade costs as a central mechanism behind abandonment.
Reproducibility and scope
Analysis code, derived data, and the 1,350-row planning grid behind the Risk Planner accompany the paper. All estimates describe generation interconnection between 2000 and 2025; dedicated large-load interconnection processes for data centers are newer and may differ. Estimates are research outputs provided for planning context only.
Delivery-risk planner
The planner evaluates the deep competing-risks ensemble described in the Method page at a configuration you choose: market region, technology, entry year, and capacity class. Solid curves are ensemble means, shaded bands are 90 percent ensemble intervals, and the grey dashed curve is the competing probability of withdrawal. The right panel re-evaluates the same project in all nine regions.
Probability of holding power
solid = ensemble mean · band = 90% interval · grey = withdrawal
Same project, every region
P(commercial operation within 5 years) · ▲ = your pick
Applying the estimates
For schedule construction the paper reduces to a set of multipliers on a mid-2010s baseline: about 2.1× on expected time in queue for entry in 2021 to 2022, about 1.25× per unit of log capacity, more than 5× for ISO New England and 0.63× for ERCOT relative to MISO. Screening the region and technology surface for combinations that have historically delivered, and holding parallel queue positions, addresses the two distinct risks the analysis separates: delay, which can be estimated, and abandonment, which cannot be forecast at entry.
Estimates describe generation interconnection under 2000 to 2025 conditions and are provided as decision support, not as a guarantee of outcomes.
About
This site is the interactive companion to Grid Interconnection as an Upstream Schedule Constraint in Data Center Delivery: Evidence from U.S. Generation Queues, a manuscript currently under anonymized peer review. Author and contact information will be added when the review concludes.
Data. Lawrence Berkeley National Laboratory Queued Up (2026) and interconnection cost series; IM3 Open Source Data Center Atlas (PNNL); United States Census Bureau Value of Construction Put in Place; EIA Form 861.
Scope. Estimates are research outputs provided for planning context only. They describe generation interconnection between 2000 and 2025; dedicated large-load interconnection processes are newer and may differ. Nothing on this site is engineering, legal, or investment advice.
dtcenter.electriai.com