The queue is not failing at the same rate it always did. It is failing worse.
The queue is not failing at the same rate it always did. It is failing worse.
Kenneth · 2026-08-12
*Source: Lawrence Berkeley National Laboratory, Queued Up 2026 Edition (CC BY 4.0),
complete interconnection request dataset through end-2025. Analysis my own.*
The composite PJM withdrawal rate is 68%. I asked whether that number is
deteriorating, or whether it is simply an average dragged down by a decade of
dead solar.
It is deteriorating, and not gently.
Only cohorts old enough to have mostly resolved can answer this. The right-hand
column counts terminal outcomes only.
| Filed | Withdrawn | Built | Still active | Resolved | Failure rate |
|---|---|---|---|---|---|
| 2003 | 32 | 18 | 0 | 100% | 64% |
| 2007 | 105 | 115 | 0 | 100% | 48% |
| 2010 | 361 | 81 | 0 | 100% | 82% |
| 2013 | 80 | 63 | 0 | 100% | 56% |
| 2016 | 250 | 108 | 21 | 92% | 70% |
| 2018 | 311 | 56 | 37 | 84% | 85% |
| 2019 | 409 | 75 | 140 | 73% | 85% |
| 2020 | 719 | 29 | 163 | 81% | 96% |
| 2021 | 819 | 0 | 49 | 94% | 100% |
The 2021 cohort is the one to look at. 819 requests withdrawn. None built.
And it is 94% resolved — this is not a cohort still waiting its turn.
The obvious objection is that recent cohorts have had less time, so of course
fewer are built. That objection holds for 2022 and 2023 and I have excluded them.
It does not hold for 2021. **Median time from request to operation in PJM is 3.2
years** (Note 001, n=1,180). A project filing in 2021 reaches the median
completion date in 2024, and this dataset runs to the end of 2025 — a full year
beyond it. There was time.
What happened instead is that 819 of them left, and 49 remain.
Pre-2014 cohorts, fully resolved, cluster around 60–80%. From 2018 the floor
moves to 85%, and 2020–2021 go to 96% and 100%.
Something changed around 2018–2019 and it did not change back.
**If you are pricing a 2019-or-later PJM request off the historical 68%, you are
using a number from a different market.** The relevant base rate for anything
filed since 2019 is 85% at best, and for the 2020–21 clusters it is
effectively total.
If you own a 2021-vintage request that is still active, you are one of 49
survivors from 868. That is not necessarily bad news — surviving a cohort that
brutal is information about your project — but it is a very different position
from the one the composite figure describes.
If you are underwriting new entries, the historical average is the wrong
anchor and the direction of travel matters more than the level.
Why. The data shows the deterioration; it does not explain it. Cluster-study
reform, network upgrade costs, speculative filing, and interconnection cost
allocation are all plausible and I have tested none of them. I would rather
publish the shape than guess at the cause.
Whether it is PJM-specific. The same dataset covers every major ISO. That is
Note 003.
*Method: LBNL's complete request file, filtered to entity = PJM, grouped by
q_year. Failure rate counts terminal outcomes only — active and suspended
requests are excluded, since a live request has neither succeeded nor failed.
Cohorts under 30 requests, and 2022 onward, are omitted as too unresolved to
interpret. Every figure is a count from published data. The file is free at
emp.lbl.gov/queues if you want to check my arithmetic.*
*This is published research. Every figure is a count from a free dataset at emp.lbl.gov/queues — check it yourself rather than taking my word for it.
Later — 2026-08-21. Joe Rand of LBNL, who publishes this dataset, suggested weighting outcomes by capacity rather than counting requests. Doing so makes every figure on this page worse, not better — PJM's 2018–21 cohort moves from 93.4% to 97.0%, and 21 of 25 operators deteriorate. ERCOT survives as the outlier either way. See Note 009. This page stands as published; it counts requests and says so.