PJM Queue — Note 009

I was counting the wrong thing. Counted properly, it is worse.

I was counting the wrong thing. Counted properly, it is worse.

Kenneth · 2026-08-21

*Source: Lawrence Berkeley National Laboratory, Queued Up 2026 Edition (CC BY 4.0),

complete interconnection request dataset through end-2025. Analysis my own.*


Where this came from

I sent Notes 001–008 to Joe Rand at Lawrence Berkeley National Laboratory, who

publishes the dataset they are built on, and asked him to correct me rather than

encourage me.

He was straight about what he had not done — *"I do not currently have capacity to

try to reproduce the analysis"* — so **nothing here is verified by LBNL and I am not

claiming it is.** But he made one suggestion:

*"One thing I would suggest is looking at outcomes weighted by capacity (MW),

rather than simply counts of interconnection requests, since the capacity metric

is more relevant for electric system operations and decision making."*

Every previous note counted requests. A 2 MW rooftop array and a 900 MW plant

each counted once. He is right that the grid does not care how many requests failed

— it cares how much capacity failed.

I expected this to soften my headline. The obvious objection to "93% of requests

fail" is that queues fill with small speculative applications that fail cheaply

while serious projects proceed. If that were true, weighting by megawatts would

pull the number down.

It does the opposite.


PJM, both ways

CohortBy request countBy megawattRequestsGW
pre-201470.4%82.5%2,449369.2
2014–1773.3%75.3%1,150108.6
2018–2193.4%97.0%2,418204.1
2022–25100.0%100.0%35335.9

Every cohort is worse measured by capacity. The headline from Note 002 moves

from 93.4% to 97.0%.


And it is not a PJM effect

2018–21 cohort, operators with at least 30 terminal outcomes:

OperatorBy countBy MWShiftnGW
IP48.8%93.4%+44.6848.8
SRP61.7%85.3%+23.64716.9
NWMT75.9%94.5%+18.6585.3
BPA88.3%99.4%+11.17730.4
ISO-NE82.6%93.1%+10.542534.9
DUKE89.6%96.6%+7.118229.5
NYISO94.6%98.8%+4.351575.9
PJM93.4%97.0%+3.62,418204.1
ERCOT60.7%63.2%+2.654493.9
CAISO96.3%98.4%+2.1510194.7
SPP96.2%97.4%+1.229048.2
FPL19.4%20.0%+0.6727.9
MISO95.2%95.0%−0.3816120.4
PNM95.8%79.6%−16.31208.6

**Twenty-one of twenty-five operators look worse weighted by capacity than by

count.** Only PNM improves materially. MISO is flat.


What that means, and what it does not

**The plain reading: larger projects are withdrawn more often than smaller ones,

nearly everywhere.**

That is the opposite of the comfortable assumption, and it was invisible while I was

counting requests. A queue is not mostly small speculative noise failing harmlessly

around serious projects that proceed. **On this data the serious large projects fail

at a higher rate.**

Why, I do not know, and the dataset does not say. Larger projects trigger heavier

network upgrade obligations, longer study timelines and greater cost-allocation

exposure. All plausible. None of them tested here. As with every previous note:

the data shows what happened, not why.

One caution about the biggest numbers. IP, SRP and NWMT show the largest shifts

on small samples, where a handful of big withdrawals dominate. Those shifts are real

but fragile. PJM's +3.6 across 2,418 requests and 204 GW is the robust one.


Two earlier findings, tested and standing

ERCOT is still the outlier under either metric — 60.7% by count, 63.2% by

capacity, on 93.9 GW of terminal outcomes. This was the claim I was least sure of

when I published Note 003, and it is the one that has held up best. **A

small-sample explanation is now closed off.**

FPL holds at 19.4% and 20.0% on 72 requests and 7.9 GW. Small, but no longer

dismissible as an artefact of counting.


Method

Terminal outcomes only — withdrawn and operational. Active requests are excluded

throughout, because a project still in the queue has neither failed nor succeeded.

Capacity is mw_1 + mw_2 + mw_3 from sheet 03. Cohorts by q_year. Operators

shown where the cohort holds at least 30 terminal outcomes.

Notes 001–008 stand as published. They count requests, they say so, and the

figures in them are correct for what they measure. **This note does not withdraw

them; it reports that they were measuring the less useful of two things.**


*The suggestion that produced this note was Joe Rand's, offered unprompted in reply

to a request to be corrected. He has not verified any of it, and any error here is

mine.*