What does the wait for grid power mean? A hands-on lab built on all 37,527 requests in Lawrence Berkeley National Laboratory's Queued Up data. It covers the odds a power project ever connects, counted correctly when most recent requests haven't finished; how long the wait really is, as a distribution; what a year of delay costs a 250 MW AI campus; and which ways around the line buy time.
Play it: https://bakulbadwal.github.io/interconnection-queue-lab/
Power, not chips, now sets the pace of AI buildout. Every new US power plant must ask its grid operator for a connection, then wait in a public interconnection queue while engineers study what the grid needs. Most requests are withdrawn, and the ones that get built are taking longer. Queue trackers show which projects are waiting. This lab answers what the wait means, in numbers you can defend.
The news hook: on Oct 6, 2026, Google signed a 20-year deal for 890 MW of uprates at 11 existing Constellation reactors in PJM (POWER). That buys new megawatts behind a grid connection that already exists. Step 4 shows why that's worth so much.
| Step | You play with | What clicks |
|---|---|---|
| 0 · The line | Every request since 1995 (the chart starts at 2000), by region and outcome, by MW or by count | Of capacity that asked to connect 2000–2020, 13% was running by end-2025 and 75% was withdrawn (LBNL's headline, reproduced). These are generator queues; a data center's own load request is a separate line |
| 1 · Odds of reaching the finish ★ | Region × technology × size × entry years; three ways to count | The naive share understates (unfinished ≠ failed). Kaplan–Meier that ignores withdrawals overstates wildly (withdrawn can't finish). For PJM requests that entered 2010–2020: naive 16%, KM 70%, competing risks (Aalen–Johansen) 21% within 10 years. Readouts stop where a slice's follow-up ends: odds beyond it are unknown, not zero |
| 2 · How long the wait is | Request→COD distribution for built projects; "already waited N years" landmark odds | Among 2010-and-later entrants in CAISO, MISO, PJM and SPP, the median wait rose from 3.2 years (built 2015–2019) to 5.7 (2021–2025), and the P90 from 6.1 to 8.6. Survivors' odds rise (3-year COD odds 3% at entry → 26% at six years), but most are still waiting |
| 3 · What a year costs | A 250 MW campus, Compute Deal Desk's price of a megawatt-year, discounting, an "idle GPUs" toggle, the measured waits of big PJM gas plants | One year late ≈ $613M at teaching prices, $1.49B if the GPUs are already bought. The null case (power on the ready date) is $0 |
| 4 · The options desk | Queue a new build (measured), PJM's expedited track, an uprate (Google–Constellation), a restart (Crane), behind-the-meter gas, flexible load | The scarce asset is the interconnection right, not the generator. Even Crane's restart ran into PJM upgrades until 2031 |
| ★ The power memo | Split 250 MW across paths; exact odds of ≥200 MW by end-2029; a six-sentence memo drafted from your plan | All in the queue: 2%. A reference mix (uprate + behind-the-meter gas + flexible load): 26%. All behind-the-meter gas: 83% under the default odds, and the step asks what would stop you |
| ✓ Field test | Eight questions answered by operating the widgets | Proof it stuck |
Each step has a predict-then-reveal question, a "say it out loud" line, a model card (the estimator, what's measured vs modeled, the Python twin), and a running, printable list of takeaways.
- Measured (tag: measured): every share, odds curve, wait distribution and landmark probability, computed in your browser at run time from
data/queue.json, a cleaned extract of LBNL's file. - Sourced (tag: sourced): deal terms, dates and equipment prices, cited where they appear and listed in DATA.md.
- Teaching values (tag: teaching value): the value of a megawatt-year ($3.48M, Compute Deal Desk's shadow price at teaching prices), the discount rate, and the odds and most-likely times for paths with no public track record. Quote the measured numbers and the mechanics, not the dollars.
Limits. These are generator queues, not load queues. Timing statistics come from the four regions whose files date nearly every outcome (CAISO, MISO, PJM, SPP; 14,951 requests entered 2010–2025). ERCOT can be counted but not timed. Details in DATA.md.
| Tool | Offers | This lab adds |
|---|---|---|
| interconnection.fyi (GridTracker) | Daily tracker of 44k requests: map, filters, status | Odds, durations, delay cost, options |
GridStatus app + gridstatus library |
Queue browser; standardized ISO queue DataFrames | Statistics on outcomes and timing |
| LBNL Queued Up + its durations dashboard | The authoritative cohort completion shares (2000–2020) and median durations, one dimension at a time | Censoring-aware odds that use recent requests too; any region × technology × size × era slice; P90s; conditional "already waited" odds; dollars; options |
| Paces | Commercial siting with power-flow studies | A free teaching lab on the queue statistics themselves |
make check is the acceptance gate (982 checks plus 288 cross-checks at the last run):
- every estimator in
js/stats.jsandjs/econ.jsis re-implemented in pure-stdlib Python (models/), and 791 values must agree to 1e-9; - the survival estimators are cross-checked against statsmodels (
CumIncidenceRight,SurvfuncRight), an implementation this repo didn't write: 288 values to 1e-9; - null cases and identities: no delay ⇒ $0; S + CIFCOD + CIFwithdrawn = 1; KM = Aalen–Johansen with no competing events; Aalen–Johansen = the naive share with no censoring; a certain, on-time portfolio ⇒ P = 1;
- LBNL's published headline (13% / 75% / 19%) is reproduced from the snapshot;
- every number the copy quotes is computed in
js/numbers.jsand written totests/numbers.json. The check fails if a step file hard-codes a percentage or dollar figure that isn't traced there or to the cited-sources list.
make check # Node + Python (statsmodels via uv)
make serve # http://localhost:8831 (ES modules need http://, not file://)
make smoke # load every route in headless Chrome, with the server running
make data # re-download LBNL's file (SHA-256 checked) and rebuild data/queue.json| Path | Role |
|---|---|
data/fetch_clean.py |
Download, hash check, clean, write queue.json and SNAPSHOT.md |
data/options.json |
Campus inputs and the options desk's anchors, tagged measured / sourced / teaching |
js/stats.js |
Outcome shares, Kaplan–Meier, Aalen–Johansen, landmark odds, quantiles |
js/econ.js |
Delay cost, triangular time-to-power ranges, the exact portfolio distribution |
js/numbers.js |
Every number the prose quotes, computed once |
js/steps/*.js |
One module per step |
models/ |
The Python twins and run_checks.py |
tests/ |
check.mjs, check_statsmodels.py, battery.json, numbers.json |
- Hyperscale Ledger: the campus (audited GPU specs, rack power, the 250 MW scenario).
- Compute Deal Desk: what a megawatt is worth (the shadow price this lab multiplies by time).
- Interconnection Queue Lab: how long until the megawatt arrives, how likely it is, and what the wait costs.
Built by Bakul Badwal (UVA Darden MBA '27) with Claude Code.
Code: MIT (see LICENSE). Data: data/queue.json is adapted from the Queued Up 2026 data file by Lawrence Berkeley National Laboratory and GridTracker, licensed CC BY 4.0. The changes are listed in DATA.md.
