THE LOADSTAR NUMBER · FREE TO CITE
Announced is not requested. Requested is not signed. Signed is not built.
Of the 1160.6 GW requested across the 7 active U.S. ISO generation interconnection queues, at least 298.8 GW — about 25.7% — carries an executed interconnection agreement, almost all of it from the 6 queues where our reading yields a meaningful signed share (CAISO, ISONE, SPP, ERCOT, NYISO and PJM). A floor, not a rate: what our reading cannot see counts as unverified, never as speculative.
Supply side: power plants and storage waiting to connect to the grid. Not a count of data-center demand.
Citation copied — paste it anywhere.
the series · live
reading as of 2026-10-01 · recomputable
PJM · ERCOT · CAISO · MISO · SPP · ISO-NE · NYISO — read from each operator’s own published queue, recomputable from our public JSON.
How to read this figure — the full picture ▾
Four questions, four straight answers
Is the AI buildout real?
Most of it is unproven — which is not the same as phantom.
Only a floor of the U.S. interconnection queue can be machine-verified as signed. The live share, and what it does and doesn’t mean:
THE NUMBER →
Does anyone publish declared AI load?
One U.S. grid operator names it, request by request.
A dedicated “AI Datacenter” class in a public file — the record, and the four things it cannot tell you:
DECLARED AI LOAD →
Can I check your numbers myself?
Yes — every figure recomputes from official filings.
Step-by-step derivation, hash-sealed ledger, and the rules that govern how any of it can change:
VERIFY IT →
Will you admit it when you’re wrong?
Yes — in public, dated, with the size of the change.
When our reading of a queue was wrong, the number went down and the correction says by how much. Every change, dated:
CORRECTIONS →
What brings you here?
Same instrument, four doors. Go straight to the read your seat needs.
Regulator / policy staff
What each grid operator actually publishes on large-load requests — read operator by operator, with the date each document was read.
Open the disclosure survey →Journalist on deadline
A citable number with its primary source and a one-click citation string. Written answers within 24 hours.
Get the number + receipt →Investor / analyst
Formation scores region by region, each linking to a recomputable receipt — announced demand vs. what public signals corroborate.
See the live index →Utility / grid planner
Energy first: where low-carbon supply is forming in the interconnection queues, and how much of the queue is machine-verifiable as signed.
Open the energy map →Who pays
The AI data-center boom is already on your power bill.
Utilities and capacity markets price forecast data-center demand into bills now, on the largest U.S. grid alone, that pricing added billions in a single year. That is the real, rising cost of the AI boom — and the part no one can yet check is how much of the demand behind it is committed, and how much is speculative.
PJM · 13 states + DC
$9.3 billion in one year
the data-center-attributable share of the rise in PJM consumers' capacity cost for one delivery year (63% of the total increase), per PJM's own market monitor.
Source: Monitoring Analytics ↗Virginia
+$11.24 a month
the 2026 residential increase Virginia's regulator approved in Dominion's rate review — and state analysts project a further $14–$37 a month by 2040 driven by data-center growth. From 2027, a new data-center rate class makes large loads cover most of their own costs.
Source: SCC · JLARC ↗The blind spot
no public field says which is real
Generation queues generally publish agreement status — in our reading, only some of the seven publish a machine-readable field — while most large-load requests are not disclosed request by request, so from the public record alone, no single field separates a committed data center from a speculative one.
The dollar figures are on the public record, reported by regulators and grid monitors, not measured by us. What we add is one missing read: how much of the capacity requested in the grid queues carries a signed agreement, queue by queue, with the source file attached.
See the number, queue by queue →Why we exist
AI is being built faster than the power to run it.
Every week brings another gigawatt-scale data-center announcement. But a press release isn't a power plant, and the grid can't tell the difference. Utilities commit billions in upgrades, regulators approve rate hikes, and the public foots the bill — all against queues of power requests whose public records rarely say which ones are committed.
We built LOADSTAR to measure that gap from the grid operators' own queue files: how much of the capacity requested carries a signed agreement. Every figure is dated, and the daily files behind it are archived, so a published number can't quietly move. We take no side and sell no project.
Others sell the answer privately to one developer. We publish the number for everyone: free to cite, sourced, and neutral.
How we stay neutral →We don't ask you to trust us.
We ask you to check.
We don't say whether it's real — we say how much we were able to verify. Every published figure comes with its source and its read date, and the daily files behind it are archived with a SHA-256 fingerprint, so anyone can recompute it.
Absence of signal isn't proof of phantom. It's a coverage gap — and we say which one.
The Reality Report
The number, what moved, and every correction — by email.
A short dated note when the number moves or a correction goes public. Free, and cross-posted in full on Substack.
Cite it, question it, or build on it.
The number is free to cite. Questions about the method get a written answer. Teams that need the daily files behind it can ask about a data license.