How we make it

Built by AI.
Powered by humans.

All figures on this page are as of 15 August 2026, the date of our last ingest from source. Every figure, by state and county · how we make it.

Ten AI departments read the American grid around the clock. People approve everything that ships. Every number traces to a public record — and our misses stay published next to our hits.

STAGE 01

Fetch

56 sources pulled on schedule — ISO queues, utility filings, dockets, disclosures.

◆ 2×/day · archived forever
STAGE 02

Parse

Split headers, merged cells, PDFs, JSON — every format normalized to one schema.

◆ 39,173 rows
STAGE 03

Reconcile

Row counts checked against the source. Unexplained drops fail the run.

◆ 0 silent drops
STAGE 04

Join

Projects linked to counties, people, rules, and money through our entity graph.

◆ 1,575 counties
STAGE 05

Human review

A person approves every report, score, and guide before publication.

◆ nothing auto-publishes
The department floor

Ten agents. One newsroom.

Each department runs on its own schedule, writes to a shared record, and escalates to a human when something matters.

Data Ingestion2×/day · freshness & QA
Load Registry2×/day · DC requests
Policy Desk2×/day · dockets & rules
Influencedaily · lobbying
Research2×/day · reports
Marketing2×/day · content
Sales Deskweekdays
Productdaily · build queue
Trust & Methoddaily · spot-checks
Chief of Staff2×/day · synthesis
Human gate

No agent sends an email, publishes a page, or moves money. Drafts queue for a person — always.

Immutable history

Scores are timestamped and never edited. Source files are archived the day we fetch them.

Public corrections

When we're wrong, the correction is dated and stays visible. That's the whole business.

Reconciliation — from the 15 August 2026 build

Our receipts, on the record

reconciliation · build 2026-08-15

These are the counts recorded by that ingest run, not a live feed. It said "live" and was not; the figures were always real, the label was not.

SourceRawParsedDroppedNote
CAISO3 sheets2,2820header row 4
PJM9,263 rows9,2630full history to 1997
MISO3,810 rows3,8100full lifecycle
ISO-NE1,751 rows1,7510A/C/W status letters
ERCOT1,971 rows1,9674blank INR rows
BPA1,465 rows1,4650incl. 455 load requests
LBNL (non-ISO)38,201 rows15,54822,653ISO rows superseded by live feeds
Corrections — public

Everything we got wrong, dated

When we're wrong, the correction is dated and stays visible. We keep two lists, because they are not the same thing: errors that reached you, and errors our own checks caught before publication. Publishing only the second kind would be its own form of dishonesty.

published — corrected on the live site
DateCorrection
2026-08-17Source count
We said we read 57 primary sources. Seven ISO/RTO queues plus 49 non-ISO utilities is 56. The wrong figure appeared on the homepage, on this page, and in llms.txt — the file written so answer engines can quote us accurately.
2026-08-17Load Ledger labelling
We labelled 488 map marks and 545 queue rows as data-center power requests. In fact 494 of the 545 rows carry no end-use classification at all; 40 are tagged as data centers and 36 of those are geocoded. We were asserting an end use for roughly 450 requests we had never classified. They are now described as large-load power requests, with the confirmed data-centre count shown separately.
2026-08-17Coverage start date
We said our records run since 1997, including in the machine-readable Dataset markup. The earliest interconnection request in our data is 1989-09-15. Corrected to 1989.
2026-08-17AI campus count
llms.txt reported 72 AI data-centre campuses while the map showed 44. Both were true of different things: 72 campuses are identified, 44 of them are geocoded and appear on the map. We now say which is which.
2026-08-17API example hostname
The copy-paste example on our developer page pointed at a hostname that carried another company's product name, and at a raw worker origin rather than api.wattstreet.net. Both corrected.
caught before publication — never shipped
DateNear miss
2026-08-17ERCOT audit scope
A draft collapsed three different figures into one. 474 GW is the full ERCOT interconnection queue; ~200 GW is the audited Batch Zero cohort; 49.8 GW is BloombergNEF's estimate of delayed data-centre load. Three numbers, three denominators.
2026-08-17ERCOT filing date
ERCOT's good-cause exception was filed 2026-08-11, not 08-10. August 10 was a Sunday.
2026-08-17Oklahoma HB 2992
Enacted 2026-05-11, effective 2026-07-01. The 2026-08-06 date reported in several outlets is a ceremonial signing three months after enactment.
2026-08-17New Mexico
Recorded as a statewide moratorium bill. It is an announced intent to introduce legislation in the 2027 session. No bill number exists.
2026-08-17CAISO abeyance motion
Recorded as stated intent only. Confirmed filed 2026-08-03 on CAISO's own FERC filings register.
2026-08-17Fort Worth, Texas
Recorded as an adopted moratorium. The council initiated the statutory process; earliest possible effect is 2027-02-16. Counting it as adopted is wrong by one and by six months.
2026-08-16Fermi — Matador site
Filed in our draft under ERCOT. The site is in SPP territory.

Nothing is removed from this page. If a correction is later itself corrected, both entries stay.

Coverage roadmap

What's connected — and what's connecting next

We publish the roadmap because buyers deserve to know the edges of the map.

The team

Humans on the line

Every report is read by a person before it reaches you. Every support email is answered by one. When we get something wrong, a human writes the correction and signs it.

function nav(e,p){ if(window.go){e.preventDefault();go(p);history.pushState({},'',{"platform": "/", "business": "/business", "reports": "/reports", "packages": "/packages", "nearme": "/near-me", "agents": "/for-agents", "method": "/how-we-make-it"}[p]||'/');return false;} return true; }