# JevonsIntel public-data intake — September 9, 2026

The public-data pull is complete as a local intake batch. The full backlog is not exhausted. No JevonsIntel application, database schema, credentials, or existing data snapshot was supplied in this workspace, so this package has not been imported into a live JevonsIntel system and cannot identify changes against its existing data.

Collected **58 dataset/API snapshots** (229.6 MB), prepared **286,596 source records**, and curated **5 relationship records covering four power arrangements**.

## What is available

| Prepared table | Records |
|---|---:|
| eia860_2024_final_generators | 34,855 |
| eia860_2024_final_plants | 16,132 |
| eia860_2024_final_utilities | 6,643 |
| eia860_2024_final_ownership | 5,496 |
| eia861_2024_final_service_territory | 11,866 |
| eia860_2025_early_generators | 36,106 |
| eia860_2025_early_plants | 16,933 |
| eia860_2025_early_utilities | 6,788 |
| eia860_2025_early_ownership | 5,668 |
| eia861_2025_early_service_territory | 9,503 |
| eia923_2026_monthly_generation_fuel | 7,925 |
| eia923_2025_early_generation_fuel | 18,080 |
| egrid_2023_plants | 12,612 |
| lbnl_queues_2025 | 38,201 |
| sec_financial_observations | 56,351 |
| sec_filing_index | 3,437 |

Raw EIA archives also contain supporting schedules and source instructions beyond the selected tables above. SEC coverage includes 25 public companies: AEP, AMZN, CAT, CEG, CMI, D, DLR, DUK, EQIX, ETN, ETR, GEV, GOOGL, IBM, META, MSFT, NEE, NRG, ORCL, PWR, SO, TSLA, VRT, VST, XEL.

## Company-to-power evidence

| Company | Relationship | Capacity meaning | Status caveat |
|---|---|---|---|
| Microsoft | Constellation / Crane | 835 MW generation restart | Restart work; current operator page targets 2027. No campus load established. |
| Amazon / AWS | Talen / Susquehanna | 1,920 MW full PPA quantity | Ramping contract, full quantity expected by 2032 in the June 2025 announcement; actual delivered load unverified. |
| Meta | Entergy Louisiana / Richland Parish | Up to 5,000 MW potential campus expansion | Construction and conditional expansion; not energized load. |
| Google | TVA / Kairos Hermes 2 | 50 MW generation project shared across two campus relationships | Announced 2030 target; Montgomery County, TN and Jackson County, AL. Count the project once. |

Primary URLs and claim-level caveats are in `data/prepared/relationships.jsonl`; retrieval evidence is in `data/relationship_sources.json`. Confirmed confidence describes the named relationship, not delivery certainty or completion.

## Financial observations

SEC extracts retain capex/PP&E, lease, acquisition-payment, and commitment-related concepts, units, reporting periods, accession numbers and filing dates. These are historical observations with overlapping periods and restatements, not additive records or unique economic events. Company Facts does not replace reading commitments and risks in filing text.

Selected latest annual cash PP&E observations are in `data/prepared/compute_company_annual_cash_capex.json`. These are company-wide cash purchases, not an AI demand leaderboard. Amazon is explicitly left unavailable for this comparable concept because its recent disclosures use a broader productive-assets concept. Fiscal periods differ.

## Interpretation and integration rules

- Keep EIA 2024 final and 2025 early-release tables separate. EIA explicitly says the early releases are incomplete/not fully edited and should not be aggregated to state, regional or national totals. Counts here describe extracted records only.
- The EIA-861 2025 early service-territory introductory text refers to calendar 2026, while its filename and row Data Year identify 2025. Preserve this source inconsistency; this package labels the rows 2025.
- EIA-860 Operable includes statuses beyond operating, including standby and out-of-service. Preserve sheet and Status; do not classify the entire sheet as operating.
- EIA-923 monthly data cover January–June 2026. Later-month dots represent unavailable observations, not zero generation. Preserve fuel and prime-mover granularity; annual and monthly vintages overlap.
- EIA-861 county coverage supports utility screening, not parcel-level campus service confirmation. No utility has been assigned from state or county alone.
- eGRID is the 2023 revision-2 vintage. It is not a measurement of 2026 carbon intensity.
- Berkeley Lab queues describe generation/storage interconnection requests through year-end 2025, not data-center load queues. Preserve q_status, entity, q_id, project_type, mw_1/mw_2/mw_3 and IA phase. Queue rows can represent modifications or related requests; neither q_id nor row count is guaranteed to equal a unique project. Do not sum components or classify an active queue as built capacity.
- SQLite is a portable staging database. Each source table stores record_id, source_id, source_row and a JSON payload. JSONL preserves original source column names and metadata. Source row IDs locate evidence within a specific snapshot; they are not stable entity IDs across refreshes.
- Join EIA and eGRID plant identifiers only after normalizing types and checking year, uniqueness and coverage. Resolve corporate parents separately from utility legal entities. Campus-to-utility links require direct evidence.
- Relationship capacity_group_id prevents counting the same Hermes 2 project twice. Campus demand, PPA volume and generation capacity must remain separate measures.

## Remaining backlog

Not collected in this batch: EIA-861M and EIA-930 hourly feeds; separate CAMD time-series; direct PJM/ERCOT/MISO/SPP/CAISO/NYISO/ISO-NE forecast/price/reliability feeds; FERC EQR/docket documents; comprehensive state commission filings, permits and tariffs; full filing-text extraction; private-company coverage; and comprehensive campus or equipment/fuel-beneficiary relationships. LBNL provides an aggregate queue backbone but does not replace direct regional feeds.

Compute Demand Leaderboard, comprehensive Campus-to-Utility Table, AI Power Exposure Map, and Energy Beneficiary/Readiness Score remain product outputs to build from expanded and reconciled evidence. No scoring methodology was supplied, so no readiness scores were invented.

## Files and refresh

- `data/jevonsintel_staging.sqlite`: local staging database.
- `data/prepared/`: JSONL tables, inventory and selected annual cash PP&E observations.
- `data/raw/`: original downloaded ZIP/XLSX/SEC JSON and available relationship-page snapshots.
- `data/source_manifest.json`: URLs, periods, retrieval times, byte sizes and SHA-256 hashes.
- `reports/validation.json`: executed validation results.
- Scripts: run `pull_public_data.py`, `prepare_data.py`, `prepare_relationships.py`, then `summarize_and_verify.py` using Python with openpyxl. Downloads reuse existing snapshot files; for a genuinely new collection use a new batch directory and recheck latest publisher release links. No recurring task was scheduled.

## Attribution

EIA and EPA datasets retain their agency attribution. Queued Up Data File © 2026, The Regents of the University of California through Lawrence Berkeley National Laboratory and GridTracker, licensed [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). This package converts selected source tabs to JSONL without changing their data values (dates serialized as strings); source notes are excluded from data rows. Original workbook and codebook are retained.

## Verification

- All 58 raw dataset/API downloads match recorded SHA-256 hashes.
- Every prepared JSONL record parses and references a registered source; counts reconcile to inventory; source footnotes excluded.
- SQLite integrity check passed; all staging-table counts match JSONL files.
- Five relationship IDs are unique and retain evidence URLs, verification time, and unknown energized load as null.