Every site, found and scored in the open.
Refinery Atlas turns public infrastructure data into a ranked pipeline of stranded-power sites. Here is exactly how a footprint becomes a scored, sourced candidate — and how you can audit or re-weight every number behind it.
Principles
The fastest MW is already on the grid
Greenfield power waits years for interconnection. A slabbed parking lot or a half-empty mill already holds energized 2–20 MW service the utility hasn't reclaimed. We hunt that.
A wide net, then corroboration
Every building footprint becomes a speculative candidate. Layoff notices, closures, and vacancy signals promote it to probable. Researched intelligence lifts it to verified. Confidence is earned, not assumed.
Every input is sourced
Each number carries its raw value, its normalization curve, and its provenance. Contacts surface only with a real source URL attached. No black boxes, no fabricated facts.
Two surfaces, one model
A searchable index for the analyst and a live map for the scout — both reading the same scored, sourced dataset. Adjust the weights and every site re-ranks.
The discovery funnel
- 01Ingest
OSM building footprints + HIFLD/SeerAI grid + ERCOT queue across 40 countries, Texas-first.
- 02Match
Each footprint joins its nearest substation (PostGIS KNN); area × archetype W/m² estimates load; generation and oversized polygons are excluded.
- 03Corroborate
WARN layoffs, closures, and vacancy signals are name- and proximity-matched to promote candidates from speculative to probable.
- 04Enrich
A research pass adds sourced owner / operator / utility / status — verified intelligence, never fabricated.
The candidate scorecard
The headline score every candidate carries is a weighted blend of three signals — the same axes you see on each site’s radar. Each maps its raw value through a piecewise-linear curve to 0–100.
Estimated load from footprint × archetype W/m². The thesis weight — energized capacity is the prize.
0 → 100 across 0–20 MW
Straight-line distance to the nearest ≥69 kV substation. Closer is cheaper to interconnect behind the meter.
100 at 0 mi → 10 at 10 mi
Voltage of that nearest substation. Higher class means more headroom to add load.
40 at 69 kV → 100 at 345 kV
Confidence tiers
Energized footprint, estimated load — no corroborating distress signal yet.
Corroborated by a WARN layoff, closure, or vacancy signal within range.
Researched intelligence confirms the opportunity — sourced owner, status, and grid context.
Profiles — the full weighted-sum model
Beyond the headline score, each use-case re-weights eight auditable categories. Every weight is exposed as a slider in the atlas; change them and the whole map re-ranks live.
| Profile | power | grid | land | fiber | water | climate | tax | community |
|---|---|---|---|---|---|---|---|---|
| Hyperscale DC | 30 | 10 | 12 | 15 | 8 | 7 | 10 | 8 |
| Colocation | 22 | 8 | 10 | 25 | 5 | 5 | 10 | 15 |
| BESS Merchant | 18 | 40 | 12 | 0 | 0 | 5 | 15 | 10 |
| BESS Offtake | 20 | 20 | 15 | 0 | 0 | 5 | 20 | 20 |
| Edge DC (2–20 MW) | 32 | 12 | 6 | 22 | 3 | 6 | 9 | 10 |
| Container BESS | 18 | 30 | 22 | 0 | 0 | 4 | 13 | 13 |
| Stranded Power | 38 | 10 | 10 | 5 | 2 | 5 | 12 | 18 |
Power
Available MW, transmission proximity, substation headroom, dual-feed redundancy.
Grid
Locational marginal price, congestion hours, ancillary stack (ECRS, RRS, Reg-Up).
Land
Acres, slope, FEMA flood overlap, NFPA 855 setbacks, zoning compatibility.
Fiber
Distinct providers within 1 mile, distance to nearest IXP, theoretical Dallas latency.
Water
Cooling water availability, distance to >50 cfs waterway, drought severity.
Climate
ASHRAE free-cooling hours, FEMA National Risk Index, severe-weather exposure.
Tax
JETI eligibility, Opportunity Zone, Chapter 312, ITC stack, paired-storage exemption.
Community
Permitting timeline, NIMBY history, construction trades availability and wages.
Why weighted-sum
Of the MCDA families — AHP, TOPSIS, ELECTRE, PROMETHEE — weighted-sum wins here for three reasons: users adjust weights live, it composes cleanly with per-category confidence, and it stays fully auditable in the UI. Composite maps to tiers (S 90+, A 80–89, B 65–79, C 50–64, D <50); confidence below 0.6 flags the weak inputs rather than hiding them.