# Data-center siting: descriptive observed-location comparison

Generated 2026-09-14T19:24:46.623Z; current method `2026-09-load-service-v6`; previous method `2026-09-battery-poi-v5`; environmental snapshot `2827bcdb6114`.

This compares locations from the [Cleanview California public inventory](https://cleanview.co/data-centers/california) with production Site Selection Atlas scores. It is a descriptive diagnostic, not an independently trained predictor, causal test, historical backtest or validation of utility capacity. No weights are fitted to this inventory. Inventory membership contributes no score.

## Scenario

Every location receives the same hypothetical **100 MW peak facility load**, 90% steady hourly utilization, PUE 1.25, WUE 0.5 liters/IT-kWh, target 2031, minimum 115 kV, 25 km study radius and 0.5 acres per facility MW. The 100 MW is facility power, not IT power. PUE translates facility demand to an IT/cooling-water envelope; it does not add a second PUE multiplier to the electricity bill. Source-reported project MW may have an unknown basis and is not substituted into this controlled comparison. These are not valuations of each actual data center.

## Source quality and grouping

The source snapshot supplies 86 records. 76 operating/construction/planned records with usable published coordinates pass the California land check; 10 records are excluded with reasons in JSON. Unknown status or missing/invalid coordinates are excluded, never inferred. Public map points remain unverified parcel locations. Multi-building campuses sharing a documented campus group contribute only one anchor: most mature status then stable ID. All members and mixed statuses are retained. This leaves 67 campus anchors, using 24 environmental grid samples and 41 price nodes. These counts are not independent geographic observations.

## Prespecified comparisons

Alternatives are mapped California substations at 115 kV or above, within 200 km of an eligible observed campus and at least 1 km from every mapped source project (including canceled projects). Nearby substations are deduplicated in 0.005-degree cells. They are comparison coordinates, not confirmed available parcels, failed projects or a complete map of places without data centers.

Matches require the same voltage band and the same broad NLCD class (developed-open/low 21-22; developed-medium/high 23-24; non-developed separately), no more than 2 km difference in substation distance, 3 km difference in line distance, five percentage points in previous-method non-queue coverage and eight points in previous-method grid score. Unknown urban class is not matched. Coverage and grid matching use the frozen previous evaluator to keep matched pairs fixed across changes. Matching never uses overall, market or operations scores.

The local comparison permits at most 50 km geographic separation; the broader regional comparison permits 200 km. Both retain the urban-context constraint to avoid comparing urban campuses against arbitrary rural land. Within each scope, alternatives are assigned without replacement, most constrained campus first. The same alternative may be used once in each separate scope. Unmatched campuses remain explicit; criteria are not relaxed to manufacture separation. Observed coordinates are not snapped to substations, so unmatched campuses farther from mapped high-voltage infrastructure are a material selection limitation.

## Results

| Scope | Cohort | Campuses | Scored pairs | Observed median | Alternative median | Median paired difference |
|---|---|---:|---:|---:|---:|---:|
| local | operating | 43 | 26 | 75.95 | 76.5 | -0.6 |
| local | construction | 0 | 0 | n/a | n/a | n/a |
| local | planned | 24 | 17 | 75.1 | 76.3 | -1.6 |
| regional | operating | 43 | 27 | 75.9 | 76.8 | -1.8 |
| regional | construction | 0 | 0 | n/a | n/a | n/a |
| regional | planned | 24 | 17 | 75.1 | 76.3 | -1.6 |

Across the **43 local pairs scored in both versions**, the previous-method median observed-minus-alternative difference is **-0.8**, compared with **-1.1** under the corrected semantics. A larger difference is not proof of better predictions. The median observed score change is **3** points and median coverage change **-21.3** percentage points; removing unsupported favorable or unfavorable proxies changes normalization.

## What changed and what remains unresolved

The former generator-queue competition proxy is not a load-service readiness measure. Load readiness now remains unknown until relevant service evidence is available. Nearby operating generation is not firm or contracted electricity for a load. Middle-mile route proximity does not establish lit carrier access, capacity, diversity or latency; groundwater pressure must be tied to the assumed cooling water use. Component-level before/after values and evidence statements are retained for every location.

Remaining important inputs include utility will-serve status and timing, actual retail tariffs and demand charges, contracted clean supply, carrier/latency commitments, parcel cost and entitlement, cooling design and water supply, backup-generation permits and project financing. Nearest-node wholesale cost is not a utility retail bill. A low score at an operating campus may identify missing model inputs or inaccurate source coordinates rather than a poor actual investment.

## Limits

- This public map is a selected inventory, not a census of all California data centers. Page totals can differ from geocoded records; consult the source-quality counts.
- The sample is geographically concentrated. Closely spaced campuses share environmental samples and price nodes; do not treat every building or campus as an independent trial. County counts and shared-sample counts are in the JSON.
- Operating sites show past feasibility. Planned sites show reported intent, not successful completion. Cancellation is not a reliable negative label and is excluded.
- 2025 prices and later environmental/inventory snapshots cannot establish that earlier siting decisions were predicted using information available beforehand. Target 2031 does not forecast utility capacity, tariff or energization timing.
- Regional five-kilometer evidence may miss the exact parcel's zoning, flood mitigation or developed land cover. A known urban campus may inherit a nearby non-urban sample; this is disclosed rather than overwritten using its outcome.
- No weights were tuned and no independent hold-out claim is made. A future predictive test should reserve whole metropolitan areas or utilities, freeze dated inputs, and observe subsequent energization outcomes.

## Reproduce

Extract [baseline-v5.zip](../data/validation/baseline-v5.zip) under `.codex-tmp/siting/repro`, then run:

`node siting/scripts/validate-datacenter-siting.mjs --baseline=.codex-tmp/siting/repro/baseline-v5/model.js`

The runner refuses non-file URLs. Raw inventory, every score/component, exact matched pairs, exclusions, module hashes and hashes of all consumed datasets are in [datacenter-validation.json](../data/validation/datacenter-validation.json). The benchmark makes zero external requests; source snapshots are refreshed separately.

Current module hashes describe source files at evaluation completion. Unscored inventory context and interface additions were made during the first run; numerical scoring semantics remained unchanged. Linked-campus values are checked against the final browser runtime separately. The previous evaluator was frozen before any changes and is preserved in the archive.