Tracing the physical footprint of a digital world.

Searching statewide imagery

Existing lists capture sites that have already been documented. To look beyond them, we divided Illinois’s 2023 USDA National Agriculture Imagery Program (NAIP) coverage into 2.91 million tiles, each 224 meters across. The imagery has approximately 1-meter resolution.

SkyCLIP represents each tile as an embedding: 512 numbers describing its visual features. We averaged embeddings from known data centers and ranked other tiles by their similarity. To evaluate the search, we withheld each known site in turn and tested whether it could still be found.

Without adjustment, the search recovered 36% of held-out sites in the top 1,000 results. Average similarity scores were close: 0.83 for random tile pairs and 0.86 for data center pairs. Shared visual features across Illinois made the distinctive signals harder to isolate.

ZCA whitening rescales the embeddings to reduce the influence of shared features, without adding labels. This increased recall, the share of known sites recovered, from 36% to 80% in the top 1,000 results.

A text-only query, such as “an aerial photo of a data center,” achieved 36% recall in the top 1,000 results without example images. This suggests a starting point for searches in regions with limited labeled data.

How image resolution affects discovery

In the top 1,000 results, approximately 1-meter NAIP imagery recovered 80% of known sites, compared with 56% for 10-meter Sentinel-2 imagery. NAIP recovered all hyperscale and large purpose-built sites in the test. Both sources recovered 79% of warehouse-style sites. Combining them added one site beyond those recovered by NAIP alone.

Finding sites missing from the inventories

Searches in Earth Index, Earth Genome’s imagery search tool, identified Illinois data centers absent from all three inventories. Each search used a small set of labeled example tiles, followed by human review of the results against NAIP imagery and public records.

Absent from inventory

Absent from inventory

Absent from inventory

Absent from inventory

The review also identified an EdgeConneX facility missing from the inventories. Searches frequently returned electrical substations, which can provide useful context for nearby data centers but do not establish a facility’s identity on their own.

Reviewed in Earth Index. Imagery: USDA NAIP via the USGS National Map. Some facilities were completed after the imagery was captured.

Search results require verification

Recall measures how many known sites a search recovers. Precision measures how many returned results are data centers; we have not yet measured it. Logistics warehouses can rank highly because they resemble colocation facilities, making human review essential. About one in five known sites was not recovered by the NAIP search at this threshold. Permits and registries remain important for facilities with limited visible features, including urban carrier hotels and small edge sites.

Prior land use varies across campuses

Of the 23 campuses assessed, 18 were classified as previously developed sites, such as factories, warehouses, or existing data centers. The remaining sites included cropland or pasture, woodland and wetland, a brownfield, and former farmland later prepared as an industrial park. These categories provide context for evaluating each development’s land impacts.

Different landscapes, different impacts

The two hyperscale campuses replaced different landscapes. Meta’s DeKalb campus developed former row cropland in phases. Microsoft’s Hoffman Estates campus replaced a mix of pasture, old field, woodland, and wetland. Estimated tree cover at Hoffman Estates fell from 62% in 2019 to 5% by 2021.

The land-cover datasets captured this change differently. NLCD did not capture the rapid tree-cover loss reflected in Dynamic World’s 10-meter Sentinel-2 classification. Comparing multiple datasets and imagery provides more detail than a single “greenfield” designation.

Measuring the wider site footprint

Building area captures only part of a campus footprint. Roads, parking, substations, generator yards, and other built surfaces also occupy land.

Across 21 campuses with mapped roofs, roof area totals 121.7 acres. The satellite-derived built-surface estimate totals 742.8 acres, approximately 6.1 times the mapped roof area.

Built surface is estimated from 10-meter satellite imagery from 2026 and can include adjacent development. Excluding campuses with that overlap reduces the ratio to approximately 5×.

From facility permits to regional questions

Nameplate capacity describes a generator’s rated output. It does not indicate how often the generator runs or how much pollution it emits. The reviewed permits set annual operating limits of roughly 180 to 1,200 hours for testing and outages. Multiple facilities could operate generators during the same outage or grid emergency, making the timing and combined effects of operation important questions for further analysis.

Of 73 final air permits in the collected records, 22 are dated 2024 or later. These recent approvals provide evidence of continued permitting activity, rather than a direct measure of installed or operating capacity.

Generator totals were extracted from 24 unique Illinois EPA records using optical character recognition (OCR). Aligned ORD-01 and ORD-02 share one record and are counted once. Two of 41 generator groups were checked manually. Generators without a recovered kilowatt rating contribute zero to the capacity total, making 1.26 GW a minimum estimate.