How we decide what counts as a fact

Our seven standards

1. Original sources first

   Every figure links to the original report, dataset, filing or study — not to an article about it. If we cannot yet find and link the original, we wait to post the figure rather than rely on a secondhand account. Where a secondhand link is temporarily unavoidable, it is marked “link to original pending.”

2. Every source labeled by type

   Each citation carries a label showing who produced the information, so readers can judge it for themselves. The six types are defined below.

3. Equal treatment for both sides

   On every issue we present the strongest sourced evidence that supporters and critics each point to — with the same space, the same visual styling and the same sourcing standard. Neither side gets the last word by default.

4. Facts kept separate from opinion

   A fact is a measurement, count, official record or study finding that can be checked against its source. Opinion pieces and debate arguments are clearly labeled as such and live in their own sections.

5. Every figure dated

   Each source shows when it was published, and each page shows when it was last reviewed. Data centers change quickly; an old number is labeled as old.

6. Uncertainty stated openly

   Where the evidence is incomplete, estimated or disputed, we say so in a “Still uncertain” note rather than presenting a single number as settled.

7. Corrections made in public

   When we get something wrong, we fix it and record the change, the date and the reason in the corrections log below. We do not silently edit published facts.

The six source types

Labels describe who produced the information, not whether it is right. Every type can be accurate; every type can have interests. Knowing the source helps readers weigh the evidence.

How the Data Center Tracker is built

The tracker combines three layers of information. Every row shows which layer it comes from with a dot rating, and every row links to its source.

1. Fresh public data every night. Each night an automated process downloads the latest versions of two public, openly licensed databases:
   • Academic Epoch AI’s data center dataset — the largest AI computing campuses, with status and power estimates drawn from satellite imagery, permits and company disclosures. Power figures are estimates, typically within a factor of 1.4 of the true value.
   • Community OpenStreetMap — every U.S. building or campus that volunteers have mapped as a data center. This is the same source the U.S. Department of Energy’s Pacific Northwest National Laboratory used for its national data center atlas. It is the most complete free list available, but it is not verified and coverage is uneven.
2. Sorted into the same fields for every site. Each record is organized into name, owner or operator, status, location, power, size and tax incentives. Counties are looked up from each site’s coordinates using the U.S. Census Bureau’s public geocoder. Building sizes from map data are footprints traced from aerial imagery, not total floor space.
3. Duplicates removed. When the same campus appears in more than one source, the more detailed record is kept.
4. Researched rows are never overwritten. Rows rated ●●● are checked by hand against original documents, including tax incentive agreements. The nightly process does not change them.
5. Safeguards. If a source fails to download, the previous night’s data is kept. If the table would suddenly shrink by more than 30 percent, nothing is published until the cause is checked.
6. Searched and sorted without favor. Every row is shown by the same rules. The default order is by status, then by how much is verified, then by size. Readers can search and filter by status, state and detail level. No project is promoted, hidden or ranked by anyone’s preference, and no company or group can pay to change how it appears.

The tracker’s data file is public. Location data from OpenStreetMap is available under the Open Database License; Epoch AI data is used under Creative Commons Attribution.

How we choose which facts to show

1. Preference order. When several sources cover the same question, we prefer, in order: government audits and national laboratory data; academic studies; official records obtained by the public; company and trade disclosures; advocacy and news accounts.
2. Strongest case on each side. For each issue we look for the best-supported evidence each side relies on — not the weakest version of either side’s argument.
3. Same source for both sides when possible. When one study supports both perspectives (as with the Virginia noise and air quality findings), we use it for both.
4. Balance check. Before publishing, we compare the source types used on each side of every issue. If one side relies mostly on industry or advocacy sources while the other relies on government or academic sources, we look for stronger evidence or note the imbalance.
5. Plain language. Figures are rounded and explained in everyday terms, with the exact figure available at the source.

Waiting on original sources

These figures are held back or marked pending until we can link the original document.

• Virginia data center tax exemption, FY2025 (~$1.6 billion). Reported in the news from the state’s annual financial report; held until we link the report itself.
• Golf course water use (Golf Course Superintendents Association of America survey) — shown on the home page chart; link pending.
• Avocado and corn irrigation figures (California Avocado Commission; Texas Tech) — shown on the home page chart; links pending.
• Kent County, Delaware noise ordinance — currently cited through news coverage; link to the ordinance pending.
• Tracker tax incentives for Stargate Abilene, AWS New Carlisle, Project Jupiter and Meta Prometheus — reported in the news; awaiting the original county and city agreements.
• Noise-by-distance comparison table on the home page — estimates; original measurements pending.

Corrections log

No corrections since launch. Each future correction will be listed here with the date, what changed and why.

Independence

We take no funding from data center developers or opposition groups, and accept no advertising from any company connected to data centers.