How the directory is built, and how often it is wrong
Every campus row has a source. Every capacity figure says who states it. We check the rows against independent sources, count the errors, and publish the result here, including where we fall short.
The short version
In audit #11 (November 1, 2026), the share of capacity (MW) figures found wrong was 1.2% when wrong means a source contradicts the figure, 1.2% when a power, utility-contract, program or phase figure filed as the campus total also counts as wrong, and 39.5% when only a word-for-word statement of the campus load counts as right.
Audit #11 meets every field pass mark. Capacity (1.2%), status (0.0%), location (0.5%) and developer (0.0%) meet theirs. Capacity must be 5% wrong or less, with the top of the 95% interval at 10% or less; status 3% or less; location 2% or less; developer 2% or less.
Estimated cost is the weakest field: 5.4% wrong (5 of 93). All five wrong costs were taken down until a source states what building the campus costs. Two counted the computer equipment inside the campus. One was a modelled cost for the first building only. One was a figure no publisher states. One did not match the developer's own filing.
We call the data proven only after two audits in a row, on independent samples, meet every mark. Audits #10 and #11 do.
16 of the 17 errors the audit confirmed were corrected through a sourced proposal, and the other one is still open. Next is audit #12.
What a row is
A row is one named data center campus. It is not every building on the site. When a campus is built in phases, each phase hangs off its campus row, and the campus is counted once.
We also track power plants, chip fabs and things that are not campuses, like a supercomputer inside another operator's hall or a cloud region. They have pages, but they never count toward campus, state or MW totals.
Coverage is named campuses, mostly 30 MW and up. Why our count is smaller than some others: they count buildings of every size, from server rooms to colocation suites. We track the large campuses being built, each with a source.
Where a number comes from
Every row links at least one source: an article, a press release, a planning record, a utility filing or the company's own page. An aggregator listing is never the source of a fact.
Nobody outside the builder can measure a campus's megawatts. So we never call a figure official. We show who said it, and when.
- Filed: a record states the figure (an SEC filing, a permit or planning record, or a utility filing).
- Stated: the developer or the press states the figure.
- Estimated: no source states the exact figure, so it is modelled or inferred. That includes Epoch AI's IT-power model. Lists and rankings mark these figures "est."
Filed and stated figures carry an "as of" date, the date of the newest source that states them. A rule sets the label: a figure only counts as stated when a source quote gives that exact number for the campus itself.
How it stays right
- Every change a person makes goes through a proposal that names one source and quotes it. It is checked before it is applied, and nobody edits a row by hand.
- A correction locks the field, so the nightly scrape cannot swap a sourced value for a modelled one.
- A daily check runs over every number the site shows, including the figures on this page.
- Sources for the Tier 1 campuses are saved to the Internet Archive's Wayback Machine.
- Possible duplicates are flagged by rule, and pairs confirmed to be different campuses are recorded as checked.
How accurate it is, measured
On November 1, 2026, in audit #11, we audited our Tier 1 rows: the 285 campuses that matter most, meaning the largest by MW plus every US campus being built. We checked all 25 of the largest and a random draw of 175 more, drawn with a fixed seed so it can be repeated.
Each field was checked against one independent second source, found by search rather than taken from the row. Each rate comes with a 95% confidence interval (the Wilson method).
Capacity, counted three ways (same 81 figures)
This is the base our pass mark uses. Figures labelled estimated are left out of it.
- Contradicted by a source: 1.2% (1 of 81, 95% interval 0.2 to 6.7%). A source contradicts the figure.
- Plus scope errors: 1.2% (1 of 81, 95% interval 0.2 to 6.7%). Also counts a power, utility-contract, program or phase figure filed as the campus total.
- Word-for-word only: 39.5% (32 of 81, 95% interval 29.6 to 50.4%). Only a word-for-word statement of the campus load counts as right.
If you are quoting the size of a campus, read the strict count: 32 of the 81 figures fail it, either because a source contradicts them or because they rest on a ceiling, a plan, a power figure or arithmetic rather than a stated campus load. Another 51 capacity figures are labelled estimated on the site, because no publisher states a figure for the whole campus, and 68 could not be checked at all, because no independent source states them. Neither group is in the 81.
Audit #10, on October 31, 2026, found 3.3%, 3.3% and 40.0% of 90 figures.
Every field, the 200 sampled rows
This table covers the 25 largest campuses and the random draw. In all we read 271 rows. That includes 19 rows picked because an earlier audit had corrected them (some are also in the sample), 61 campuses that joined Tier 1 after it was frozen, and a few open questions left from the last audit. Those extra rows are reported on their own, not pooled here. On the re-checked rows, 0 of 7 capacity figures (0.0%) were wrong.
| Field | Wrong | Rate | 95% interval | Pass mark |
|---|---|---|---|---|
| Capacity (MW) | 1 of 81 | 1.2% | 0.2 to 6.7% | 5% or less, interval to 10%: pass |
| Status | 0 of 179 | 0.0% | 0.0 to 2.1% | 3% or less: pass |
| Location | 1 of 197 | 0.5% | 0.1 to 2.8% | 2% or less: pass |
| Developer | 0 of 196 | 0.0% | 0.0 to 1.9% | 2% or less: pass |
| Expected completion | 1 of 84 | 1.2% | 0.2 to 6.4% | Not gated |
| Tenant | 1 of 118 | 0.8% | 0.1 to 4.6% | Not gated |
| Estimated cost | 5 of 93 | 5.4% | 2.3 to 12.0% | Not gated |
| Campus name | 0 of 200 | 0.0% | 0.0 to 1.9% | Not gated |
Capacity (MW): The mark also needs the top of the interval at 10% or less. One figure was wrong. We showed a campus's 75 MW first stage, while the city says the whole site is approved for a 150 MW load. It was corrected.
Location: The one miss is a campus on federal land that spans three counties. No source names which county the site is in, so it is still open.
Tenant: Most campuses have no public tenant, so fewer could be checked and this rate is loose.
What we are fixing before audit #12
- Checking the estimated cost of every Tier 1 campus. A cost should cover building the campus. It should not include the chips inside it, a wider program or only the first phase. Cost was the weakest field in audit #11. Three of the costs audit #10 had corrected turned out to be wrong.
- Making sure every saved copy of a source really exists before we link to it. Audit #11 found 13 links to saved copies with nothing behind them. Each one was saved again or taken down.
- Looking again for the county of the Savannah River Site data center. The site is on federal land that spans three counties. We can name the county once the lease names the tract.
- Looking again at three other open questions when new sources appear. They are the exact site of one Texas campus, whether one Virginia campus has started building, and a saved copy of one local news page.
- Watching two campuses that may change soon. One becomes cancelled if no appeal is filed. The other may stop mining at the site after its owner reports results on November 5.
Audit #12 runs on a fresh random draw from the same frozen set of Tier 1 campuses, so the results can be compared. It has no fixed date.
Found an error?
Open the campus page in the directory, expand "sources" at the bottom, and use "Suggest a correction or add a source". Include a link that states the right figure. You can also email hello@datacenterindex.ai.
Licences and credit
- Epoch AI, Frontier Data Centers, CC BY 4.0.
- AI Data Center Index (SheetGenius, Inc.), CC BY 4.0. We use its public browse pages only.
Both credits travel with every export row and every API response. More on the site and its house rules: About.