What Is the Data Center Index? A Sourced Campus Directory

The Data Center Index is a directory, and it moved again this week. Over the last seven days it took in 78 new campuses, changed the status on 31 of them, and corrected 145 figures that were already in the file. The word index here is the kind you find at the back of a book, a list you look things up in. The things are data center campuses.

Right now it counts 618 of them. Those rows hold roughly 250 gigawatts of capacity between them (a gigawatt is a thousand megawatts, and one large campus draws a few hundred of those). Only 40.6 GW of it is running today, so most of what you're reading is still a plan with a permit attached. Announced cost comes to $1.9 trillion across the 214 campuses that carry a cost figure at all. You can find aggregate numbers like that one anywhere and do almost nothing with them, so this is built the other way round. Every campus gets its own page with an address, a developer, a status, a capacity figure, and the article, permit or filing that put it there. When a number has no publisher behind it, the page says so.

What's in the Data Center Index right now

Of the 618 counted campuses, 190 are already operational and another 179 are under construction. Earlier in the pipeline, 92 have permits in hand and 157 have only been announced. The file is mostly American, with 490 rows across 43 states, while the other 128 sit in 48 countries and the biggest groups outside the US are in China, France and Germany. You can browse them by state market or straight down the project list.

Some rows are tracked and deliberately never counted. Nine paused campuses and two cancelled ones keep their pages and their sources, because a project that stalled in front of a county commission is worth knowing about, but neither one adds to a state total. The same goes for 147 rows that aren't campuses at all, like a supercomputer installed inside somebody else's hall or a company-wide spending program, plus 17 power plants and 11 chip fabs that show up in the same news cycle. They're listed, they just never count as data centers.

Coverage runs campus by campus rather than building by building. A county with forty halls going up can show as six named campuses here, because a campus only enters the file when a source names it, and plenty of buildings go up without anyone naming them in print.

Every capacity figure says who stated it

Nobody outside a campus's own engineering team can verify its megawatts, so the directory doesn't pretend to. Instead every capacity figure carries a label naming who stated it, and those labels are the most useful thing on the site.

• Filed means a record states the number: a permit, a planning docket, a utility filing, an SEC filing.
• Stated means the developer said it, or the press reported it from the developer.
• Estimated means somebody modelled it, including the figures we take in from Epoch AI, and no publisher ever stated the number outright.

Each label carries the date of the newest source that earned it, so you can see whether a number is from July or from two years ago. And the tiers are strict about what counts as proof. A quote only lifts a figure out of estimated when it states that exact number for the whole campus, so a phase figure can't stand in for a campus total, and "100 MW+" doesn't state 100 MW.

Only three campuses carry a capacity figure that a filing or a permit states outright, and the largest of them draws 75 MW, in Mantsala, Finland. Permits mostly describe acreage and water, while press releases describe ambition, so 174 rows sit in the estimated tier because no publisher ever stated a load for them. Anything in that tier prints an "est." badge next to it on every list, ranking and export, so you never have to remember which is which.

A good row looks like Meta's Hyperion campus in Holly Ridge, Louisiana: 5,000 MW, labelled stated, dated July 13, 2026, which is the day Meta published the expansion on its own site. You can click through to the source and read the sentence the number came from.

What a project page shows

Open a campus and you get the developer, the tenant when there's a public one, the city and state, the status, the capacity with its label and date, the announced cost, and a completion date. That last field is narrower than it looks. Expected completion means the date the next building goes live, so a 2029 county deadline covering an entire site never goes in it, because that deadline tells you nothing about what energizes next spring.

Under that sits the source list, with every URL, what it said, and when we last fetched it. Then the public companies tied to the campus, the campuses nearby, and where this one ranks by size. Those company links are the part investors use most: 66 public companies are attached to campuses today, and each one has a page carrying capex, SEC filings and price history.

When two sources disagree, the newer primary record wins the field and the other reading goes into the notes underneath, with its publisher and its date. Nothing gets quietly dropped to make a page look tidy. The whole file also comes out as CSV or GeoJSON, and there's a REST API plus an MCP server (the protocol an AI assistant uses to pull data straight from a source) if you'd rather have software read it than read it yourself.

How wrong it is, and how we know

A directory that tells you it's accurate without showing its arithmetic is asking for trust it hasn't earned. So the errors get counted. The last audit closed on September 24, 2026, with 129 rows re-read against independent sources: 90 errors found, 87 of them corrected the same week.

Capacity was contradicted by a source on 5.7% of the figures that could be scored, and because the sample is small the 95% confidence interval runs from 2.2% to 13.8%. Developer came in at 1.2% wrong and location at 4.9%, while status missed on 7.3%. The worst field by a distance was expected completion at 36.6%, and nearly every one of those was a date set to the full build-out instead of the next building, which is the rule we changed on September 21. Our own pass mark is 5% with a confidence ceiling of 10%, so capacity hasn't cleared it yet and the page says that out loud.

The methodology page carries the full table, the sourcing rules, the sample design and the rate for every field. It gets rewritten after each audit rather than after each good quarter.

What I'm watching

618 campuses is where the count sits this week, and it'll read differently by Friday. The next audit lands in late October, and capacity is the number I want under 5%.

Tags: directory, methodology, capacity, pillar