Data centers, explained

How much is a megawatt?

A megawatt is a rate of power draw, the way miles per hour is a rate of speed. A data center’s megawatt figure is how much electricity it pulls at full tilt, around the clock, every day of the year. Across this site we use one conversion everywhere: a megawatt of data center load draws roughly what 750 American homes draw.

• 5 MW
  A small colocation site

  About 3,750 homes’ worth of demand. Most dots on this map are this size.

• 100 MW
  One hyperscale building

  75,000 homes’ worth, drawn by a single building running around the clock.

• 1,000 MW
  One gigawatt

  The output of a large nuclear reactor, feeding roughly 750,000 homes.

• 10,000 MW+
  The announced mega-campuses

  Ten reactors’ worth. Nothing this size exists anywhere yet; every figure like this on the map is still a plan.

We don’t publish a total for the map, and the reason belongs in a page about vocabulary: about four in ten of these facilities never disclose a capacity, because operators aren’t required to, and when we read a random sample of the rest against their own sources the published figure was wrong about half the time. Adding those numbers up would produce a gigawatt total that looks like a census of American compute and is nothing of the kind. Per-facility capacity appears wherever a source states it. For scale, the entire United States runs on about 470 gigawatts on an average day.

The announcement gap

Capacity gets quoted at four different stages, and a headline number rarely tells you which stage it came from. The map draws every project by stage, as a light, so you never have to guess.

Proposed

A company says it intends to build. Sometimes that means permits and purchased land, sometimes a press release and a land option. The biggest numbers on this map live here, and some of them will never break ground.

Under construction

Earth is moving. Even here, the quoted figure usually describes full buildout years away, and a first phase can be a tenth of the headline.

Operational

Built and drawing power. The firmest numbers on the map, though operators rarely disclose exact loads.

Canceled or withdrawn

Proposed, then stopped: rejected by a county, pulled by the developer, or blocked by a moratorium. There are 469 of these on the map, drawn as lights that are off.

There’s a second trap inside the big announcements. Mega-campus press releases usually quote the power the developer plans to generate on site rather than computing load anyone has signed for. When we re-verified the largest announced projects on this map, we found one 10 gigawatt campus whose only signed tenant so far is a 50 megawatt deal. That’s a 200-to-1 gap between the headline and the contract.

The announced figure still matters: it’s what the developer is asking a county to approve, and what the grid is being told to prepare for. But it’s a request, and where a project is still on paper, you should read its capacity as ambition. Our project notes say which one you’re looking at.

How do we size a facility that doesn’t publish capacity?

About two thirds of the facilities on this map never publish a megawatt figure, because operators aren’t required to. Rather than leave all of them in one unlabeled pile, we class each one by what the record does hold, and we label the result for what it is: an inference, never a measurement. A carrier’s cable landing station, a university’s machine room, and an unannounced hyperscale campus are different things, and a map that renders them identically is hiding what it knows.

The strongest signal is the operator’s own disclosed sites. When a company does publish figures for some facilities, the median of those figures says what kind of thing that company builds. These are the medians in our record today, recomputed nightly, for every operator with at least ten disclosed sites:

1. Flexential27 of 53 sites disclose a figure6 MW
2. DataBank34 of 57 sites disclose a figure6.7 MW
3. Lumen Technologies12 of 99 sites disclose a figure6.8 MW
4. EdgeConneX10 of 27 sites disclose a figure9 MW
5. Equinix18 of 95 sites disclose a figure13.5 MW
6. Iron Mountain13 of 22 sites disclose a figure14 MW
7. NTT Global Data Centers16 of 21 sites disclose a figure36 MW
8. T5 Data Centers12 of 19 sites disclose a figure36 MW
9. Digital Realty16 of 93 sites disclose a figure38 MW
10. CyrusOne29 of 59 sites disclose a figure40 MW
11. Sabey Data Centers10 of 16 sites disclose a figure54 MW
12. QTS Realty12 of 28 sites disclose a figure75 MW
13. Stack Infrastructure10 of 16 sites disclose a figure92 MW
14. Aligned Data Centers14 of 29 sites disclose a figure102 MW
15. Vantage Data Centers21 of 33 sites disclose a figure120 MW

So an unannounced Google site is classed large, because Google’s disclosed sites run hundreds of megawatts, and an unannounced Lumen site is classed a network site, because Lumen’s run about two and a half. Status matters too: a facility already operating is existing load, not new demand. Name matters where it’s explicit, so “cable landing station” reads as carrier infrastructure and “AI campus” on a proposal is never quieted. And where nothing in the record signals scale at all, a single-site operator already in service with no published figure, we rank it small and say exactly why on its page.

Every inferred class is labeled “inferred from the record” on the facility page, with the evidence sentence beside it, and an inferred class never becomes a megawatt number anywhere on this site. Where a source states a figure, the figure is shown and marked measured. Where it doesn’t, the capacity field says “not disclosed” and the class carries the reader instead.

The vocabulary

Hyperscale

A campus built by or for one giant tenant, think Microsoft or Meta. Hundreds of megawatts and up, often several buildings on one site.

Colocation (colo)

A landlord data center. The operator owns the building and rents rack space and power to many tenants. Most of the existing fleet is colo.

Interconnection queue

The waiting line to plug into the grid. Every big new load or generator files a request with the regional grid operator. The queue is public in aggregate, and it’s the earliest signal of what’s coming.

Large load

Grid-operator term for a single customer big enough to matter to the whole system. In Texas, ERCOT tracks loads of 75 MW and up in a dedicated queue.

Generation vs. load

A power plant generates electricity; a data center consumes it. Mega-campus announcements usually quote the generation the developer plans to build on site rather than computing load anyone has signed for.

Behind the meter

On-site power feeding the campus directly, bypassing the public grid. Faster to build, and much harder for regulators and neighbors to see coming.

PUE

Power Usage Effectiveness: total facility power divided by computing power. A PUE of 1.5 means half again as much electricity goes to cooling and overhead as reaches the servers.

MGD

Million gallons per day, the unit for water draw. Evaporative cooling at a large campus can run to millions of gallons on a hot day, from the same aquifers nearby farms depend on.

How to read this map

• Every light links to its public record. Open any project and you’ll find the source we traced it to: industry reporting, a company announcement, local news, or a planning filing. If a source stops holding up, the listing comes down.
• The verification badge records a check, not a vibe. “Source-checked” means a person read the facility back against the document it cites, and the badge carries the date. Everything else honestly reads “Source on file,” which is what makes the checked ones worth something.
• The protected-land flag. Projects within 10 miles of protected public land carry a flag; 496 currently do. That threshold used to be fifty miles, and fifty caught 1,606 facilities, close to seven facilities in eight, which is a description of the United States rather than a finding about data centers. We moved it to ten in September 2026 for that reason. The whole distribution stays on the page underneath the flag, because a threshold is a choice and a reader is owed the shape it was cut from: 14 sit within a mile, 62 within three, 182 within five, 496 within ten, 739 within fifteen, 1,189 within twenty-five, 1,606 within fifty. Distance is measured to the boundary of the protected unit, not to a point at its center, which is a correction we made in August 2026 after the old method put an Alaska project at zero miles from a refuge it was 48 miles away from. Proximity is context either way, and it says nothing about a project by itself.
• Found something wrong? Email ask@morethanjustparks.com and we’ll check it against the record and correct it.

Who pays for a data center

A data center arrives as an economic development story: jobs, tax base, a company nobody expected to come here. The costs arrive later and they don’t land on the company. They’re worth knowing before a county votes, because almost every fight on this map is really an argument about this.

Electricity. A large campus draws what a small city draws, and the wires and plants to serve it get built by the utility. A regulated utility earns a guaranteed return on what it builds, so the more it builds the more it earns, and the bill goes to everyone on the system. If the projected demand never shows up, the infrastructure is still there and the customers still pay for it.

Water. Evaporative cooling consumes water rather than borrowing it. A closed-loop site consumes far less on site, but the power plants feeding it consume their own, and the on-site figure is the one that gets quoted. Construction is its own draw, and it runs for years: concrete, dust control, flushing and filling the systems in each new building.

Public money. Most states exempt data center equipment from sales tax, and many counties add abatements on top. Those are real dollars that a school district or a county budget doesn’t collect, handed to an industry with some of the largest profits in the world, for a facility that employs a few dozen people once it is running.

The land and the people next to it. Substations, transmission corridors, generator testing, security lighting and around-the-clock cooling noise. These land on the nearest neighbors, who are rarely the people who voted.

We reported one of these from beginning to end. In Fayette County, Georgia, an unelected development authority brokered the sale of 615 acres for $153.8 million and kept $75 million of it in an account no elected body controls. The operator told a chamber of commerce audience the campus used about as much water as an office building. The county water system had already found unauthorized and unbilled connections that ran to roughly 40.7 million gallons, during a drought, after a neighborhood lost water pressure. The whole account is in The Great Data Center Swindle.