The headline number is $500 billion over four years, scaling to ten gigawatts of AI compute by 2029, roughly New York City's peak demand. Three years into the build, six of the seven sites are dirt. One is running at about a third of its planned capacity.
The headline number is $500 billion over four years, scaling to ten gigawatts of AI compute by 2029, roughly New York City's peak demand. Three years into the build, six of the seven sites are dirt. One is running at about a third of its planned capacity.
Three are in Texas. One in New Mexico. One each in Wisconsin, Michigan, and Ohio. A satellite would show you graders, fences, and concrete being poured in four of them.
Only Abilene is operational. About 0.3 gigawatts. The other six combined produce zero.
That is the physical reality of Stargate today.
Abilene was built by a company called Crusoe, which was best known until recently for using flared natural gas at oil wells to mine bitcoin. The relevant insight Crusoe brought to AI infrastructure was that they knew how to spin up a gas turbine fast.
The first 200 megawatts came online with a mix of on-site GE Vernova gas turbines and West Texas grid power, which draws heavily on local wind. On-site gas is the primary supply, and the entire site is projected to reach 1.2 gigawatts by Q4 2026.
Of the 9 gigawatts on the map, roughly half is going to two Texas sites and the New Mexico site. All three of them rely on private on-site natural gas. They are not connected to the grid in any meaningful sense.
The remaining sites are split into smaller portions across Ohio, Wisconsin, and Michigan, all grid-connected, all considerably smaller. The skew matters because it tells you that the planners optimised for time.
Here's the strange part of the New Mexico decision. The local grid is around 38% wind and 11% solar, one of the cleanest electricity mixes in the country. STACK Infrastructure, the developer, still chose to build two private natural gas microgrids on site.
Texas is the same story. Texas leads the United States in wind generation. Stargate's three Texas sites all bypass the grid in favour of on-site fossil fuel.
It's because connecting to the grid takes too long.
ERCOT, the Texas grid operator, is currently sitting on roughly 432 gigawatts of generation in its interconnection queue. About 1.8% of that is actually operational. More than half of the requests have not submitted enough information for ERCOT even to begin a study. There is a term for the rest: phantom load.
If you cannot reasonably believe the grid will deliver, you build your own plant. That is what Texas does to you.
Three sites avoid on-site gas. Lordstown, Ohio. Port Washington, Wisconsin. Saline Township, Michigan. Look at what their grids actually run on.
Ohio's grid is 44% coal and 40% gas. 84% fossil-fueled, the dirtiest electricity mix of any Stargate site. Wisconsin sits at 40 per cent gas and 32 per cent coal. Michigan at 45% gas and 21% coal, with about 22% nuclear, softens the picture. The grid-connected sites are drawing from grids that run on roughly the same fuels Texas and New Mexico chose to build privately.
Grid-connected doesn't mean what people assume. It means someone else owns the gas plant.
In November 2025, the Lordstown village council voted six to nothing to ban future data centres. The Stargate project itself slipped through because it was announced earlier and is technically a Foxconn equipment factory with a small data centre attached. A separate $3.6 billion AI project from a different developer is now suing Lordstown in the Ohio Supreme Court.
Around eighteen Ohio municipalities are considering or have already passed similar moratoriums. Ohio's largest utility forecasts that data centre demand could grow from 600 megawatts in 2024 to 5 gigawatts by 2030, even after policies that cut its pipeline in half.
Michigan's Saline Township tried to refuse the 250-acre Stargate site outright. Some of the township's own residents sued, the ones with land to sell. The township settled. It got water-use limits and money for the local fire department, in exchange for letting the project proceed.
In March 2026, OpenAI and Oracle quietly cancelled a planned 600-megawatt expansion in Abilene. The official reason was that capacity was being redirected to other sites. The real reason, according to people tracking ERCOT's queue, is that even Stargate couldn't get the additional power approved fast enough. Microsoft picked up the adjacent 900 megawatts through a new partnership with Crusoe.
If you'd told an AI executive in January 2025 that the flagship Stargate site would have its own expansion cancelled by March 2026, they would have laughed at you. The headline number is still $500 billion. But the first crack has appeared.
SoftBank and OpenAI each committed $19 billion in equity. Oracle and MGX each put in $7 billion. The remaining roughly $450 billion has to come from debt and limited partners. KeyBanc estimates Oracle alone may need to borrow up to $100 billion over four years to meet its commitment. Then, Nvidia announced a $100 billion investment in OpenAI. OpenAI plans to use much of that to buy Nvidia chips.
It is a sophisticated arrangement of equity, debt, and supply commitments between aligned partners. Or as a circle. Both are technically correct. Short sellers like Jim Chanos and Michael Burry have started using the second word.
The interesting thing about Stargate isn't the chips or the buildings. It's what the seven sites tell you about the country where those buildings are being built.
The grid wasn't ready. The permitting system wasn't ready. The political coalition turned out to be less stable than it looked at the White House launch. The financing depends on circular flows between a small number of companies. The physical reality, three years in, is that one site is partially running and six sites are in foundation work.
None of this means Stargate fails. It probably gets built, mostly. The thing sitting in the data is that even at $500 billion, with full administration backing, what you're actually buying is a wager against your own infrastructure. You're betting the gas turbines arrive on time, that the courts rule the right way, and that the small towns don't organise.
Those are industrial bets, the kind we mostly stopped making sometime around 1972. We're going to find out, in public, whether we still know how.
Site capacities, power source assignments, construction timelines, and projected completion dates from Epoch AI's report "OpenAI Stargate: where the US sites stand" by Elliot Stewart and Ben Cottier, published April 2026 and updated April 23 to reflect Oracle's revised Abilene operational figure.
Grid generation mix figures derived from the US Energy Information Administration 2024 state-level electricity data. Interconnection queue figures from Lawrence Berkeley National Laboratory and ERCOT board materials cited by Dave Friedman in January 2026. Local opposition coverage from Datacenter Dynamics, Signal Ohio, Spectrum News, the WFMJ News local desk, and the Washington Post. Financing details and circular-funding critique from S&P Global, Wall Street Journal, KeyBanc Capital Markets research notes, and CNBC interviews with OpenAI CFO Sarah Friar.
The framing of Stargate as a power-and-permitting story rather than a compute story draws on Enverus Intelligence Research's October 2025 outlook on US data centre capacity and on the Logistics Viewpoints analysis of FERC Order 2023 and the rise of behind-the-meter microgrids.