The $500 Billion Build-Out Is Already Under Way
The epoch AI report makes the scale plain: OpenAI, Oracle, and SoftBank are together investing $500 billion in what they're calling Stargate, a nation-spanning suite of data centers that by 2029 should surpass 9 gigawatts of capacity. That's roughly the peak power demand of New York City. As of the report's cutoff—April 2026—0.3 GW was already operational in Abilene, Texas, and six more sites were under active construction. The pace is fast, but the road is long. OpenAI recently pulled back from expanding Abilene from 1.2 GW to 2.1 GW, redirecting that capacity elsewhere—a concrete illustration that plans can shift even after construction begins.
Seven Sites, One Unprecedented Goal
The project has seven locations across the United States. Abilene, Texas, is the most advanced, with four of eight buildings already hosting Nvidia Blackwell chips. Projected to reach 1.2 GW by Q4 2026, Abilene was supposed to go bigger, but OpenAI nixed the 2.1 GW expansion not long ago. Microsoft subsequently partnered with Crusoe on an adjacent 900 MW site, so the Abilene footprint remains substantial even if the original ambition scaled back.
Shackelford County, Texas, sits just across the county line from Abilene. Vantage is building a 1,200‑acre campus there with 10 buildings, projected to hit 2.0 GW by Q4 2028. Power will come from an on‑site natural gas microgrid. Roofing was already visible in satellite imagery as of mid‑April 2026, and Vantage expects the first building to deliver by late 2026.
Doña Ana County, New Mexico, is another Vantage project—Project Jupiter, four large buildings projected for 2.2 GW by Q4 2028. Two on‑site natural gas microgrids are intended to limit impact on the local grid. Foundation work was underway when the satellite imagery was captured in February 2026.
Milam County, Texas, about 70 miles northeast of Austin, is a "fast-build" site developed by SB Energy, a SoftBank subsidiary. Projected capacity is 1.2 GW by Q4 2028. SB Energy plans to fund and build new energy generation and storage to supply the majority of the campus's power. Steel framing and roofing were visible in a March 2026 satellite image.
Port Washington, Wisconsin, north of Milwaukee, is Vantage's "Lighthouse" campus, projected for 1.3 GW by Q4 2028. It's described as "sustainable‑by‑design," with 70 percent of power drawn from solar, wind, and battery storage. Foundation work was visible in February 2026 imagery.
Saline Township, southwest of Detroit, is Related Digital's "The Barn," projected for 1.4 GW by Q4 2028. DTE Energy will provide 100 percent of the power, augmented by a battery storage system financed by the project. Foundation work was underway in March 2026.
Lordstown, Ohio, is the seventh and most uncertain site. Some land has been cleared, but no large‑scale data center construction is visible. SoftBank and Foxconn have a joint venture there, primarily a manufacturing facility for AI servers and data center equipment. Projected capacity is under 0.3 GW. The site has already faced a ban on future data centers, and local opposition has been vocal. Power will come from the grid, as Foxconn already has a substation connected.
Power Choices: On‑Site Gas and Closed‑Loop Cooling
The report highlights two design choices that shape both timeline and cost. To avoid lengthy grid-connection queues, at least three of the seven sites will use on‑site natural gas plants. That shortcut buys time but raises facility costs. To address public concern about water usage, at least six sites will employ closed‑loop liquid cooling systems, which don't evaporate water. Milam County's closed‑loop design wasn't confirmed in the report, but it's described as designed to minimize water usage.
These choices are practical compromises. On‑site gas avoids the interconnection queue lottery, but gas infrastructure isn't free. Closed‑loop cooling saves water but may be less energy‑efficient than evaporative systems, especially in Texas's hot summers—the report notes that Shackelford County's estimated total facility power was bumped up to about 2 GW precisely because of closed‑loop cooling's efficiency trade‑off.
The Road Ahead: Financing, Politics, and Plan Changes
At this point, Stargate is more than pure ambition. Construction is underway across the US, and there's enough time to finish by 2029. But the road is long. Plans can change after construction begins—Abilene's expansion reversal is the clearest example. Financing and procuring equipment at this unprecedented scale will be challenging. Political opposition is real: Lordstown has a ban on future data centers, and the Michigan site has faced local opposition.
The notes section of the source adds some comparative context. The NYISO 2025 Gold Book forecasts about 11 GW of peak summer demand for New York City (Zone J) from 2026 through 2030, the single highest annual hour of demand. The H100 reference is just that—a reference point. The actual chips in these data centers will probably be Nvidia Blackwell, and later Nvidia Rubin. Total global AI compute by the end of 2025 was estimated at about 20 million H100‑equivalents, based on the AI Chip Sales database. Stargate's projected compute is estimated from power capacities and the trend in energy efficiency for leading machine‑learning hardware, except for Abilene, where Crusoe disclosed 50,000 Blackwell GPUs per building. All stated power capacities refer to total facility power, including GPUs, cooling, lighting, and so on—not just IT power. For some sites, total facility power was estimated from reported IT power.
Conclusion
The Stargate initiative puts $500 billion on the line across seven states, with capacity projected to more than nine gigawatts by 2029. The build-out is already moving—0.3 GW operating in Abilene, six sites under construction—but the final shape will depend on financing, equipment delivery, and whether political or technical setbacks alter the course. Epoch AI says it will keep following the project, and so should anyone tracking the intersection of AI ambition and U.S. electricity demand.
Source: Elliot Stewart and Ben Cottier (2026), "OpenAI Stargate: where the US sites stand". Published online at epoch.ai. Retrieved from https://epoch.ai/publications/openai-stargate-where-the-us-sites-stand