Eight months ago, Stargate was a podium and a press release. This week, the Stargate data centers became poured concrete. On September 23, standing at the flagship site in Abilene, Texas, OpenAI, Oracle, and SoftBank announced five new US data center locations. It confirmed that Abilene itself is now up and running, full of Oracle Cloud infrastructure and racks of Nvidia chips. When I wrote about the half-trillion-dollar Stargate announcement back in January, my main worry was the gap between the number and the buildings. This update is the first real answer to that worry, and it is more convincing than I expected.
What changed
The five new sites are spread across Texas, New Mexico, and Ohio: Shackelford County and Milam County in Texas; Doña Ana County in New Mexico; Lordstown in Ohio; and an additional Midwest site. Three are Oracle-led, two are SoftBank-led. Combined with the Abilene flagship and the ongoing CoreWeave projects, that brings Stargate to nearly 7 gigawatts of planned capacity and over 400 billion dollars of investment over the next three years. To put 7 gigawatts in perspective, New York City runs on about 6. These are not data centers in the old sense; they are power-plant-scale industrial sites that happen to compute.
The companies say this puts them ahead of schedule to meet the full $ 500 billion, 10-gigawatt commitment by the end of 2025. There was also a July agreement between OpenAI and Oracle to develop up to 4.5 additional gigawatts of capacity, a partnership the two say will exceed $ 300 billion over five years. And the five sites were picked from over 300 proposals across more than 30 states, which tells you every county in America with land and power now wants a piece of this.
Why Abilene actually matters
Here is the part that moved the needle for me. Abilene is live. Oracle began delivering the first Nvidia GB200 racks in June, and OpenAI says it has already started early training and inference workloads on that capacity. That is the difference between this update and the January one. In January, we had a number; now there is a running supercluster about 180 miles west of Dallas with real chips drawing real power. OpenAI’s CFO Sarah Friar made a sharp point on the timeline, too: the shovels going into the ground now are laying foundations for compute that will not come online until 2026, starting with Nvidia’s next-generation Vera Rubin chips. So what you are watching is a buildout with a multi-year fuse, where today’s groundbreaking is 2026’s compute.
Personally, seeing actual GB200 racks in operation is what makes me take the 7-gigawatt figure more seriously than the original 10-gigawatt figure. A number on a White House lectern is a promise. A humming data center in Texas is a fact. C’est the proof-of-life the project needed.
Where I am still skeptical
I am not fully sold, though, and it would be dishonest to pretend otherwise. Much of this capacity is still planned rather than delivered, and how many of these projects ultimately get funded remains an open question. OpenAI is still losing billions of dollars a year, and “nearly 7 gigawatts of planned capacity” is doing a lot of work in that sentence; planned is not the same as built. The pattern with Stargate has been to announce capacity in big round numbers and let the delivery catch up later. One running site in Abilene is genuine progress. It is not yet 7 gigawatts.
Still, the trajectory is real and fast. In eight months, Stargate went from a contested announcement to a live supercluster plus five new sites chosen from hundreds of bids. Whatever happens to the final tally, the AI infrastructure race is clearly accelerating, not stalling, and the map of where America’s computing will physically live is being drawn right now, county by county. I will keep watching whether the planned gigawatts turn into running ones, because that conversion rate is the whole ballgame. And the field was about to get more crowded: weeks later, Anthropic and Microsoft made their own massive infrastructure moves, with the model labs themselves turning into landlords.