President Trump announced in January 2025The $500 billion Stargate AI data center projectAt that time, the three companies responsible for the development had promised to advance at lightning speed - immediately investing US$100 billion to build new data centers with a total computing power of 10 gigawatts.
More than a year later, according to three people involved in the stranded plan,The Stargate joint venture has not yet established a team, nor has it built any data center for OpenAI. Within weeks of the project being announced, Stargate had stalled and lacked leadership and coordination, these people said. The three parties of Stargate - AI model developer OpenAI, cloud service provider Oracle, and Japanese conglomerate SoftBank - have serious differences in division of labor and cooperation structure.
Therefore, OpenAI, which is extremely hungry for computing power, can only break out on its own. The company was preparingDo most of the construction work yourself, plans to directly lease, and even in some projectsautonomously heldLarge-scale data center campus to reduce dependence on other cloud vendors and strengthen control over computing power.
But the plan ultimately ran aground - lenders were unwilling to provide multi-billion-dollar project financing for a company with an unproven business model that was burning through billions of dollars a year.
It wasn’t until several months later that OpenAI and its CEO Sam Altman faced this financial reality and returned to advancing the project with Stargate partners. Even so, cooperation is mostlyOpenAI and SoftBank, OpenAI and Oraclebilateral agreement rather than tripartite coordination.
Despite a bumpy start, Stargate's data centers have finally begun to take off. OpenAI no longer aggressively builds and operates itself and uses debt to expand;Heavy reliance on collaboration with cloud partners, adopting a model that can improve the control of computing power without investing a large amount of principal.
At the World Economic Forum in Davos last month, OpenAI Chief Financial Officer Sara Flea made this shift clear. In an interview with Jessica Lessing, editor-in-chief of "The Information", she said that OpenAI is "Leverage cloud partners to maintain a lighter balance sheet".
"We currently have no plans to fully build it ourselves. Frankly speaking, we have very good partners." She said, "But what will happen in three years, no one can say."
According to two OpenAI employees, the company still intends to build its own data center, butNo longer a near-term priority. Instead, it is signing a series of cloud service and infrastructure agreements under the banner of "Stargate", some of which give OpenAI special permissions in data center design.
Spokespersons for OpenAI, SoftBank and Oracle declined to comment.
The impact of project stagnation
It is difficult to determine whether the obstruction of Stargate progress has slowed down the development of OpenAI models, but in the past year, its competitors Google and Anthropic have become more serious threats.
Last year, OpenAI failed to complete its set goal: locking in the next three years through SoftBank and Oracle10 GWComputing power commitment. However, as Stargate negotiations have taken a long time, OpenAI has also signed contracts with other suppliers such as Amazon Cloud and Google Cloud to obtain much-needed computing power. The company also said it would use non-Nvidia chips and reached an agreement with AMD and chip startup Cerebras to expand its computing power supply.
OpenAI’s computing power dilemma has also dragged down its financial performance: last year, it purchased computing power at a temporary high price.Gross profit margin lower than expected. The company last year expected to invest in computing power by 2030$450 billion, but that forecast has been revised upwards in recent weeks to$665 billion.
After the White House announced the Stargate project, OpenAI executives began planning a path to build their own data center. Employees travel to various parts of the United States to choose a location that can support800 MW to 1.2 GWFor single-park-scale locations, priority will be given to areas that can provide large-scale power in 2026-2027.
But exactly how the Stargate joint venture operates and what roles Oracle and SoftBank play have remained unclear.
According to people involved in the plan, some executives have proposed splitting Stargate into an independent entity, responsible for building facilities and leasing them back to OpenAI, and some employees will also be transferred to Stargate. Management has also discussed whether Stargate can be used as a financing entity to raise funds for chips and infrastructure.
Self-construction plan was postponed
None of the above ideas have been implemented. OpenAI realized after site selection and calculation that the cost of self-financing was too high. If a company with better credit qualifications such as Oracle signs a lease and then sublease it to OpenAI, the cost will be much lower - this is also the model it adopts in other data center cooperation.
It wasn't until the spring of that year that the OpenAI Stargate team met with Oracle's senior management team (including Clay Magulk, then head of the cloud business and now co-CEO), and the two parties finalized a new plan:Oracle reaches blockbuster data center agreement directly with OpenAI, totaling 4.5 GW.
According to participants at the meeting, this overall structure no longer signs projects one by one, but spreads risks across multiple parks.
About two months later, OpenAI suspended plans to build its own data center. Parks that had been finalized previously (such as the Vantage data center project in Wisconsin) were also included in the new agreement and handed over to Oracle.
Finally, at the end of July, Oracle and OpenAI announced a cooperation to build AI in many places in the United States.4.5 GWdata center.
According to two people familiar with the matter, the two parties reached an unusual arrangement:Sharing the economic risks of this large-scale project.
This means that if the project is delayed or overrun, the costs are shared between the two parties; if the final cost is lower than budget, the savings are also shared between the two parties. These details have never been disclosed before.
Only control power, not self-restraint
After Oracle announced the cooperation, OpenAI employees restarted preparations for building their own data center and restarted negotiations with equipment suppliers in order to quickly advance new projects.
One of the key sites: one in Milam County, Texas1 gigawattcampus, located between Austin and Waco.
During an inspection in early 2025, OpenAI employee Keith Hyde once joked with his colleagues that he could judge the clay content by tasting the soil. After that, he squatted down, pinched a little soil and put it in his mouth. The scene became an internal joke and reflected the team’s extreme dedication to site selection. (Hyde left OpenAI a few weeks ago.)
OpenAI had drafted plans for the project earlier that year, but negotiations with Oracle halted the project, disrupted construction, and even lost its general contractor.
Even so, OpenAI has not chosen to develop it on its own. Instead, it has cooperated with SoftBank, another Stargate partner, and announced that it will jointly develop the park with SoftBank subsidiary SoftBank Energy.
But according to two people involved in the negotiations, the two sides werecontrolConflict breaks out: OpenAI wants Texas campus to become itsThe first self-built data center, while SoftBank hopes to develop and hold the project.
From September to October of that year, the Stargate team went to Japan several times to negotiate with SoftBank founder Masayoshi Son.
One of the talks took place in a large conference room downstairs from Son's main office and lasted several hours. OpenAI employees frequently ran to the nearby 7‑11 to buy Coke and milk tea, and by the end of the meeting, the table was covered with empty bottles.
The two sides finally reached a compromise:
OpenAI signs long-term lease andTake control of facility design
SoftBank Energy is responsible for developing and owning
Inside OpenAI, this moment marks:The ambition to build your own data center is no longer a near-term goal.
The project started in October. Although this is not OpenAI’s first truly self-built data center, employees are still relieved that the deal, which has been in the works for nearly a year, has finally come to fruition.
OpenAI hopes the project will serve as a model for subsequent expansion of computing power: the agreement gives it control over the design without having to fully invest in or own the project. This control includes decisions such as cluster architecture, cooling systems, rack configurations and power infrastructure.
new person in charge
The evolution of the Stargate project and the realistic judgment of “not building self-operated infrastructure for the time being” are also accompanied by changes in the top management of OpenAI’s infrastructure department. In November, the company hired former Intel chief technology officer and AI chief Sachin Carty to lead the infrastructure team, reporting to president and co-founder Greg Brockman.
One of Kati's core tasks is to help OpenAI build data center-related intellectual property rights so that infrastructure protocols can accurately match the company's needs. He is also responsible for the overall computing power roadmap, including which chips OpenAI will use.
Several executives who previously reported to Brockman - including data center head Chris Malone and vice president of industrial computing Peter Hoschele - have now reported to Carty.
According to people familiar with the matter, Carty is also reorganizing other teams within the computing department.