Stargate or not, the most important thing is to secure compute capacity.
This pivot is evident in OpenAI’s strategy. The joint venture, rolled out with much fanfare in early 2025 with Washington’s blessing, is not being overturned. But addressing compute needs is not as consistently tied to building datacenters.
Setbacks and U-turns for Stargate’s Texas showcase
Stargate is intended to marshal $500 billion over four years to build an AI infrastructure for OpenAI. Oracle, SoftBank and the Emirati sovereign wealth fund MGX are among its partners. Arm, Microsoft and NVIDIA are technology partners.
The centerpiece of the initiative is a campus of datacenters in Abilene, Texas. Site work had begun as early as July 2024 under Crusoe, the infrastructure outfit. The plan was to ultimately reach about 1.2 gigawatts of compute power across roughly ten buildings.
As of the latest reports, under 300 MW were online. Most of it had been up since September 2025 (Oracle had delivered the first GB200 racks in June). There was a plan to extend capacity to 2.1 GW. But in early 2026, OpenAI and Oracle chose not to proceed. Behind the scenes, funding challenges, delays in connecting to the electrical grid—and more broadly, a cooling of enthusiasm for the Stargate idol itself—factored into the decision, especially with an IPO looming and the push to rationalize spending. Microsoft ultimately “took back” that 900 MW tranche.
Scale-back abroad
In May 2025, Stargate started to unfold beyond U.S. borders. OpenAI formalized a project in the United Arab Emirates with the G42 holding company. The plan called for deploying 1 GW in Abu Dhabi, of which 200 MW would be operational in 2026.
In July 2025, a second international initiative was announced. In Norway, with the British company Nscale, the target was 500 MW, starting with 100,000 GPUs and potentially rising to 250,000. OpenAI ultimately did not secure an agreement to operate all of the capacity. Microsoft stepped in to take on about 30,000 GPUs.
The Stargate UK project, announced in September 2025 and developed with Nscale, was paused in the spring of 2026. Officially, the reasons cited were high energy prices and what regulators deemed overly restrictive rules. OpenAI had discussed a purchase agreement for up to 31,000 GPUs.
From Michigan to New Mexico, Stargate campuses… in full tilt
Texas also became the seat of two additional Stargate projects. One, announced in August 2025 and developed by Vantage Data Centers, was to deliver 1.4 GW of IT power across 10 datacenters, with the first building slated for completion in the second half of 2026. SoftBank would supply the chips. The other project, located in Burlington, saw its construction handled by SB Energy. Phase one was scheduled to wrap up by October 2026, with a goal of 1.2 GW.
In September 2025 another site, named Project Jupiter, was unveiled in Santa Teresa, New Mexico. It aimed for 2 GW, with Blue Owl as the main investor, committing $50 billion over five years. STACK Infrastructure was in charge of building. The project faced lawsuits over the opacity of the permitting process.
The two most recent Stargate campus announcements came in October 2025. One project, named The Barn, is in Benton, Michigan. Related Digital is handling the construction. The first buildings were expected to be operational early 2027, targeting 1.4 GW with Oracle supplying the chips. The other project, named Lighthouse, is entrusted to Vantage Data Centers, with Oracle again providing the chips. Construction was set to begin in 2028, aiming at 900 MW across four datacenters.
Gigawatts outside the Stargate perimeter
SB Energy is currently building in Ohio a datacenter outside Stargate’s orbit. OpenAI would target 10 GW of compute capacity based on NVIDIA GPUs, with the first tranche of 800 MW slated to come online in 2028.
Also in Ohio, OpenAI is seeking to lease resources on the Mount Pleasant Fairwater campus. Microsoft has just completed the build with Vantage Data Centers, two years after the project was announced.
Instinct, Trainium, Jalapeño… OpenAI sees beyond NVIDIA
In September 2025, OpenAI committed to deploying at least 10 GW more NVIDIA GPUs. The initial step, in the second half of 2026, would be 1 GW of Vera Rubin. NVIDIA signaled its intent to invest up to $100 billion in OpenAI.
In October followed an agreement with AMD to deploy up to 6 GW of Instinct GPUs. The first phase would be 1 GW in MI450 in the second half of 2026. In return, AMD offered OpenAI the possibility of holding up to 10% of its equity under favorable terms.
In November, OpenAI signed with Amazon a $38 billion deal over seven years, focused mainly on using GB200 and GB300 via UltraServers EC2. Early 2026 they added another $100 billion, stretched over eight years. This included a commitment to operate around 2 GW of Trainium chips through the AWS infrastructure.
At the same time, OpenAI has been working on designing its own chips with Broadcom. The first product from this collaboration, the Jalapeño ASIC, is dedicated to inference. It is not yet deployed at scale, but it is already handling some workloads, notably with GPT-5.3-Codex-Spark.
OpenAI and Broadcom formalized their partnership in October 2025. The goal is to deploy, by 2029, 10 GW of accelerators. Greg Brockman noted that the collaboration between the two firms had been ongoing for about 18 months.
* Several top executives from the Stargate initiative left for Meta earlier this year.
