OpenAI abandons first-party Stargate
- OpenAI has moved away from owning Stargate data centers itself, and now treats Stargate as a broad compute strategy built on leases and partner-built sites. - The physical buildout is still huge — a $16 billion Oracle campus in Michigan just secured financing, and Vantage-linked projects have chased roughly $38 billion more. - That shift matters because OpenAI wants capacity faster than ownership allows, but it also strains ties with SoftBank and other infrastructure partners.
OpenAI is changing what “Stargate” means. The original pitch sounded like a classic megaproject — OpenAI, SoftBank, Oracle, and others would build giant AI data centers together and own the future, basically. But the company has drifted toward something looser and more practical. Instead of insisting on first-party ownership, OpenAI now seems happy to let other companies finance and build the sites, then lease the compute it needs. (finance.yahoo.com) ### What was Stargate supposed to be? Back in January 2025, Stargate was framed as a $500 billion push to build 10 gigawatts of AI infrastructure in the US over four years. The idea was scale, permanence, and control — dedicated campuses built for OpenAI’s next generation of models. By July 2025, OpenAI was already describing Stargate less as one joint venture and mo(finance.yahoo.com)t first looked. (openai.com) ### So what changed? The cleanest version is this: OpenAI stopped treating joint ownership as the goal. Reporting this week says the company has, in practice, moved away from the original first-party structure and now prefers long-term leasing and bilateral deals. In plain English, OpenAI wants the chips and power without necessarily carrying the construction risk, financing complexity, and governance fights that come with owning the buildings. (finance.yahoo.com) ### Why would leasing win? Because AI infrastructure is weirdly time-sensitive. A data center that opens late can miss an entire hardware cycle. If Nvidia’s next platform is landing soon, waiting for a perfect ownership structure can cost more than paying someone else’s margin. Leasing also helps OpenAI stay flexible while it burns cash and keeps chasing more capacity than almost anyone else on earth. The tradeoff is obvious — less control, but faster access. (finance.yahoo.com) ### Is OpenAI still building anything big? Yes — this is not a retreat from scale. It is a retreat from owning the scale in the old way. A $16 billion Oracle-focused campus in Saline Township, Michigan, just secured financing from Related Digital, Blackstone-backed funds, and long-term debt anchored by PIMCO-managed money. The project is explicitly described as being delivered for Oracle and OpenAI, and major construction is already underway. (newsroom.bankofamerica.com) ### What about the even bigger financing? That is the other tell. Separate Oracle-linked campuses developed by Vantage Data Centers have been tied to an enormous debt package reported at about $38 billion, aimed at projects in Texas and Wisconsin. Whether or not every dollar closes on the first try, the important point is t(newsroom.bankofamerica.com) rather than a neat in-house build program. (octus.com) ### Did this upset partners? Seems like yes. The original structure required SoftBank, Oracle, OpenAI, and site developers to agree on control, economics, and timing. That is hard even in normal real estate. It is much harder when the tenant is an AI lab whose demand can jump by gigawatts in months. Some partners reportedly felt misled as projects in places like the UK and Norway shifted or stalled, with Microsoft stepping into at least one lease path. (finance.yahoo.com) ### Is OpenAI at least getting the capacity? Yes — and that is the strongest argument for the new approach. OpenAI hit a key 10-gigawatt contracted-capacity milestone years ahead of its original schedule, with 3 gigawatts added in the prior 90 days alone. So the strategy may be messy politically, but operationally it is working. (bloomberg.com). It got redefined. OpenAI still wants absurd amounts of compute — maybe more than ever — but it increasingly looks like a tenant and orchestrator, not the owner-builder at the center of a single grand project. (openai.com)