OpenAI, Oracle, and SoftBank add five Stargate sites, pushing planned capacity toward 7 gigawatts

What the September 23 announcement actually commits to

OpenAI, Oracle, and SoftBank announced five new U.S. data center sites under Stargate. Combined with the flagship Abilene, Texas campus and ongoing CoreWeave projects, OpenAI states this brings Stargate to nearly 7 gigawatts of planned capacity and over $400 billion in investment across the next three years.

The stated goal is to secure the full $500 billion, 10-gigawatt commitment first announced in January by the end of 2025 — described as ahead of schedule. A July agreement between OpenAI and Oracle to develop up to 4.5 gigawatts of additional capacity is cited as a partnership exceeding $300 billion over five years.

The sites break into two groups. Three Oracle-linked locations — Shackelford County, Texas; Doña Ana County, New Mexico; and a Midwest site later confirmed as Wisconsin (developed by Oracle with Vantage) — plus a potential 600-megawatt expansion near Abilene can deliver over 5.5 gigawatts. Two SoftBank-partnered sites, in Lordstown, Ohio and Milam County, Texas, can scale to 1.5 gigawatts over 18 months.

The GB200 line is the one that touches production apps

Most of the announcement is about power, dollars, and jobs — over 25,000 onsite jobs are projected across the sites. But one sentence is directly relevant to anyone running inference in production: Oracle began delivering the first NVIDIA GB200 racks in Abilene in June, and OpenAI says it has already started early training and inference workloads on that capacity.

For frontend and application teams, training capacity is invisible; it shows up months later as a better model. Inference capacity is not. It is the constraint behind rate limits, response latency, and whether a feature stays available under a traffic spike. The announcement explicitly ties new sites to inference, not only training.

That distinction matters because the physical footprint being built here is what backs the API a product depends on. When a Stargate site turns on inference capacity, the practical downstream effect is more headroom on the endpoints application code already calls.

'Fast-build' and cost efficiency as an infrastructure claim

The SoftBank-partnered sites are described in operational terms: Lordstown uses an 'advanced data center design' on track to be operational next year, and Milam County is a 'fast-build' site with powered infrastructure supplied by SB Energy. OpenAI frames both as steps toward 'faster deployment, greater scalability, and improved cost efficiency—making high-performance compute more widely accessible.'

To meet this enormous demand, we continue to expand OCI's footprint at an unrivaled pace to deliver the most performant and cost-effective AI training and inferencing.— Clay Magouyrk, CEO of OracleMontana Labs

'Cost-effective inferencing' is the phrase to hold onto. The announcement offers no per-token pricing, so nothing here changes a budget today. But the stated intent — cheaper inference at larger scale — is the mechanism by which capacity announcements eventually reach product economics.

What frontend teams should watch as sites come online

None of this is deliverable to a product on announcement day. Abilene is running on OCI; the other five sites range from operational next year to under evaluation. The selection process itself — over 300 proposals from more than 30 states, with more sites to come — signals this is a multi-year buildout, not a switch being flipped.

The concrete implication for teams building on OpenAI: inference capacity is being expanded and geographically distributed, and the first GB200 inference workloads are already live. The useful posture is to treat rate-limit headroom and latency as things that should loosen over the buildout window, while designing for the reality that today's constraints still apply until specific sites report inference online.