OpenAI and Oracle ink 4.5 GW mega-deal for Stargate

While most AI startups are still counting their compute in thousands of GPUs, OpenAI is now measuring its infrastructure in gigawatts and power grids.

OpenAI just announced a massive deal with Oracle to build an additional 4.5 gigawatts of data center capacity across the United States. Paired with its existing Stargate I site in Abilene, Texas, the company now has over 5 GW of AI data center capacity in active development—enough juice to power more than 2 million chips.

Back in January, OpenAI pledged at the White House to pour $500 billion into 10 GW of U.S. AI infrastructure over four years. Thanks to rapid progress with Oracle and SoftBank, OpenAI now expects to comfortably beat that target.

GB200 racks are already humming

Stargate is not just a long-term concept on a slide deck. Parts of the Abilene site are already live and running.

Oracle delivered the first batch of Nvidia GB200 racks last month, and OpenAI confirmed it has already begun running early training and inference workloads to test next-generation frontier research. The Abilene project alone has pulled in technicians and specialized electricians from over 20 states.

OpenAI estimates that building and operating this extra 4.5 GW capacity will generate more than 100,000 U.S. jobs, from construction crews to long-term facility engineers. Meanwhile, SoftBank is actively scouting sites and rethinking data center blueprints to house these power-hungry clusters, while Microsoft will continue to provide cloud services alongside Oracle and CoreWeave.

Why it matters

We think this partnership proves that frontier AI is no longer just a software race—it is a heavy-industry logistics game where power generation and physical rack delivery dictate the timeline.

For developers and everyday users, running over 2 million advanced chips will eventually translate into dramatically lower inference costs, near-instant responses, and the raw muscle needed to train true next-generation frontier models. But do not expect overnight miracles; building out 4.5 GW takes years of construction and grid interconnection, so current capacity bottlenecks will linger for a while.

Strategically, OpenAI is playing a brilliant hand by diversifying its dependencies. Instead of relying solely on Microsoft Azure, bringing Oracle, SoftBank, and CoreWeave into the fold gives OpenAI serious leverage and ensures it is not caught waiting on a single cloud partner.

The real risk, however, is execution on the physical grid. Finding 4.5 gigawatts of clean, reliable electricity in the U.S. is notoriously difficult. If local power infrastructure cannot keep up, or if the commercial returns on frontier models fail to justify half a trillion dollars in capital outlay, OpenAI could find itself sitting on an extraordinarily expensive power bill.