OpenAI, Oracle and SoftBank Add Five Sites to Stargate Buildout

OpenAI, Oracle and SoftBank have announced five new datacenter sites for Stargate, the joint infrastructure venture building computing capacity for next-generation AI in the United States.

The expansion accelerates a program the partners size at $500 billion and 10 gigawatts of capacity — a scale of investment that puts AI compute buildout in the same financial league as national infrastructure programs rather than ordinary corporate capex. The companies say the buildout will create tens of thousands of jobs.

Stargate unites three distinct balance sheets. OpenAI supplies the models and the demand signal. Oracle brings enterprise datacenter construction and operations. SoftBank provides capital. Five additional sites means the venture is now spreading its footprint across multiple US locations, a design choice that spreads power procurement, grid negotiations and construction timelines across regions rather than concentrating them in one.

The stakes are straightforward. Training and serving frontier AI models requires gigawatt-scale electricity, land and cooling that no single facility can deliver. If US-built capacity falls behind demand, model development slows or moves elsewhere. A 10-gigawatt commitment — roughly the generating capacity of several large nuclear plants — signals that the partners expect AI compute demand to stay steep for years, not quarters.

The jobs figure matters for the policy context. Tens of thousands of construction and operations roles give the project a political constituency at a moment when Washington is weighing how aggressively to support domestic AI infrastructure through energy and permitting policy.

The three companies have not yet disclosed the locations of the five new sites, their individual capacities or their construction timelines in this announcement.

With the expanded footprint, Stargate becomes the benchmark against which rival US and international AI infrastructure programs will be measured — and the first test of whether the industry can actually deliver gigawatt-scale compute on schedule.