$500B Is Hardening Into Concrete and Power — What the Buildout Leaves Founders

Stargate’s $500B, four-year pledge to U.S. AI infrastructure from OpenAI, SoftBank, and Oracle is hardening from promise into steel. The Abilene, Texas flagship is already live, and five new sites push it toward 7 gigawatts. This compute buildout reshapes founders’ cost curves, GPU access, and the power and datacenter constraints under all of it.

What Happened

Stargate is a standalone company built to invest $500 billion over four years in U.S. AI infrastructure. It was first unveiled in January 2025 at a White House announcement. The initial equity funders are SoftBank, OpenAI, Oracle, and MGX. SoftBank and OpenAI each put in $19 billion and hold 40%; Oracle and MGX each contributed $7 billion. SoftBank carries financial responsibility, OpenAI handles operations, and Masayoshi Son chairs the venture. When announced, the number was so large it read as aspiration. Eighteen months later it’s concrete. The flagship campus in Abilene, Texas is already running on Oracle Cloud Infrastructure, and in September 2025 five new sites were added — two in Texas, plus New Mexico, Ohio, and an undisclosed Midwest location. That brings planned capacity to nearly 7 gigawatts, with more than $400 billion to be deployed over the next three years — putting Stargate ahead of schedule on its full 10-gigawatt, $500 billion commitment. The headline figure is turning into things you can touch: sites, power, GPUs.

What This Means for Founders

This buildout is double-edged. One edge is plainly opportunity, because it bends the cost curve per unit of compute downward. As gigawatts of capacity come online in waves, inference prices fall structurally, and workloads that were uneconomic last year pencil out this year. Even if you’re not training models yourself, cheaper and more abundant compute opens a wider market for the founder building products on top. The other edge is dependency. The number-one customer of this infrastructure is OpenAI. When the priority, pricing, and allocation of a $500 billion pipeline is organized around a single company, startups outside that ecosystem may not get the same GPUs on the same terms. That’s exactly why hyperscalers and well-capitalized labs are racing to lock their own power contracts and silicon. And the real bottleneck isn’t chips — it’s electricity and land. Ten gigawatts rivals the output of several large nuclear plants. If you’re building AI infrastructure, “how do I get GPUs” is fast becoming an easier question than “how do I secure power, cooling, and permits.” The constraint is migrating from the chip to the grid.

What You Can Do Now

First, map your product’s compute dependency explicitly. If inference scales linearly with revenue, a falling cost curve is a tailwind — but single-supplier dependency is a fatal risk. Second, diversify your compute supply. Don’t lash yourself to one cloud’s GPU pool; secure fallback options before pricing or availability shifts under you. Third, hunt for opportunity downstream of the buildout. More data centers grow adjacent markets — power, cooling, networking, and operations automation. You can build a real business in the layers that hold this infrastructure up without ever entering the model race. Fourth, treat power and land as the first-order constraint of any infrastructure play; if you’re standing up capacity, the energy contract and permitting timeline will govern your speed more than chip procurement. Fifth, don’t postpone efficiency just because compute is getting cheaper. Abundant capacity makes waste easy. Saving every token and every watt is what protects margin over the long run.