When President Trump stood alongside Sam Altman, Larry Ellison, and Masayoshi Son in January 2025 to announce Stargate, it was billed as America’s answer to Chinese AI infrastructure - a $500 billion moonshot that would secure U.S. dominance in artificial intelligence. Fourteen months later, the flagship expansion has collapsed, Oracle is hemorrhaging cash and cutting tens of thousands of jobs, and Nvidia is playing kingmaker to decide who gets the abandoned capacity.
The gap between Stargate’s promise and its reality tells us something important about the economics of AI infrastructure - and whether the current buildout is sustainable.
What Fell Apart
Oracle and OpenAI have abandoned plans to expand their data center in Abilene, Texas from 1.2 gigawatts to 2 gigawatts of capacity. According to Bloomberg, two factors killed the deal: financing disputes and “OpenAI’s frequently changing demand forecasting and shifting view of Stargate.”
That second part is striking. OpenAI, the company at the center of the AI boom, apparently cannot predict how much compute it actually needs. When you’re asking partners to commit billions to infrastructure that takes years to build, that’s a fundamental problem.
A winter outage made things worse. Liquid cooling equipment failures left part of the Abilene site offline for several days, reportedly damaging the relationship between OpenAI and Crusoe, the site’s developer.
The existing 1.2-gigawatt project remains on track, valued at $300 billion over its lifetime. But OpenAI had committed to 5 gigawatts of Stargate capacity. It now needs to find alternative sources for what the expansion would have provided.
Oracle’s Impossible Position
The Stargate collapse is one symptom of a larger crisis at Oracle. The company is trapped in what analysts call an “unequal relationship” with OpenAI - providing infrastructure on terms that strain its finances while OpenAI retains the option to shift demand elsewhere.
The numbers are brutal. Oracle faces negative free cash flow of $13.1 billion and total debt of $108 billion. To fund AI infrastructure expansion, it recently announced plans to raise $45-50 billion in 2026 through debt and equity. Its stock has fallen 54% from September 2025 highs.
The human cost is equally stark. Oracle is evaluating layoffs of 20,000 to 30,000 employees - potentially this month - to generate $8-10 billion in cash flow for AI infrastructure. The company committed to a $156 billion OpenAI deal requiring 3 million GPUs over five years, but the revenue from that deal won’t materialize fast enough to cover the upfront costs.
Oracle’s calculus was that AI infrastructure would eventually justify the massive capital outlays. But “eventually” requires surviving until the revenue arrives. The Abilene expansion collapse suggests the company is starting to triage its commitments.
Nvidia Plays Kingmaker
Enter Nvidia. After Oracle and OpenAI walked away from the expansion, developer Crusoe had 0.8 gigawatts of planned capacity with no tenant. Rather than let that capacity potentially go to a competitor’s chips, Nvidia put down a $150 million deposit before approaching Meta about taking over the site.
Meta is now in talks to lease the abandoned capacity. A deal hasn’t been signed, but the arrangement makes sense for all parties involved - except perhaps Oracle. Meta has committed to up to $135 billion in 2026 capital expenditures on AI infrastructure. The Abilene site is already being developed. And Nvidia gets to ensure its chips power the facility rather than AMD’s.
This dynamic illustrates how the AI infrastructure landscape is fragmenting. When Stargate was announced, it represented a coordinated partnership between leading AI companies. Now we’re seeing opportunistic dealmaking, with Nvidia using its position as the indispensable chipmaker to influence which companies get which capacity.
The Structural Problems Were Always There
The Stargate joint venture has been dysfunctional from the start. More than a year after the announcement, the partnership hasn’t hired any staff and isn’t actively developing any data centers as a unified entity. The three partners - OpenAI, Oracle, and SoftBank - have been arguing over responsibilities and structure since last summer.
The financing model was always shaky. SoftBank, the primary financial backer, could only secure 10% in equity funding. The remaining $448 billion was supposed to come from debt financing that hasn’t materialized. As SoftBank’s CFO admitted last August, the project was “taking longer than anticipated.”
OpenAI initially tried to build its own data centers but couldn’t get financing. Lenders weren’t willing to back billion-dollar projects from a company projecting $14 billion in losses for 2026. The current arrangement - where Oracle borrows money and leases capacity back to OpenAI - offloads financial risk onto a partner that’s now drowning in it.
What This Means for AI Infrastructure
The Stargate collapse matters beyond the specific companies involved. It signals that the current AI infrastructure buildout may be hitting financial limits.
The industry collectively announced hundreds of billions in data center spending based on assumptions about AI demand that may not hold. When OpenAI itself can’t reliably forecast its compute needs, what does that say about everyone else’s projections?
Several scenarios become more likely:
Consolidation accelerates. Companies with the strongest balance sheets - Microsoft, Google, Amazon - gain advantage as weaker players struggle to finance buildouts. The AI infrastructure market may consolidate around a smaller number of hyperscalers.
Stranded assets emerge. Data centers take years to build. If demand forecasts prove wrong or financing dries up, we could see partially-built facilities or completed centers running below capacity. Ratepayers and investors absorb the losses.
China gains ground. Stargate was explicitly framed as competition with China. If U.S. infrastructure buildout slows due to financing problems while Chinese investment continues, the geopolitical calculus shifts.
The Bottom Line
Stargate hasn’t failed entirely - the existing 4.5 gigawatt agreement between Oracle and OpenAI continues, and other sites remain in development. But the flagship expansion’s collapse, combined with Oracle’s financial crisis and the underlying partnership dysfunction, suggests the $500 billion vision was always more aspiration than plan.
The AI infrastructure race continues, but the runners are starting to tire. Oracle is cutting workers to fund chips. Nvidia is depositing millions to control who gets capacity. Meta is picking up pieces that OpenAI couldn’t finance. And somewhere in Abilene, a data center that was supposed to symbolize American AI dominance is being divided among whoever can still afford to build.