The $500 Billion Gamble: Inside the Debt Deal Remaking America’s AI Infrastructure
An $18 billion loan for a massive data center in the New Mexico desert reveals how OpenAI is quietly building a new kind of economy—and why the real battle isn’t about AI models anymore.
In the sunbaked plains of Doña Ana County, New Mexico, twenty of the world’s biggest banks—led by Sumitomo Mitsui, BNP Paribas, Goldman Sachs, and Mitsubishi UFJ—are making a daring move. As reported by Bloomberg, they’re backing an $18 billion project finance loan to build a data center campus so huge it’ll draw as much electricity as the entire city of Seattle.
This deal, expected to close in late November, is part of a wave of enormous financings linked to Stargate, a $500 billion joint venture between OpenAI, SoftBank, and Oracle. The goal? To build 10 gigawatts of AI computing capacity across America. Oracle will anchor the project as its main tenant, while Blue Owl Capital—through its STACK Infrastructure platform—will handle both equity and construction.
But make no mistake: this isn’t just about real estate. It’s a signal that the AI race has hit a new phase. The choke point isn’t code or data anymore—it’s power and steel.
The Compute Wars Go Vertical
For years, OpenAI leaned heavily on Microsoft’s Azure cloud to train its models. That partnership brought huge computing bills and growing strategic tension. Now, the relationship is evolving—not with a breakup, but with a bold expansion.
In October, OpenAI inked a $250 billion Azure deal while Microsoft agreed to give up its right of first refusal on OpenAI’s compute purchases. Weeks later, OpenAI signed another massive agreement—a $38 billion, seven-year contract with Amazon’s AWS—unlocking access to hundreds of thousands of Nvidia GPUs. The message couldn’t be clearer: OpenAI is diversifying fast, and it plans to sell computing power, not just use it.
Executives have hinted at the company’s next move: building an “AI cloud” business that rents out GPUs-as-a-service. That would put OpenAI head-to-head with Amazon Web Services, Google Cloud, and even its own investor, Microsoft Azure. The entire strategy revolves around Stargate—constructing enough compute infrastructure to handle both its own training needs and lease specialized capacity to other companies and governments.
The financial logic makes sense. OpenAI earns around $13 billion in annual recurring revenue but burns roughly $2.5 billion every quarter. It needs new cash flows. Selling excess data center capacity, especially optimized for AI-heavy workloads that traditional cloud providers struggle with, could bring in tens of billions a year by the end of the decade.
There’s also a global chess match at play. OpenAI has urged the incoming Trump administration to back public-private partnerships that could add 100 gigawatts of new power generation each year. The argument: Stargate isn’t just business—it’s a matter of national competitiveness. In 2024, China added 429 gigawatts of new power capacity. The U.S.? Only 51.
The Smart Money’s Calculus
For investors, the New Mexico project crystallizes three clear investment theses, each with its own risks and rewards.
First, AI infrastructure has become a new asset class. Project finance structures—non-recourse loans secured by long-term leases—are drawing in banks and institutional investors who typically fund utilities or toll roads. If Oracle’s tenant commitments are strong enough, with long contracts, utilization minimums, and rent escalators, the syndication should sell smoothly. A successful retail phase to pension funds and insurers would confirm investor appetite for similar deals in Stargate’s upcoming projects in Texas, Ohio, Wisconsin, and Michigan.
The biggest challenge lies in power. High-density AI data centers, costing $12–20 million per megawatt, require guaranteed electricity, cooling water rights, and transmission access that can take years to arrange. A year’s delay can slash returns by hundreds of basis points. Savvy investors will demand proof—interconnection queue positions, on-site power generation plans, and long-term power purchase agreements—before signing off.
Second, Oracle is the hidden winner. The market still sees Oracle as a database and cloud software firm, but through Stargate, it’s morphing into a physical compute landlord with decades of visibility into OpenAI’s needs. If Oracle manages these massive facilities efficiently and captures profit from both leasing and operations, its upside could be far bigger than investors currently expect. The downside? Tenant concentration and the risk of managing something on an unprecedented scale.
Third, OpenAI’s pivot to selling AI cloud capacity won’t replace the hyperscalers—it’ll squeeze their margins. The Azure and AWS mega-deals prove demand is real. But by offering competing GPU services, OpenAI forces everyone else to get sharper on pricing. That gives customers more bargaining power, pushing AWS and Azure to sweeten their bundles. Nvidia, however, wins either way—every outcome means more GPUs flying off the shelves. Still, inventory control and production timing will be critical as inevitable construction delays kick in.
The most compelling trade may not even be in tech stocks. It’s in energy and infrastructure—utilities in regions friendly to data centers, high-voltage equipment makers, and companies controlling water rights near major fiber networks. In this new race, the real bottleneck isn’t chips—it’s the grid.
Structured credit investors are watching the loan’s syndication pricing closely. If it clears between 250–400 basis points above benchmarks with strong terms, secondary participation could look very attractive. But if demand falters or pricing widens, it’ll be a warning shot for the entire AI financing pipeline.
What Comes Next
By late November, when syndication commitments close, the market will get its first true read on investor appetite for AI infrastructure debt. If this deal sails through, expect a flood of similar financings to follow. If it stumbles, that means investors are starting to worry about overcapacity or thin margins.
Beyond New Mexico, the Stargate roadmap aims for seven gigawatts of new capacity across multiple sites in just three years. That’s ambitious, especially given the power grid bottlenecks that have frustrated big tech builders for over a year. OpenAI is betting that by vertically integrating—owning the models, the compute, and now the physical infrastructure—it can secure the independence worth hundreds of billions.
This isn’t a gamble on whether demand for AI exists—it clearly does. The real question is whether the power and the people can arrive fast enough to keep up.
NOT INVESTMENT ADVICE