Finance
The Small Slice of Data-Center Money That’s Making Everyone Nervous
There’s a pivotal scene in *The Founder*, the movie about Ray Kroc’s success in building the McDonald’s hamburger empire.
Kroc is meeting in a small, cramped office, talking about his low profit margins and lack of control over how franchisees operate. After a brief exchange, he’s told, “You are not in the burger business. You are in the real estate business.”
Instead of earning slim margins on commodity hamburgers, Kroc could collect rent from franchisees while gaining control over nearly every aspect of their operations. The real opportunity wasn’t the burger—it was the land.
And it worked. Kroc turned McDonald’s into a global icon not just by selling Big Macs, but also by buying dirt.
Keep that in mind as you watch the American data center industry expand.
A small but growing—and increasingly meaningful—share of its financing comes from junk bonds: billions of dollars in high-yield, high-risk debt purchased by investors who understand that the risk is elevated but are attracted by the potential for higher returns.
Most of the money pouring into America’s AI data centers is fairly conventional. Roughly $600 billion will be spent on the buildout this year. About half will come from cash already sitting on the balance sheets of Microsoft, Meta, Google, and Amazon. Another quarter will come from investment-grade corporate bonds, the sort pension funds buy routinely.
The part attracting attention is much smaller: about $18 billion, or roughly 3% of this year’s data-center spending, raised through high-yield bonds. In less polite circles, they are called junk bonds. They offer higher returns because the risk of not getting repaid is real.
Three percent is not much. But it is growing quickly, and this is usually where financial strain appears first.
The deals follow a familiar structure. A smaller operator such as Applied Digital, TeraWulf, or the newly formed Meridian Arc borrows several billion dollars to build a data center. Because the operator cannot borrow cheaply on its own, it first signs a lease of ten to fifteen years with a creditworthy tenant such as Amazon, Google, Oracle, or Nvidia.
The lease payments then back the bonds. In practical terms, lenders are betting on the tenant’s promise to pay rent, not on the financial strength of the company building the facility.
That structure has produced some enormous deals. A Google-backed project in Sullivan County, Indiana, raised $5.7 billion, the largest financing of its kind on record. An Nvidia-linked project received $14 billion in orders for a $3.8 billion offering. Named data-center junk-bond deals had already exceeded $15 billion in gross issuance through the first half of 2026. (That figure and the $18 billion above measure different things: the $18 billion is an estimate of full-year net new borrowing, while the named deals are counted as gross issuance.)
Demand has been ferocious. The compensation for taking the risk has not.
In December 2025, data-center bonds paid roughly 2.2 percentage points more than the broader high-yield market. That premium compensated investors for construction delays, unproven cash flow, and dependence on a single tenant. By the middle of 2026, it had nearly disappeared.
Penn Mutual Asset Management, which tracks these spreads, concluded that investors were no longer receiving a meaningful premium for owning the debt. The spread began widening again in late July, suggesting some of that confidence may now be fading.
Aswath Damodaran, the NYU finance professor, put the concern more bluntly: “Who are these lunatics who are lending money to the data centers?”
He compared the boom to shale-oil producers borrowing heavily when oil traded at $120 a barrel. The financing worked only if favorable conditions continued. They did not.
Damodaran’s larger point is that the economics are especially unforgiving for bondholders. If a project fails, they absorb the losses. If AI becomes the gold mine its backers expect, their return is still limited to the interest on the loan. They take much of the downside and receive none of the upside.
“You can’t make interest payments with potential and promise,” he said.
Timing makes the problem worse. Take that same $5.7 billion Sullivan County project: it is expected to be operational in about ten months, yet developers must begin servicing the debt almost immediately, often before a data center is complete or generating revenue. If it misses a payment, lenders may have the right to demand repayment of the entire balance.
None of this amounts to a crisis. AI-linked bonds still make up only about 2% to 3% of the investment-grade and high-yield bond benchmarks, and the tenants signing the leases are among the strongest companies in the world.
Still, there is a historical warning. Telecom debt accounted for about 1% of the bond market in the mid-1990s. It exceeded 20% just before the dot-com bust. New industries can scale remarkably quickly, and debt can accumulate long before the revenue expected to support it arrives.
There is an old warning about junk bonds, often associated with Warren Buffett: too often, these securities live up to their name.
For now, only a small part of the AI buildout is testing that proposition. But it is the part worth watching.
Kroc is meeting in a small, cramped office, talking about his low profit margins and lack of control over how franchisees operate. After a brief exchange, he’s told, “You are not in the burger business. You are in the real estate business.”
Instead of earning slim margins on commodity hamburgers, Kroc could collect rent from franchisees while gaining control over nearly every aspect of their operations. The real opportunity wasn’t the burger—it was the land.
And it worked. Kroc turned McDonald’s into a global icon not just by selling Big Macs, but also by buying dirt.
Keep that in mind as you watch the American data center industry expand.
A small but growing—and increasingly meaningful—share of its financing comes from junk bonds: billions of dollars in high-yield, high-risk debt purchased by investors who understand that the risk is elevated but are attracted by the potential for higher returns.
Most of the money pouring into America’s AI data centers is fairly conventional. Roughly $600 billion will be spent on the buildout this year. About half will come from cash already sitting on the balance sheets of Microsoft, Meta, Google, and Amazon. Another quarter will come from investment-grade corporate bonds, the sort pension funds buy routinely.
The part attracting attention is much smaller: about $18 billion, or roughly 3% of this year’s data-center spending, raised through high-yield bonds. In less polite circles, they are called junk bonds. They offer higher returns because the risk of not getting repaid is real.
Three percent is not much. But it is growing quickly, and this is usually where financial strain appears first.
The deals follow a familiar structure. A smaller operator such as Applied Digital, TeraWulf, or the newly formed Meridian Arc borrows several billion dollars to build a data center. Because the operator cannot borrow cheaply on its own, it first signs a lease of ten to fifteen years with a creditworthy tenant such as Amazon, Google, Oracle, or Nvidia.
The lease payments then back the bonds. In practical terms, lenders are betting on the tenant’s promise to pay rent, not on the financial strength of the company building the facility.
That structure has produced some enormous deals. A Google-backed project in Sullivan County, Indiana, raised $5.7 billion, the largest financing of its kind on record. An Nvidia-linked project received $14 billion in orders for a $3.8 billion offering. Named data-center junk-bond deals had already exceeded $15 billion in gross issuance through the first half of 2026. (That figure and the $18 billion above measure different things: the $18 billion is an estimate of full-year net new borrowing, while the named deals are counted as gross issuance.)
Demand has been ferocious. The compensation for taking the risk has not.
In December 2025, data-center bonds paid roughly 2.2 percentage points more than the broader high-yield market. That premium compensated investors for construction delays, unproven cash flow, and dependence on a single tenant. By the middle of 2026, it had nearly disappeared.
Penn Mutual Asset Management, which tracks these spreads, concluded that investors were no longer receiving a meaningful premium for owning the debt. The spread began widening again in late July, suggesting some of that confidence may now be fading.
Aswath Damodaran, the NYU finance professor, put the concern more bluntly: “Who are these lunatics who are lending money to the data centers?”
He compared the boom to shale-oil producers borrowing heavily when oil traded at $120 a barrel. The financing worked only if favorable conditions continued. They did not.
Damodaran’s larger point is that the economics are especially unforgiving for bondholders. If a project fails, they absorb the losses. If AI becomes the gold mine its backers expect, their return is still limited to the interest on the loan. They take much of the downside and receive none of the upside.
“You can’t make interest payments with potential and promise,” he said.
Timing makes the problem worse. Take that same $5.7 billion Sullivan County project: it is expected to be operational in about ten months, yet developers must begin servicing the debt almost immediately, often before a data center is complete or generating revenue. If it misses a payment, lenders may have the right to demand repayment of the entire balance.
None of this amounts to a crisis. AI-linked bonds still make up only about 2% to 3% of the investment-grade and high-yield bond benchmarks, and the tenants signing the leases are among the strongest companies in the world.
Still, there is a historical warning. Telecom debt accounted for about 1% of the bond market in the mid-1990s. It exceeded 20% just before the dot-com bust. New industries can scale remarkably quickly, and debt can accumulate long before the revenue expected to support it arrives.
There is an old warning about junk bonds, often associated with Warren Buffett: too often, these securities live up to their name.
For now, only a small part of the AI buildout is testing that proposition. But it is the part worth watching.
Ad slot · in-article
Leave a Reply