The AI boom was paid for with cash. Now it runs on credit
By Sayed A.
Chief Business Officer
Graystone Capital
Dek: Big Tech’s AI build-out is now being paid for with bonds, private loans and off-balance-sheet deals. The warning signs are in the debt markets, and they are easier to read than share prices.
For three years, the AI boom was a story about stock prices. If the bulls were wrong, shareholders would lose money and everyone else would watch.
That story ended this autumn. The AI boom is now a credit story, and credit stories travel. When debt goes bad, lenders pull back, refinancing gets harder, and the trouble reaches companies that never bought a single chip.
The turn came into focus on November 5, when the Nasdaq 100 lost roughly $2.5 trillion in market value in one session. The more telling moves came in the bond market. Within days, a European bank was shopping data center loans and an AI cloud company’s bonds were trading at distressed levels. Some banks that had lined up to fund AI projects told Reuters they were now reluctant to finance more.
From cash machines to borrowers
The spending is vast. Amazon expects about $220 billion of capital expenditure this year, and Microsoft about $170 billion for the calendar year. TwentyFour Asset Management estimates total AI capital spending above $750 billion in 2026. Morgan Stanley expects $1.66 trillion next year.
For most of this boom, the world’s most profitable companies paid from their own pockets. No longer. Goldman Sachs says the hyperscalers had negative free cash flow in the third quarter. The gap is being filled with debt.
Goldman counts about $296 billion of hyperscaler bond issuance in 2026. It expects $440 billion next year, plus $300 billion of project-level debt. The IMF notes that hyperscalers’ share of global investment-grade bond issuance has jumped to 10%, from less than 1% before 2024.
Borrowing to build is not wrong. Railroads and power grids were built on credit. But it changes who carries the risk if the returns come late.
Where the risk is hiding
The danger is not Microsoft defaulting. It sits in the weaker parts of the system, and in debt that is hard to see.
The leveraged middle. Oracle is building data centers for a reported $300 billion OpenAI contract. It has issued $118 billion of long-term debt and carries $39 billion in lease obligations, according to FactSet data cited by CNBC. Analysts expect its free cash flow to be negative $23.7 billion this year. S&P rates it BBB-, one notch above junk. Its default-insurance costs hit records in October, at levels one S&P Global analyst likened to junk credit.
The neoclouds. CoreWeave, which rents out Nvidia chips, is the system’s stress gauge. Its 2032 bonds yielded above 16% in early November, even as its business grew. Rival Nebius shelved a planned high-yield bond.
Debt off the books. Meta’s Hyperion campus in Louisiana is being built by a joint venture with funds managed by Blue Owl, at an estimated cost of about $27 billion. Meta owns 20% and does not consolidate the venture. But Meta will lease the campus and has guaranteed its residual value. Its latest quarterly filing puts the maximum it could owe under such guarantees, across its data center ventures, at $41.5 billion. That is up from $28.1 billion at the end of 2025.
Across the system, the IMF counts 14 deals since 2023 in which hyperscalers invested about $300 billion of equity in financially weaker AI firms. Those deals carried $650 billion of debt, in exchange for $1.4 trillion of future revenue commitments. The same few companies are investors, lenders, suppliers and customers to one another. That works while demand rises. It concentrates losses if demand falls short.
The lenders are blinking
The clearest signal comes from banks. ING is seeking buyers for 12 data center loans worth about $500 million, Bloomberg reported, including a piece tied to the Stargate project in New Mexico. At JPMorgan, loans backed by income-producing properties, a category that includes data centers, have grown to 12% of its commercial real estate book, from 8% three years ago.
A lawyer who works on these deals told Reuters that investors’ books “have become full.” A strategist at SMBC described “financiers looking to exit before the door shuts.” Bank of America says spreads on 10-year hyperscaler bonds have widened about 50 basis points this year. Comparable non-AI companies have seen about 20.
Politics now matters too. After this month’s midterms, Texas’s governor-elect called for a data center moratorium until consumer protections are in place. Goldman estimates moratoriums could affect 45% of planned US capacity. A data center without a permit or a grid connection cannot generate the cash its lenders were promised.
Regulators saw this coming. The IMF warned in October that an AI correction could be amplified by “rising leverage, concentrated exposures, and circular financing arrangements.” The Bank of England has flagged the “complexity and opacity” of AI debt structures.
This is not 2008, yet
The optimists have a case. The biggest borrowers are among the strongest companies on Earth, and investors still compete for their bonds. Morgan Stanley says spreads have tightened on most names since the latest deals, and sees good value in top-quality hyperscaler debt. The IMF’s stress test found the global banking system would stay resilient in a severe AI-led correction. Demand is real, too: CoreWeave beat expectations and raised its guidance this week.
But the better comparison is the dot-com bust, and it is not reassuring. The telecom builders of 1999 also had real customers, and their fiber was eventually used. The problem was that the debt came due before the demand arrived, and lenders broke before the technology did. Today, chips wear out in a few years, while much of this debt runs 15 to 30 years. That mismatch is the core of the risk.
What needs to happen now
None of this means a crash is coming. It means the important questions have moved to the bond desk, and policymakers should follow.
- Show the full picture. Companies should report data center lease guarantees and joint-venture obligations as one clear figure, not scattered across footnotes.
- Follow the risk. Private credit funds are buying the loans banks sell. Regulators need to know where that risk ends up, and how much returns to banks through lending to those funds.
- Price the politics. Lenders and rating agencies should stress-test for permitting delays, power shortages and local moratoriums.
The next tests are close. Nvidia reports earnings on November 17, and ING’s loan sale closes for bids two days later. Morgan Stanley expects far heavier AI borrowing in 2027. If investors keep lending on today’s terms, the boom rolls on. If they demand more, the weakest borrowers will feel it first.
For years, the question was whether AI would change the world. The question now is more prosaic, and more urgent: who is holding the debt if the payoff takes longer than promised?








