Private capital is financing an AI infrastructure boom that cannot yet pay its own interest bills.

CoreWeave rents out Nvidia GPU clusters to AI companies. It funded the buildout by borrowing against the chips themselves as collateral. Today it carries roughly $50 billion in debt, pays around $640 million in interest every quarter, and reports a net loss despite growing revenue. This week its unsecured bonds fell below 80 cents on the dollar. Its debt insurance costs hit 1,200 basis points. A company with a $104 billion contracted backlog is in distressed bond territory. Know this: contracted backlog is not cash.

CoreWeave's bonds due 2032 were priced at 9.625 percent in June. They now yield roughly 15 to 16 percent. That repricing travels. Every private credit fund, pension pot, or insurance company that bought this paper has a mark-to-market loss today. UK insurers and asset managers are among the buyers of discounted AI infrastructure debt being offloaded by US banks. If your pension or insurance provider has private credit exposure, this is the story underneath the quarterly statement you have not been sent yet.

Law of the Addict. CoreWeave could not slow the borrowing because the model only works at scale. Every GPU cluster requires another loan to stay competitive. The next chip generation is already coming. The collateral depreciates faster than the debt matures. Investors kept lending because the yield was high and the backlog looked enormous. Now the bonds are distressed and nobody wants to admit they mistook contracted revenue for solvency. What is being done to your head: the backlog number was real but irrelevant. Revenue commitments do not pay interest.

JPMorgan and MUFG led a $38 billion construction loan for data centres in Texas and Wisconsin leased to Oracle. Six months later they are still trying to move it. Some lenders are now offering slices to non-bank buyers at a discount just to get it off their books. Oracle's five-year credit default swap, the cost of insuring against Oracle defaulting, hit 151 basis points in early October, the highest since 2009. S&P cut Oracle to BBB-minus in July, one notch above junk. Bond order books for AI hyperscalers have thinned from five times oversubscribed in February to under two times by July.

A credit default swap, or CDS, is insurance against a borrower failing to pay. Oracle's CDS at 151 basis points means insuring $10 million of Oracle debt now costs $151,000 a year. That is the market telling you risk has risen sharply. The $38 billion loan is £28 billion. For context, the entire annual capital budget of Network Rail is around £7 billion. US and UK institutional lenders, including insurance funds and asset managers, are being offered pieces of this at a discount. The discount is the cost of urgency.

Law of the Trap. Banks raced to underwrite data centre loans when the yield was good and the borrower looked solid. Internal exposure limits were not a problem when every deal looked like a winner. Now those limits are real, the debt is sticky, and the exit costs money. The trap is that the same competitive pressure that drove the underwriting is now driving the discounted sale. Every bank that holds too long loses relative to the one that sold first at a small discount. What is being done to your head: the original confidence in these loans was manufactured by the same institutions now quietly reversing.

The IMF October 2026 Global Financial Stability Report flagged AI infrastructure debt as a top systemic risk. The concern is circular: hyperscalers are providing equity to weaker AI firms, lending them money, and also collecting future revenue commitments from those same firms. Fourteen disclosed deals involve roughly $300 billion in equity investments, $650 billion in debt, and $1.4 trillion in future revenue commitments, all looping back through the same set of companies. The IMF called it a potential triple collapse: equity falls, debt cannot be serviced, and revenue commitments evaporate together.

The IMF flagged a specific number: hyperscalers now account for 10 percent of all global investment-grade bond issuance, up from less than 1 percent in 2024. Your pension fund holds investment-grade bonds. Your insurer holds investment-grade bonds. The New York City Teachers pension alone has poured $3.5 billion into 50 digital infrastructure funds. US public pension funds, which pay teachers, firefighters, and state workers, have billions allocated to this sector. If the IMF's scenario materialises, the loss does not stay on Wall Street. It moves to public sector payrolls and retirement income.

Law of Entropy. Complex systems built on circular dependencies do not fail gradually. They hold, and then they do not. The IMF is not warning about one bad company. It is warning about a financing architecture where the same small group of institutions are simultaneously the lenders, the equity investors, and the customers. When the largest debt issuers are also the largest revenue sources for the companies they lent to, a shock at any point in the loop hits every other point instantly. What is being done to your head: the sheer scale of the numbers makes the risk feel abstract. It is not. It is structural.