AI's Money Runs in a Circle — and One Lender Just Showed How Thin It Is

In February, a lender was asked to put up $4 billion for a data center in Pennsylvania. The tenant would be CoreWeave, the company that rents out Nvidia chips and has become one of the biggest names in the AI buildout. The lender looked at the tenant's credit rating — B+, one notch below investment grade — and walked away. Blue Owl, the asset manager assembling the deal, could not find anyone to take its place. ^[1] The refusal is worth sitting with, because of what sat behind the number. CoreWeave carries 6.71 times net debt to EBITDA and is not expected to turn a profit until fiscal 2028. ^[2] Its halls are built by Applied Digital, which borrowed $2.35 billion at 9.25 percent and leased the finished buildings back to CoreWeave. ^[3] So one indebted company borrows at 9.25 percent to build for a second indebted company with a sub-investment-grade rating, and the second company's ability to pay the first rests on demand that has not arrived yet. That demand is supposed to come from artificial intelligence, but look at who funds it. Nvidia has invested $100 billion in OpenAI, $6.3 billion in CoreWeave, and $2 billion in xAI — companies that then buy Nvidia chips. ^[4] Meta, having built so much capacity that it is now selling the excess, is negotiating a $10 billion lease to rent compute to Anthropic. ^[5] The chipmaker's biggest customers are increasingly its own investments, and the companies that were rivals are becoming each other's landlords. Beneath that is a second layer, the one that hides how much debt is really there. Nvidia's $500 billion infrastructure fund, built with BlackRock and Goldman Sachs, offers "residual value support" — a guarantee that lets a customer borrow against Nvidia's credit rating without recording the debt on its own books. Rating agencies have begun calling these "phantom liabilities," and the total across the major AI companies runs to roughly $70 billion. ^[6] Michael Burry, who has shorted Oracle and Nebius, put a name to it. ^[7]

understating depreciation by extending the useful life of assets artificially boosts earnings — Michael Burry

Jim Chanos has been asking a version of the same question for a year. ^[4]

[Don't] you think it's a bit odd that when the narrative is 'demand for compute is infinite', the sellers keep subsidizing the buyers? — Jim Chanos

What Chanos is describing, and what Burry is shorting, is the same machinery Enron's broadband unit relied on to make a real business look larger than it was — deals that circle between related parties, debt kept where the rating agencies cannot see it, and demand forecasts that only hold if the financing keeps flowing. ^[8] None of it is illegal, and none of it is imaginary. The demand is real, in the way Enron's broadband demand was real. Vacancy in Northern Virginia has dropped to 0.3 percent, and landlords now quote space in kilowatts instead of square feet. ^[9] The industry is short roughly 12 gigawatts of capacity against about $2 trillion in contracted orders. Companies are signing 20-year power agreements and restarting dormant nuclear plants to feed it. ^[10] That is the trap the analogy points at, not the comfort. A business can have genuine customers and genuine revenue and still be built on a circle, because the customers are paid by the suppliers and the suppliers' balance sheets are the customers' collateral. The question was never whether anyone wants compute. It is whether the people renting it can keep paying, and the people building for them can keep borrowing, without the loop feeding itself. A chain like this does not need every link to break. It needs one. In February, one lender looked at a B+ credit rating and declined to be the next link. The money leaves Nvidia, becomes an investment in a customer, returns as chip revenue, and is lent out again against Nvidia's guarantee — a closed loop that only stops when someone outside it says no.