- Bloomberg puts outstanding AI data center debt above $500 billion right now
- CoreWeave isolates each loan within its own special purpose vehicle
- Parent companies report only a fraction of their actual total exposure, hiding the rest in shell entities.
A growing group of analysts now warn that AI data center debt is increasingly resembling the subprime mortgages that triggered the 2008 financial crisis.
Much of that debt is issued through special purpose vehicles, structures that keep billions of dollars off corporate balance sheets entirely.
Bloomberg estimates more than $500 billion in outstanding AI data center debt, of which approximately $200 billion is held by private credit funds.
A debt structure based on theoretical income
Special purpose vehicles (SPVs) generate debt to build data centers and then pay creditors only once paying customers start generating revenue.
That structure is exactly why CoreWeave has raised billions through separate SPVs for individual loans, including an $8.5 billion line of credit tied to the Meta contract, as each loan remains isolated within its own entity.
The same logic explains why Nikkei Asia reported that Meta, Google, Amazon, Microsoft and Oracle have accumulated about $1.65 trillion in debt over five years, much of it spread across similar vehicles rather than on a single balance sheet.
That gap between actual exposure and reported debt exists because these vehicles are jointly owned by outside investors, allowing the parent company to report only a fraction of the risk.
Meta’s Hyperion data center clearly shows the pattern: it is 80% owned by Blue Owl and only 20% owned by Meta itself, so most of the debt resides with Blue Owl on paper, even though Meta is the intended tenant.
Google has used the same approach, backing debt-funded data centers built by Fluidstack, Cipher Mining and TeraWulf without those obligations ever affecting its own balance sheet.
That type of arrangement is precisely what drew scrutiny from auditor Ernst & Young, which flagged Meta’s structure as a critical audit matter, questioning who ultimately bears its economic risk.
The stakes extend far beyond the companies involved, because pension funds and insurers are also directly exposed, and many now rely on data center returns to fund future payments.
Echoes of the 2008 mortgage collapse
The comparison with 2008 holds because both bubbles were based on the same flawed premise: that demand would continue to grow forever and would never need to be tested.
Subprime mortgages were proof of that thinking at the time, and by 2006, about 20% of all new mortgages issued in the United States were already classified as subprime, according to government data.
Instead of treating this as a warning sign, financial institutions bundled those loans into complex securities, a move that obscured the true underlying risk for both investors and rating agencies.
Financier Michael Milken captured the mood of the time when he publicly described such securities as a “financial innovation” that would vastly increase national prosperity and employment.
Reality caught up with that optimism once mortgage defaults began to rise sharply in 2005, and from there the damage spread to the entire financial system.
Lehman Brothers embodied how unbridled that confidence had become, trading with more than 25 times leverage in 2005 without serious reaction from regulators or rating agencies.
Today’s numbers reflect that same pattern of unexamined risk: analysts estimate more than $1.4 trillion in bank exposure to private credit, of which $300 billion is held by major banks.
Some estimates suggest that planned AI data center capacity exceeds actual annual computing demand by a factor of about 15 times.
Unlike 2008, this risk is not driven by derivatives but by the enormous scale of construction costs for individual data centers.
Whether this debt is paid off gradually or all at once likely depends on how quickly AI’s major customers can pay their bills.
For now, the scale of exposure across banks, pensions and insurers suggests that the comparison with 2008 is not merely rhetorical.
Via Ed Zitron
Follow TechRadar on Google News and add us as a preferred source to receive news, reviews and opinions from our experts in your feeds.
