AI data centre debt tops $500bn — analysts warn of 2008-style crisis

Full text

Bloomberg pegs outstanding AI data center debt above $500 billion right now

CoreWeave isolates each loan inside its own separate special purpose vehicle

Parent companies report only a fraction of their real total exposure, hiding the rest in shell entities

A growing body of analysts now warns that AI data center debt increasingly resembles 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, with roughly $200 billion held by private credit funds.

A debt structure built on theoretical revenue

Special Purpose Vehicles (SPVs) raise debt to build data centers, then repay creditors only once paying customers begin generating revenue.

That structure is exactly why CoreWeave has raised billions through separate SPVs for individual loans, including an $8.5 billion facility tied to Meta's contract, since each loan stays isolated inside its own entity.

The same logic explains why Nikkei Asia reported that Meta, Google, Amazon, Microsoft and Oracle have accrued around $1.65 trillion in debt over five years, much of it spread across similar vehicles rather than sitting on any single balance sheet.

That gap between real exposure and reported debt exists because these vehicles are jointly owned with outside investors, letting the parent company report only a fraction of the risk.

Meta's Hyperion data center shows the pattern clearly: it is owned 80% by Blue Owl and only 20% by Meta itself, so the bulk of the debt lives with Blue Owl on paper even though Meta is the intended tenant.

Google has used the same approach, backstopping debt-funded data centers built by Fluidstack, Cipher Mining and TeraWulf without those obligations ever touching its own balance sheet.

That kind 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 well beyond the companies involved, because pension funds and insurers are also directly exposed, with many now relying on data center returns to fund future payouts.

Echoes of the 2008 mortgage collapse

The comparison to 2008 holds up because both bubbles rested on the same flawed premise: that demand would keep growing forever and never needed to be tested.

Subprime mortgages were the proof of that thinking at the time, and by 2006 roughly 20% of all new mortgages issued in the United States were already classified as subprime, according to government data.

Rather than treat that as a warning sign, financial institutions bundled those loans into complex securities, a move that obscured the true underlying risk from investors and rating agencies alike.

Financier Michael Milken captured the mood of the era when he publicly described such securities as a "financial innovation" that would broadly increase national prosperity and jobs.

Reality caught up with that optimism once mortgage defaults began rising sharply in 2005, and the damage cascaded through the entire financial system from there.

Lehman Brothers embodied how unchecked that confidence had become, operating at more than 25 times leverage in 2005 without serious pushback from regulators or rating agencies.

Today's numbers echo that same pattern of unexamined risk: analysts estimate more than $1.4 trillion in bank exposure to private credit, with $300 billion of it held by major banks alone.

Some estimates suggest planned AI data center capacity exceeds actual annual compute demand by a factor of roughly 15 times.

Unlike 2008, this risk is not driven by derivatives but by the sheer scale of individual data center construction costs.

Whether this debt unwinds gradually or all at once likely depends on how quickly major AI customers can pay their bills.

For now, the scale of exposure across banks, pensions and insurers suggests the comparison to 2008 is not merely rhetorical.

Via Ed Zitron