Carlyle Research Head: AI Data Center Financing Model Draws Parallels to Eve of Subprime Crisis
nashnova research
Carlyle Group's head of research Jason Thomas warns that banks are lending to AI data centers the same way they lent before the 2005–2007 subprime crisis — underwriting the sponsor's credit rating, not the project itself. That 'don't judge the underlying asset' logic is what broke last time.
What exactly is he saying? What does "like subprime" mean?
Thomas says banks financing AI data centers today do not assess whether the projects themselves will generate returns. They lend against the investment-grade credit of the financial sponsor behind the deal.
In plain terms = the bank's position is "I don't care if this data center makes money — your credit rating is good enough, so you'll cover any losses."
That is the same logic that drove pre-2007 subprime mortgage lending — banks didn't evaluate the collateral (the houses); they trusted that the originating bank could absorb losses on its balance sheet.
Why did that logic collapse last time?
Before the subprime crisis, lenders claimed they didn't need to judge underlying mortgage quality because the originating banks could always take bad loans back onto their own balance sheets.
This means → the entire system's safety net was one layer deep: the originating bank's capacity to absorb losses. Once liabilities grew beyond that capacity, the logic snapped.
In plain terms = imagine a climbing team where every rope is anchored to a single bolt — fine while the bolt holds, but add enough weight and the bolt rips out.
Where does AI financing stand right now?
Thomas concedes the system has not yet reached a danger threshold — investment-grade sponsors still have strong credit quality.
But he warns: as AI infrastructure buildout accelerates, data-center-related liabilities keep rising and the implied volatility of associated returns is widening.
This means → if the probability of underwhelming project returns keeps growing while leverage keeps climbing, the "don't judge the project, just trust the sponsor" logic will break the same way subprime did.
Is the market pricing this risk correctly?
Thomas notes that AI risk premiums in both credit and equity markets are narrowing — in other words, the market is growing less worried about AI investment risk, not more.
This reflects a classic boom-cycle pattern: when capital floods in, risk pricing gets systematically compressed.
His bottom line is blunt: AI investment is more likely to slow down than not.
市场有风险，内容仅供研究参考，不构成投资建议。