The ‘subprime AI CDO’ label is doing a lot of work in Jim Rickards’ argument, and understanding what a collateralized debt obligation actually is tells you where the comparison holds and where it stretches. Rickards’ AI Black Paper campaign has introduced the phrase in the context of synthetic leases, data-center securitizations, and off-balance-sheet debt structures that he says mirror the financial engineering that blew up the global banking system in 2008.

What a CDO Actually Is

A collateralized debt obligation is a financial product that bundles together loans, bonds, or other debt instruments and sells slices of the bundle to investors. Each slice carries a different risk level and a different return. The structure is called a tranche — French for ‘slice’ — and the analogy that works best is a layer cake.

The senior tranche gets paid first from the cash flows of the underlying loans. Because it is the safest piece, it earns the lowest return. The rating agencies assign it a AAA rating. The top tranche must be safe enough for pension funds that cannot hold speculative-grade paper.

The middle layer is the mezzanine tranche. It gets paid after the senior tranche. It carries more risk and a higher return. Institutional investors and hedge funds buy this layer.

The bottom layer is the equity tranche. It absorbs losses first. If any of the underlying loans default, the equity tranche takes the hit. It is the riskiest piece and earns the highest return. The bank that created the CDO usually holds it or sells it to specialist investors who understand they may lose everything.

The entire structure depends on one assumption: that the underlying loans do not all default at the same time. That is the same assumption that failed in 2008.

How 2008 Subprime CDOs Worked

The 2008 crisis began with mortgages — specifically subprime mortgages given to borrowers with poor credit, no verified income, and sometimes no down payment. Banks bundled thousands of these mortgages into CDOs. The rating agencies applied their models. The models said the senior tranches were safe because the mortgages were diversified across regions and borrowers, and because housing prices had never fallen nationally since the Great Depression.

The models were wrong on two counts.

First, the mortgages were not actually diversified. The banks had packed the CDOs with loans that shared the same underlying risk: borrowers who could not afford the payments when the teaser rates reset. When housing prices stopped rising, those borrowers could not refinance. They defaulted.

Second, the correlation assumption failed. The models assumed that a default in Florida was independent of a default in California. But when the housing market turned nationally, all the mortgages went bad at the same time. The diversification that was supposed to protect the senior tranche did not exist. The AAA-rated tranches collapsed alongside the equity tranches.

The damage spread beyond the CDOs. AIG had sold credit default swaps on them — insurance policies that promised to pay if the CDOs defaulted. AIG sold more than $440 billion in CDS protection without setting aside capital to cover the losses. When the CDOs failed, AIG needed a $182 billion government bailout. The banks that held each other’s CDOs created a second layer of contagion. When one bank’s CDO portfolio collapsed, it threatened the bank’s ability to honor its obligations to other banks, which threatened the entire system.

The total notional value of CDOs outstanding peaked at roughly $1.2 trillion in 2007. The damage was not caused by the size of the market but by the concentration of the risk and the leverage applied to it.

What Rickards Describes in the AI Context

Rickards’ AI Black Paper presentation describes a different debt structure operating on the same principles.

AI companies have raised at least $236 billion in debt in the first five months of 2026 alone, per his July 2026 campaign figures. Some of this debt sits in off-balance-sheet structures: special purpose vehicles, synthetic leases, and data-center lease securitizations. The largest example is Meta’s Hyperion campus in Louisiana, financed through a $27 billion SPV co-owned with Blue Owl Capital. The debt sits off Meta’s balance sheet, and the cash flows from the data center leases are what repays the bondholders.

Data-center lease securitizations bundle the cash flows from data center leases into bonds. This is structurally similar to how mortgages were bundled into CDOs. A special purpose vehicle holds the leases, issues bonds against the expected cash flows, and sells the bonds to investors in tranches. The senior tranche gets paid first from the lease payments. The equity tranche absorbs losses first if the tenants stop paying.

The tenants are the hyperscalers: Amazon, Microsoft, Meta, Google, and Oracle. They are the largest companies in the world. They have never defaulted on a lease. The question is what happens if they stop needing the capacity.

Rickards also points to what he calls circular financing — companies investing in each other to generate reported revenue. This is a separate claim from the CDO structure but it compounds the risk. If Microsoft is leasing space from a data center REIT that Microsoft also financed through a bond that sits in a CDO that Microsoft’s pension fund holds, the risk is not diversified. It is concentrated inside a closed loop, and the loop breaks when any single node fails.

Where the Parallel Holds

Opacity. The structures are complex enough that most investors cannot assess the underlying risk. The data-center lease securitizations are private placements, not public securities. The leases are commercial contracts, not standardized mortgage documents. You cannot pull up a prospectus and read the terms.

Concentration. A small number of issuers account for most of the debt. Five hyperscalers dominate the market. In 2008, a small number of originators dominated the subprime mortgage market. Concentration means the risk is not diversified. It is sitting in the same place.

Rating agency reliance. The data-center securitizations are rated by Moody’s, S&P, and Fitch — the same agencies that rated subprime CDOs AAA in 2006. The agencies have a structural conflict of interest: the issuer pays for the rating, and the rating affects whether the bond can be sold. The agencies have improved their models since 2008, but the fundamental incentive problem has not been fixed.

Correlation risk. This is the most important parallel. If AI spending slows, do all the data center leases go bad at the same time? The answer is yes, because the spending is driven by the same macro thesis. If AI adoption plateaus, the spending that justifies all this debt slows simultaneously. The diversification that protects a lease securitization in normal times evaporates when the macro thesis breaks — exactly the same correlation failure that killed the subprime CDOs.

Where the Parallel Breaks

Scale. The 2008 CDO market peaked at $1.2 trillion. The AI debt market is large — $236 billion in five months, $570 billion projected for 2026 — but the structures are not yet as deeply layered. There are no CDO-squared equivalents.

Counterparty risk. In 2008, AIG’s CDS exposure created a systemic cascade. When the CDOs failed, AIG failed, and the banks that held AIG’s paper failed with it. No equivalent insurer sits behind AI data center debt. There are credit default swaps on hyperscaler debt — JPMorgan built a basket in February 2026 — but the CDS market is smaller and the counterparties are better capitalized.

Underlying assets. This is the critical difference. Mortgages in 2008 were given to people who could not pay. The loans were structurally unsound from origination. Data center leases are backed by hyperscaler revenue from actual businesses. Microsoft, Amazon, and Google have real earnings, real cash flow, and real pricing power. The leases are not subprime loans. They are commercial contracts between the largest companies in the world.

Transparency. Post-2008 regulations require more disclosure than existed in 2007. Dodd-Frank mandates risk retention for securitizations — the issuer must keep some of the risk on its books. FASB’s ASC 842 and IFRS 16 pull many off-balance-sheet lease obligations back onto the balance sheet or into footnotes. The disclosure has gaps, but it is better than the complete opacity of 2007.

The Revenue Gap

The strongest argument against the AI debt structure is not the CDO comparison. It is the revenue gap.

A March 2026 Vanderbilt working paper titled ‘After the AI Crash’ compared the AI investment cycle to 2008 and argued that the economy’s overreliance on AI investment, coupled with opaque financial engineering, means a correction could look more like 2008 than the dot-com bust. The paper notes that JPMorgan anticipates $5 trillion in AI infrastructure investment over the next five years. Bain estimates the industry needs $2 trillion in annual revenue to justify that spending.

The actual numbers are far smaller. OpenAI earned roughly $13 billion in revenue in 2025. Anthropic earned roughly $4 billion. The hyperscalers combined earn more, but their AI infrastructure spending is also larger. The gap between the buildout cost and the revenue base is where the structural risk lives.

Morgan Stanley estimates a $1.5 trillion financing gap for AI infrastructure through 2028. That gap has to be filled by external capital — corporate bonds, private credit, and structured products. If the revenue falls short, the debt service does not stop. The leases keep running. The bonds keep needing to be paid.

The 2008 parallel works best at this level. Both cycles involved a massive buildout financed by debt that assumed the cash flows would materialize on schedule. In 2008, the cash flows were mortgage payments. In 2026, the cash flows are AI revenue. In both cases, the debt structure is resilient in normal conditions and fragile when the macro thesis breaks. The assets are different. The mechanism is the same.

The subprime AI CDO label is useful as a warning about structure and opacity. It is not a prediction about timing or magnitude. The 2008 CDOs had a specific trigger — resetting teaser rates on millions of mortgages — that the AI debt structures do not share. The 2008 crisis had a specific counterparty — AIG — that created a cascade the AI debt market does not currently have.

The structural risk is real. The historical parallel is instructive. But the parallel is about the shape of the risk, not the destination. The 2008 CDOs collapsed because the underlying loans were bad. The AI debt structures will be tested by whether the revenue arrives, and that question has not been answered yet.

Flak Jacket Finance covers investment newsletters as an independent third party. We do not reveal paid picks, we do not trash the gurus or products we cover, and we do not sell the promos we cover.