DEAN BAKER: As in the housing bubble, complex financing structures cannot hide weak underlying assets, and their spread is itself a warning sign

The Gist

In the housing bubble, banks bundled shaky mortgages into complicated securities, but the securities were only as good as the mortgages. Lenders also ignored that payments would jump. Baker sees the same warning in AI's growing use of complicated off-balance-sheet deals. This packet steelmans the arguments made by Ed Zitron and Dean Baker in their best light for analysis and is not an endorsement of their conclusions.

Conclusion

The housing bubble showed that layered financing structures cannot make weak underlying assets sound and that lenders can fail to test whether borrowers can meet payments scheduled to rise, so the spread of special purpose vehicles, leasebacks, and similar structures in AI finance is a warning sign that deserves scrutiny of the underlying cash flows.

Premises

1. In the housing bubble, mortgages were pooled into securities and repackaged into tiered collateralized debt obligations, yet the value of every layer still depended on whether the underlying mortgages would be paid.
2. Many subprime mortgages carried low initial rates that were scheduled to reset two to three percentage points higher, and lenders often did not assess whether borrowers could pay the reset rate.
3. When the resets arrived, defaults and foreclosures followed and the structures built on those loans lost value, even though the risk was visible in advance.
4. Firms with strong, simple cash flows rarely need elaborate structures, so a shift toward special purpose vehicles and off-balance-sheet financing often indicates that the simple path is costly or unavailable.
5. AI infrastructure is now financed through project vehicles, leasebacks, and private credit whose repayment depends on a small number of tenants and their customers.

Assumptions

• Special purpose vehicles and leasebacks also serve legitimate purposes, such as isolating project risk and matching long-lived assets to long-term investors, so their use is a reason for scrutiny rather than proof of a problem.
• AI data center debt is not currently tranched and resold on the scale of mortgage securities in 2006 and 2007, so the analogy concerns underlying asset quality rather than identical structures.

Analysis

Argument type: Inductive.

Potential Fallacies

• Weak analogy / unexamined disanalogies (Conclusion; inference from Premises 1-3 to Premise 5) — The housing case is described in granular, documented detail, while the AI case gets one general sentence. The argument shows a shared surface feature (layered or off-balance-sheet structures) but not the causally relevant ones: weak underlying credit, scheduled payment shocks, or untested repayment capacity. The stated assumptions narrow the analogy to asset quality, which leaves that quality, the very thing at issue, undemonstrated. Housing involved many households and bank-held leverage, whereas…
• Transposed conditional / structure as diagnostic signal (Premise 4) — Premise 4 moves from "distressed or constrained firms often use complex structures" to "complex structures indicate distress." Without base rates for benign use in project finance, aircraft leasing, utilities, and real estate, the signal's diagnostic value is unknown. The author's concession of legitimate uses limits this to a probabilistic weakness rather than a strict error, but the vague quantifiers ("rarely," "often") leave it hard to evaluate.
• Hindsight bias (Premise 3) — Saying the housing risk was "visible in advance" judges what was knowable with the outcome in hand. A minority of analysts did warn and underwriting weaknesses were documented, but most regulators, raters, and forecasters did not see the scale. This overstates how easily comparable signals can be read now.
• Selective reference class / guilt by association (Premises 1 and 4; source framing around Enron) — Anchoring on Enron-era vehicles and the 2007 CDO collapse, which are notorious failures, samples on the outcome and lends those connotations to AI vehicles. Many SPV, leaseback, and project-finance booms did not end in crisis, and these are not weighed. The author's concession of legitimate uses partly offsets this, but the framing still does persuasive work the evidence has not earned.
• Presupposition of weak assets (and near-unfalsifiability) (Title/conclusion; Premise 5) — The title asserts that structures cannot hide weak assets, implying the AI assets are weak, whereas Premise 5 shows only concentration of tenants. Concentration is a risk factor, not evidence of inability to pay. Because "warning sign deserving scrutiny" has no stated threshold, almost any outcome can be read as consistent with it.

Counterarguments

These challenges are not ranked.

• Premise 5 (High impact) — The principal tenants are investment-grade hyperscalers with very large free cash flow, often on long-term, take-or-pay leases. This is the opposite of subprime borrowers, so tenant concentration does not imply weak underlying assets.
• Premise 4 (High impact) — Project finance, sale-leasebacks, and securitization are routine for strong, simple cash flows in infrastructure, aircraft leasing, utilities, and real estate. Without base rates, SPV growth could be a healthy sign of capital-intensive buildout, and treating it as a signal would flag nearly all modern capital markets.
• Conclusion (High impact) — The subprime reset mechanism has no clear AI counterpart. The real risks in AI finance are demand uncertainty, GPU obsolescence versus debt tenor, residual values, and refinancing. The analogy risks directing scrutiny to the wrong failure signals.
• Premises 1-3 (Medium impact) — Housing became a systemic crisis through household leverage, bank balance-sheet exposure, ratings failures, and funding runs. AI debt is held mainly by sophisticated institutional investors and is not tranched at that scale, so losses could be real but contained, or fall on equity holders rather than the banking system.
• Conclusion (Medium impact) — Once the assumptions are granted, the claim reduces to "scrutinize the cash flows," which nobody disputes. The "warning sign" and Enron framing then adds alarm that the premises do not support, and the claim is hard to falsify.
• Premise 3 (Low impact) — The risk was visible only to a minority at the time, and the historical record shows defaults driven partly by price declines and negative equity rather than resets alone. The lesson about how to separate real warnings from noise is underdeveloped.

Suggested Improvements

• Supply AI-side evidence — Cite specific deals, volumes, leverage, tenant identities and credit ratings, lease lengths, take-or-pay terms, and the share of AI capex financed through these vehicles. The housing side is richly documented while the AI side is one sentence; concrete data would let readers judge whether the underlying cash flows are actually weak.
• Identify the AI analogue to payment shocks — Name the mechanism that corresponds to scheduled resets, such as floating-rate private credit, balloon or refinancing risk, rent escalators, or GPU depreciation against longer debt terms, and show it is present and untested. The lesson about lenders failing to test payment capacity only transfers if a comparable vulnerability exists.
• Quantify the structure signal — Replace "rarely" and "often" with base rates or historical evidence on how often rapid growth in SPV and leaseback use has preceded distress versus accompanied healthy buildouts, including telecom and aircraft leasing. This would establish the diagnostic value of Premise 4, which currently carries the main inferential load.
• Distinguish benign from concerning structures — Offer concrete diagnostics such as disclosure quality, sponsor recourse and guarantees, related-party and circular flows, tenant concentration, leverage against contracted revenue, and stress-test coverage. This makes the heuristic actionable and falsifiable, and prevents it from flagging every project-financed asset.
• Address systemic transmission and the strongest rebuttal — Engage directly with the creditworthy-tenant defense, and examine who holds the debt and whether losses would reach banks, insurers, or retail-linked funds. Compare with the 1990s telecom buildout. This would show whether the concern is investor losses or a broader crisis, and would make the argument more persuasive to skeptics.
• State what would reduce concern — Specify thresholds or observable outcomes, such as coverage ratios or tenant credit quality, that would indicate the cash flows are sound. A warning with no disconfirming evidence is hard to distinguish from a standing narrative.

Scenario Tests

• Key tenants are investment-grade hyperscalers whose long-term leases are comfortably covered by strong cash flow, and the vehicles perform as expected. (Challenges) — The underlying cash flows would be sound, and the warning would look like pattern-matching. The call for scrutiny would still be harmless, but the "warning sign" claim would be discredited.
• A large share of the debt is tied to loss-making AI labs or GPU-cloud firms whose revenue depends on venture funding, with customers who are themselves unprofitable and circular financing links among suppliers, tenants, and investors. (Supports) — Correlated repayment risk would be real, and the housing lesson about untested payment capacity would transfer. Scrutiny of cash flows would be well justified.
• Stress arrives through a GPU obsolescence cliff or demand shortfall rather than payment resets or sloppy underwriting, and losses fall mostly on equity and private-credit investors. (Neutral) — Scrutiny of cash flows would be vindicated, but the reset-based mechanism and the 2008-style systemic framing would not apply. The analogy would misdirect attention to the wrong risk.
• SPV and leaseback use grows rapidly in another capital-intensive sector such as aircraft leasing or power projects, with no subsequent distress. (Challenges) — This would show that growth in such structures is a noisy, high-false-positive signal, weakening Premise 4's diagnostic value.
• Disclosures reveal opaque off-balance-sheet commitments, related-party guarantees, and lenders skipping stress tests on concentrated tenants. (Supports) — This would restore the parallel with the housing lenders' failure to test payment capacity and justify heightened scrutiny.

Coherence & Relevance

The argument has a clear three-part architecture: a documented historical case, a general heuristic linking structure to constrained financing, and a present-day instance. The conclusion is hedged to "warrants scrutiny," and the stated assumptions are consistent with that scope, so it is coherent and defensible as an inductive case for inquiry. The main gap is an unstated bridging premise that AI cash flows are potentially weak or untested, along with the absence of an AI counterpart to the payment-shock mechanism. As a result, the argument supports a modest conclusion (look closely at the cash flows) much better than a stronger one (that this resembles pre-crisis conditions).

• In the housing bubble, mortgages were pooled into securities and repackaged into tiered collateralized debt obligations, yet the value of every layer still depended on whether the underlying mortgages would be paid. (Strong) — Establishes the principle that structure cannot create value, directly supporting the first half of the conclusion. It does not itself show that AI assets are weak.
• Many subprime mortgages carried low initial rates that were scheduled to reset two to three percentage points higher, and lenders often did not assess whether borrowers could pay the reset rate. (Moderate) — Supports the lesson that lenders can fail to test repayment capacity, but no AI parallel to scheduled payment increases or documented underwriting failures is shown.
• When the resets arrived, defaults and foreclosures followed and the structures built on those loans lost value, even though the risk was visible in advance. (Moderate) — Shows the consequence of ignoring weak assets, but the causal role of resets is oversimplified and "visible in advance" is not tied to anything observable in AI finance.
• Firms with strong, simple cash flows rarely need elaborate structures, so a shift toward special purpose vehicles and off-balance-sheet financing often indicates that the simple path is costly or unavailable. (Strong) — This is the bridge that turns the spread of structures into evidence of weakness. It lacks base rates and does not address non-distress motives beyond the conceded general caveat.
• AI infrastructure is now financed through project vehicles, leasebacks, and private credit whose repayment depends on a small number of tenants and their customers. (Strong) — Supplies the only present-day fact. It lacks specifics on volume, tenant credit quality, and contract terms, and concentration does not by itself imply weakness.

View this argument on LogicFirst.ai