ED ZITRON AND DEAN BAKER: The AI build-out rests on obligations its main customers cannot yet fund, and an unwinding would push losses onto lenders, investors, and ordinary savers

The Gist

The AI build-out is paid for with promises from two companies, OpenAI and Anthropic, that can't yet afford them. The risk has been passed along to banks, private credit funds, and the pension funds and savers behind them. Oracle is the weakest link. No one can say when it unravels, but the longer it runs, the bigger the losses. 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 AI data center build-out is financed on long-term obligations that its two principal customers cannot currently fund, through structures that move the risk from the technology firms to banks, private credit, bondholders, and ultimately the insurers, pension funds, retail investors, and workers behind them; Oracle is the most exposed point, and if the build-out unwinds, at a time no one can predict, the losses will be larger the longer it continues.

Premises

Amazon's reported plan to move about $8 billion of installed Nvidia chips into an investor-funded vehicle and lease them back is, in economic substance, secured borrowing against fast-depreciating assets, and a company of Amazon's scale reaching for that structure while its free cash flow is negative is better read as a sign of funding strain from AI spending than as routine balance-sheet management.
A large share of the AI chips sold since 2024 appears not yet to be installed in powered data centers, so headline chip sales and capital spending overstate how much working AI capacity exists and how much paying demand it serves, and chips waiting for power lose economic value as newer generations arrive.
Oracle has made AI data center rentals, concentrated in OpenAI, its main source of growth and has financed them with heavy borrowing and leases it reportedly cannot terminate, so it cannot easily retreat if that demand falters, and its credit position already reflects that exposure.
Oracle is to the AI build-out what Bear Stearns was to the 2008 crisis, the smallest and most leveraged major participant, and its distress could spread through investor confidence, the funding of other AI infrastructure firms, OpenAI's growth plans, and Larry Ellison's own collateralized commitments, even though the larger hyperscalers have far stronger businesses.
OpenAI and Anthropic have signed largely non-cancelable compute commitments whose payments ramp up sharply in 2027 and 2028, much as adjustable-rate mortgages reset in 2007, and neither their current revenue, nor venture capital, nor a strained high-yield bond market offers a credible source for the hundreds of billions of dollars a year those commitments will require.
The labs' commitments assume revenue that compounds rapidly for years, but recent evidence points the other way: Anthropic's annualized revenue appears to have flattened, OpenAI's widely quoted headline figure turned out to be an investor gross-up about $20 billion above OpenAI's own net figure, showing how loosely these metrics are reported, competition from low-cost Chinese models is forcing price cuts, and heavily promoted free consumer apps do not yet show a path to matching revenue.
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.
If AI demand falls short, losses will concentrate in a relatively small number of very large data center loans held by banks, private credit funds, and the insurers and pension funds behind them, and recoveries are likely to be poorer than in the housing crash because AI facilities are specialized assets whose demand depends heavily on two unprofitable tenants.
AI data center finance has become a self-reinforcing loop: construction delays add interest during build-out, heavy buying drives up chip, memory, and materials prices, heavy borrowing pushes up the cost of debt, and the end demand that must repay it all depends on venture-funded customers whose funding is under strain, so the system grows more fragile the longer it runs.
The fact that large banks, asset managers, and technology firms are funding the AI build-out is weak evidence that it is sound, because lenders earn fees on each deal regardless of its long-term outcome, and well-resourced firms have recently made very large misjudgments, such as Meta's metaverse spending.

Assumptions

The argument does not predict timing; both speakers say the housing and dot-com bubbles lasted longer than they expected.
The largest hyperscalers fund most of their AI spending from strong cash flows and remain highly rated, so the risk is concentrated in Oracle, neoclouds, the two labs, and project lenders rather than spread evenly.
Revenue at the labs is still growing; OpenAI told investors that its total run rate grew 77 percent during the third quarter of 2026, as CNBC reported from a person familiar with the presentation, and the October change from a reported $70 billion to $50 billion reflected a switch from an investor gross-up to OpenAI's net figure, not a decline in sales, so the funding gap could narrow.
Several figures cited in the conversation were corrected by research, including copper costs, tenant concentration, Oracle's revenue history, credit default swap levels, and the basis of OpenAI's headline revenue figure, without changing the direction of the argument.
As of early October 2026, no major AI lab, hyperscaler, or AI data center lender had defaulted.

Analysis

Argument type: Inductive.

Potential Fallacies

Argument from analogy treated as mechanism (P4, P5, P7, P8) — Bear Stearns, adjustable-rate resets and the housing bubble are used to suggest how AI finance will behave. Bear was a short-term-funded broker-dealer exposed to runs, and resets hit households with no bargaining power. Here the borrowers are corporations that can renegotiate, raise equity or be backstopped, and the lenders are fewer and better capitalized. The analogies show a possible shape of risk but do not establish that the same transmission channels exist or how large they are.
Ambiguous evidence read as confirmation (P1, P2, P6) — A reported leaseback, chips awaiting power, and one lab's flat annualized revenue each have benign explanations (cost of capital, tax and accounting, power-interconnection lead times, metric basis changes, lumpy contracts). Reading them uniformly as distress signals without weighing alternatives risks confirmation bias. The OpenAI gross-up in particular shows unreliable reporting, not falling sales, as the assumptions themselves state.
Unsupported universal negative / appeal to ignorance (P5) — The claim that no credible funding source exists is inferred from a short list of candidates (current revenue, VC, high-yield debt). It leaves out equity raises, strategic and sovereign capital, hyperscaler support, renegotiation and revenue growth. The absence of an identified source is not proof that none will appear.
Present-tense overstatement (future gap framed as current inability) (Conclusion vs. P5 and A3) — The conclusion says the customers 'cannot currently fund' the obligations, but the premises describe a projected shortfall against payments that ramp in 2027-28. Assumption A3 concedes that revenue is growing and the gap could narrow. A projected gap is a weaker claim than a present inability.
Single-instance generalization (P1, P10) — One reported $8 billion transaction by one very large, highly rated firm is read as evidence of funding strain. Without a base rate for how often comparable firms use such structures at similar free-cash-flow levels, the inference is thin. Similarly, Meta's metaverse is a single example of institutional misjudgment.
Unfalsifiable framing (Conclusion, A1, A5, P9) — With no timing, no thresholds, and the claim that losses grow the longer the build-out lasts, continued non-default can be read as accumulating risk, so almost no observation counts against the thesis. The hedging avoids a failed timing prediction but lowers the thesis's testable content.
Appeal to consequences / vulnerability conflated with probable harm (Title, P8, Conclusion) — The prominence of pensioners, savers and workers lends emotional weight to the conclusion, but the severity of a possible outcome does not make it more likely. The transmission path from a few large loans to household losses is asserted, not quantified relative to the balance sheets of insurers, pensions and banks.

Counterarguments

These challenges are not ranked.

Conclusion / P5 (High impact) — The 'cannot currently fund' claim conflates a projected 2027-28 gap with present inability. OpenAI's net revenue is growing about 70% a quarter, and the labs have repeatedly raised equity, strategic and sovereign capital beyond skeptics' expectations. The gap may be met, or narrowed, without default.
P5 / P8 (High impact) — 'Non-cancelable' commitments between interdependent counterparties (Microsoft, Nvidia, Amazon, Google, Oracle) are commonly renegotiated, extended or restructured, because suppliers are also equity holders with reasons to keep the labs solvent. Losses would then fall on well-capitalized equity holders rather than lenders and savers.
P4 / P7 / Conclusion (High impact) — The 2008 analogies fail on key mechanisms. AI debt is smaller, concentrated among a few sophisticated parties, and not embedded in household balance sheets or the payments system, and banks are better capitalized. The closer precedent is the telecom and fiber overbuild, where lenders and equity investors took heavy losses but there was no system-wide crisis.
P1 (Medium impact) — Sale-leasebacks are routine tools for tax, accounting and cost-of-capital reasons, and negative free cash flow is the product of a chosen capex surge. A highly rated firm using the structure is at least as consistent with optimization as with strain.
P8 (Medium impact) — Powered sites, grid interconnections and shells are scarce and re-leasable, and GPUs retain value for inference, so recoveries may be better than the 'specialized asset' framing implies. Falling compute prices from a glut could also stimulate new demand.
P6 (Medium impact) — The OpenAI restatement was a change of metric, not a sales decline, and Anthropic's apparent flattening is one noisy data point. Price cuts from Chinese competition may raise volumes, so the evidence can be read as consistent with continued growth.
P2 (Medium impact) — Chips awaiting power are normal pipeline inventory in a supply-constrained build-out, and power bottlenecks indicate demand outrunning capacity. Delay is not evidence of absent paying demand.
P7 / P10 (Medium impact) — If SPVs and layered financing signal a bubble, so do most utility, aircraft, shipping and REIT financings; and if lender participation carries no evidential weight, no market signal could ever count against the thesis. Both lines of reasoning prove too much without a criterion separating AI structures from ordinary ones.

Suggested Improvements

Wording of the central claim — Replace 'cannot currently fund' with a precise claim such as 'has no identified funding path for the 2027-28 payment ramp under plausible revenue growth', and specify the growth rate required. This matches the claim to what the premises show and what A3 concedes, and removes the easiest objection.
Quantification of the revenue gap — Set the scheduled 2027-28 payments against net-basis lab revenue under explicit growth scenarios, and state what growth rate closes the gap and how likely that is. The whole argument rests on whether revenue grows into the commitments; the comparison is asserted, not shown.
Exposure data for the transmission to savers — Provide aggregate data center loan and bond balances, holdings by banks, private credit, insurers and pensions, and compare these to those institutions' capital and to 2007-08 mortgage-linked assets. The claim that losses reach savers and workers depends on size, which is currently unaddressed.
Contract terms and renegotiation precedent — Document take-or-pay terms, termination rights, parent guarantees and strategic-investor backstops, and cite how comparable commitments have been restructured historically. This tests whether the adjustable-rate analogy is apt or overstates rigidity.
Falsifiability — State leading indicators and thresholds, such as Oracle spreads and coverage ratios, compute rental prices, lab fundraising terms, and whether the labs meet 2027 payments without renegotiation, that would weaken the thesis if observed. This keeps the timing humility while giving the claim testable content and answering the 'perma-bear' critique.
Alternative explanations for ambiguous signals — Address benign explanations for the Amazon leaseback and for uninstalled chips, and compare against how often large firms use leasebacks at comparable cash-flow levels. Strengthens the evidential value of P1 and P2 by showing they discriminate between strain and routine management.
Disanalogies and the better precedent — Engage the telecom and fiber overbuild as a competing reference class and explain why the AI case would be more severe or more systemic. Choosing the most apt historical comparison makes the argument more credible than relying on 2008 alone.

Scenario Tests

Lab revenue keeps growing at 70% or more, and OpenAI and Anthropic meet the 2027 step-ups from revenue and new equity raises without renegotiation. (Challenges) — The central funding-gap premise fails and the risk-transfer claim remains hypothetical. The thesis would then need to retreat to Oracle-specific leverage and the claim that structures were fragile but not tested.
OpenAI's growth stalls, its fundraising falters, and Oracle cannot terminate its OpenAI leases while its debt service rises. (Supports) — Oracle becomes the first visible point of failure, as the argument predicts, and the loan-concentration and self-reinforcing-loop premises gain substance. Spread to other firms would depend on shared lenders and investor confidence.
The labs' commitments are restructured with hyperscaler or strategic investors absorbing the shortfall through equity or extended terms. (Challenges) — Losses stay with well-capitalized equity holders, so the claim of transfer to lenders, pensions and savers is overstated, even though the commitments were indeed unaffordable on their original terms.
An Oracle default or restructuring occurs and is orderly, with limited effect on other AI infrastructure lenders and no systemic stress. (Neutral) — Confirms Oracle as the weak point but undercuts the Bear Stearns contagion analogy. The argument's narrower claim holds while its systemic claim does not.
The build-out continues for several more years with no defaults and rising lab revenue while leverage accumulates. (Neutral) — Because the thesis gives no timing and claims losses grow with duration, this outcome is consistent with it, which highlights the thesis's low falsifiability. Only measurable indicators such as coverage ratios and loan exposures could distinguish a growing-into-it outcome from a deferred bust.
A demand collapse leaves powered data centers and GPUs with little alternative use, and loans held by insurers and private credit funds are written down heavily. (Supports) — Supports the specialized-asset, concentrated-loss and poorer-recovery claims, and shows how losses reach savers through private credit allocations. The scale relative to institutional capital would determine whether this is systemic.

Coherence & Relevance

The argument forms a convergent cumulative case: P3 and P5 set out present exposure and a future funding gap, P6 and P9 offer reasons the gap may not close and may worsen, P7 and P10 supply analogy and neutralize a counterargument, and P8 delivers the distributional claim. The stated assumptions are consistent with a narrowed conclusion of concentrated fragility rather than universal collapse, and the explicit refusal to predict timing is honest. Coherence is weakest at three joints: the shift from a projected 2027-28 gap to 'cannot currently fund'; the shift from a concentrated, commercial lender exposure to housing-scale transmission to workers and savers; and the dependence of several premises on one unresolved question, whether lab revenue grows into its obligations. The conclusion is conjunctive and its joint probability is lower than any single limb, with the Oracle-exposure and conditional 'worse the longer it runs' limbs the best supported.

P1: Amazon leaseback as secured borrowing and a sign of strain (Moderate) — Relevant to the claim that risk is being moved through financing structures, but the inference to strain is ambiguous, and Amazon sits in the group A2 describes as strong. It supports the structure claim better than the 'cannot fund' claim.
P2: Chips sold but not powered (Moderate) — Bears on whether capex overstates working capacity and on asset obsolescence, but does not directly show insufficient paying demand or the labs' inability to pay. The magnitude is unmeasured.
P3: Oracle's concentrated, leveraged and locked-in exposure (Strong) — Directly supports the 'Oracle is the most exposed point' limb. Its note that credit already reflects the exposure sits in some tension with treating Oracle as an underappreciated weak spot.
P4: Oracle as Bear Stearns (Moderate) — Supports the contagion portion of the conclusion, but depends on an analogy rather than evidence of specific transmission channels, and is limited by the concession that hyperscalers are strong.
P5: Non-cancelable lab commitments ramping in 2027-28 (Strong) — Central to the claim that the principal customers cannot fund the obligations, but it shows a future gap rather than a present inability, and it omits renegotiation and equity-funding possibilities.
P6: Evidence that revenue is not compounding (Moderate) — Meant to show that the revenue needed to fund P5 will not arrive, but its strongest datum (the OpenAI gross-up) concerns metric quality, not growth, and A3 reports strong net growth. Only the reporting-unreliability point holds firmly.
P7: Housing lessons on layered financing (Moderate) — Supports the call for scrutiny of structures and cash flows, but the structures are common in sound finance, and the premise presupposes the weakness of the underlying assets.
P8: Concentrated losses and poor recoveries (Strong) — Directly supports the distributional limb of the conclusion, but offers no exposure or recovery data and does not show the scale needed to reach pension funds and ordinary savers.
P9: Self-reinforcing loop (Strong) — Supports the 'losses larger the longer it continues' limb. It omits damping feedbacks and shares the same underlying dependence on lab revenue as P3, P5 and P6, so its support is not independent.
P10: Lender participation is weak evidence of soundness (Moderate) — Defuses an obvious objection but provides no positive evidence for the conclusion, and it ignores that many lenders retain credit risk.

View this argument on LogicFirst.ai