ED ZITRON: AI data center losses would land on banks, private credit, and their investors, and specialized buildings serving two main tenants recover less than houses did

The Gist

The money lent to build AI data centers comes largely from private credit funds, whose backers include insurers and pension funds. The loans are huge and few, so a handful of failures could do real damage. And unlike a house, an empty AI data center has few other uses. 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

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.

Premises

1. Private credit already holds over $200 billion of AI-related loans and is projected to supply around $800 billion of data center financing through 2028, with insurers and pension funds among the ultimate investors.
2. Single data center financings now run from billions to tens of billions of dollars each, such as $18 billion of bank loans for New Mexico's Project Jupiter, $14 billion of notes for Oracle's Michigan campus, and about $27 billion of debt for Meta's Hyperion vehicle.
3. Because the exposure is concentrated in a smaller number of very large loans, a few failures can produce large losses, unlike a mortgage book spread across millions of loans.
4. A foreclosed house can still be lived in and sold, but an AI data center is built for specialized chips, power, and cooling, and its value depends on finding tenants willing to pay for AI compute.
5. OpenAI and Anthropic account for a large and growing share of AI compute demand, and a rented GPU fleet loses money quickly when underused, so the tenants behind much of this debt are concentrated in two unprofitable firms.
6. Private credit discloses less than banks and public markets, so the full size and location of the exposure is hard to measure in advance.

Assumptions

• Estimates of the two labs' share of AI compute differ: one industry estimate puts them at about 30 percent of new capacity in 2026 and 40 to 50 percent of 2027 additions already committed, below the 70 to 80 percent cited in the conversation.
• Banks' direct AI-related exposure has been estimated at under 1 percent of assets, although banks also lend to private credit funds.
• The count of 500 to 1,000 loans of $1 billion to $50 billion is the speaker's rough estimate and was not verified.
• Data center shells, power connections, and land retain some value for non-AI cloud computing, so recovery is lower than for housing but not zero.
• Running costs for a facility do not fall much over time, so an idle AI data center remains expensive to keep.

Analysis

Argument type: Inductive.

Potential Fallacies

• Weak analogy / idealized benchmark (P4 and conclusion) — The comparison treats foreclosed houses as a benchmark of good recovery, but recoveries in overbuilt housing markets were often poor, and the 2008 crisis was driven mainly by leverage, securitization and correlated defaults. Without recovery-rate data for either asset class, 'worse than housing' is illustrative rather than demonstrated. The comparison would be stronger against telecom fiber, shipping or commercial real estate, where recovery data exist.
• Conflating exposure or concentration with expected loss (P3) — Fewer, larger loans raise the variance and tail risk of losses. They do not by themselves raise the chance that loans fail. Large loans are often made to strong sponsors with long-term leases and guarantees. Which loans are exposed depends on credit quality, seniority and correlation, none of which the argument examines. Concentration of loans is also not the same as correlation among them.
• Equivocation between demand share and debt backing (P5 and title) — That two labs account for a large share of compute demand does not mean they are the credit obligors on the loans. Many facilities are leased to or guaranteed by investment-grade hyperscalers, and the labs often contract indirectly through them. The title's 'two main tenants' framing is stronger than the evidence, even granting the lower share estimates in the assumptions.
• Hasty generalization from flagship deals (P2 to P3) — Three of the largest financings are used to characterize a market of purportedly hundreds of loans, and the loan count is itself an unverified estimate. The biggest deals may differ from the typical loan in structure, tenant quality and credit support.
• Opacity treated as evidence of danger (P6) — Limited disclosure widens the range of possible outcomes in both directions. It is a reason for more inquiry and does not by itself show that the exposure is large or poorly structured. It also means the argument's own numbers are hard to verify.
• Unsupported causal bridge (missing premise) (Conclusion) — The argument never states how a demand shortfall becomes tenant distress, lease default, loan default, collateral shortfall and finally loss to end investors. Each link depends on contract terms (take-or-pay leases, guarantees, residual value support) and on the equity layer that absorbs first losses. The step to pension funds and insurers is similarly assumed, since many hold senior, investment-grade tranches.
• Projection treated as exposure (P1) — The $800 billion figure is a forward estimate that depends on the continued demand growth whose failure the argument supposes. It should not be read as realized exposure.

Counterarguments

These challenges are not ranked.

• Premise 5 (High impact) — Many of the biggest financings are backed by long-term leases to cash-rich hyperscalers (Meta, Microsoft, Google, Oracle), often with guarantees or residual value support. Credit risk then sits with those firms, not the two labs, so lab unprofitability does not map onto loan repayment.
• Conclusion (High impact) — The housing comparison overstates the case. Housing recoveries in bust markets were poor, and the 2008 crisis came from leverage, securitization and correlated defaults across millions of loans. A few hundred billion dollars of data center debt is small against the financial system, and the telecom/fiber bust is a more apt analogy, painful for investors but not a banking crisis.
• Premise 4 (High impact) — Powered sites, grid interconnects, shells and land are scarce and reusable for general cloud or colocation, and GPUs can be redeployed to other workloads. Recoveries on the real-estate collateral could be good, especially while power is the binding constraint. Specialization applies mostly to chips, which are typically financed separately.
• Premise 3 (Medium impact) — Concentrated loans are underwritten and monitored loan by loan, are structured with equity cushions and covenants, and can be restructured through negotiation with a single borrower. Any 'first loss' falls on sponsors and equity before lenders. Concentration is therefore not clearly worse than diffuse, opaque mortgage pools.
• Conclusion (Medium impact) — Direct bank exposure is under 1% of assets, and private credit funds are largely locked-up, long-horizon vehicles whose investors can absorb losses. Insurers and pensions typically hold senior, investment-grade tranches. Losses may therefore stay contained instead of reaching ordinary savers.
• Conclusion (Medium impact) — AI compute demand may continue to grow and diversify (enterprise, sovereign buyers, other labs), so the shortfall scenario may not occur or may be partial, and the argument never assigns it a probability.

Suggested Improvements

• Tenant and credit structure — Break the debt down by who is liable: hyperscaler-leased with guarantees, lab-dependent, and speculative developments. Report how much of the exposure actually depends on OpenAI or Anthropic. This would replace the 'two tenants' shorthand with a verifiable measure and show how much of the debt is truly at risk.
• Recovery comparison — Give loss-given-default figures for foreclosed housing in bust markets and for analogous specialized assets (telecom fiber, shipping, aircraft, merchant power, office CMBS), and estimate recoveries separately for powered shells and for GPUs. A quantified, apples-to-apples comparison would show whether recoveries are really worse than housing and by how much.
• Scale and systemic transmission — Set the $200B and $800B figures against the size of private credit, insurer and pension assets and bank capital. Trace how losses would flow through bank lending to private credit funds, and consider redemption and fire-sale dynamics. This would show whether the losses are merely painful for some investors or systemically important, and it would address the correlation point that matters more than loan count.
• Scenario probability and magnitude — Define 'demand falls short' with thresholds (utilization, rental prices, lab funding access) and run graded scenarios with the debt-service coverage that results. A graded, falsifiable scenario would let readers tell a mild repricing from a collapse and make the claim testable.
• Sourcing — Replace the podcast-sourced aggregates with primary data (regulatory financial stability reports, rating agency data, loan-level deal documents) and define what counts as 'AI-related'. Unverified headline numbers carry much of the argument's weight, so firmer sourcing would materially raise its reliability.

Scenario Tests

• The labs' funding window closes and their compute commitments are cut sharply, while hyperscalers pull back too. (Supports) — A correlated demand shock would strain many loans at once, and distressed supply would push down resale and re-lease values. This is the scenario in which concentrated, lumpy losses and below-expected recoveries are most plausible.
• Lab growth slows, but Meta, Microsoft, Google and Oracle honor long-term leases or guarantees on most large financings. (Challenges) — Loan servicing would continue, with losses falling on equity holders and weaker developers. The 'two tenants' framing and the idea that lender losses will concentrate in banks and pensions would not hold.
• Power-ready sites stay scarce, and distressed shells and interconnects are bought quickly for general cloud and colocation use. (Challenges) — Recoveries on the real-estate collateral could match or exceed other commercial property, which undercuts the specialized-asset argument. Losses would be concentrated on GPU-related collateral.
• A telecom-style overbuild bust in which lenders are repaid late and at a discount, and assets are later reused. (Neutral) — Investors lose money and specialized assets sell at deep discounts, which partly supports the argument. But the later reuse and the lack of a banking crisis suggest losses could be painful without matching the housing crash in systemic impact.
• Private credit vehicles face redemption pressure and bank credit lines to them are drawn during an AI downturn. (Supports) — Opacity and bank-to-fund linkages could turn contained losses into a confidence problem. The risk then comes from uncertainty and correlation, not from the concentration the argument emphasizes.

Coherence & Relevance

The argument hangs together as a conditional, inductive risk narrative: large loans, correlated demand risk, specialized collateral and opaque holders make a downturn painful for lenders and their investors. Its weak points are the missing links from demand shortfall to default and loss, the unquantified housing benchmark, and the 'two tenants' framing, which its own assumptions weaken. A defensible reading is that a real AI demand shortfall could produce lumpy and hard-to-observe losses in private credit, with recoveries that are uncertain and possibly weak. It does not establish a housing-scale crisis or recoveries clearly worse than housing.

• Private credit already holds over $200 billion of AI-related loans and is projected to supply around $800 billion of data center financing through 2028, with insurers and pension funds among the ultimate investors. (Moderate) — Establishes that exposure is large and who ultimately holds it, but not the probability of default or the size relative to investors' portfolios. The forecast is not a measure of current exposure.
• Single data center financings now run from billions to tens of billions of dollars each, such as $18 billion of bank loans for New Mexico's Project Jupiter, $14 billion of notes for Oracle's Michigan campus, and about $27 billion of debt for Meta's Hyperion vehicle. (Moderate) — Supports the claim that deals are lumpy, but cherry-picked flagship deals with different tenant credit (Meta in particular) do not show that a typical loan is fragile.
• Because the exposure is concentrated in a smaller number of very large loans, a few failures can produce large losses, unlike a mortgage book spread across millions of loans. (Moderate) — Bears on the distribution of losses if failures occur, not on whether they occur or exceed housing-scale losses. It overlooks correlation, which drove the 2008 outcome.
• A foreclosed house can still be lived in and sold, but an AI data center is built for specialized chips, power, and cooling, and its value depends on finding tenants willing to pay for AI compute. (Strong) — Directly supports the recovery claim, but without recovery-rate data, and by treating building and chips as one asset. The stated residual-value assumption limits how far it goes.
• OpenAI and Anthropic account for a large and growing share of AI compute demand, and a rented GPU fleet loses money quickly when underused, so the tenants behind much of this debt are concentrated in two unprofitable firms. (Moderate) — Moves from demand share to debt backing and from unprofitability to inability to pay. The hyperscaler credit support and guarantees behind many loans are not examined.
• Private credit discloses less than banks and public markets, so the full size and location of the exposure is hard to measure in advance. (Weak) — Functions as a caveat on measurement, not support for a larger loss. Uncertainty cuts both ways.

View this argument on LogicFirst.ai