ED ZITRON AND WILL GUYATT: Lenders keep funding AI debt because they follow one another and lack alternatives, so the turn will show when a debt sale fails to place

The Gist

Ed Zitron says banks keep lending to AI projects because everyone else is and the money has to go somewhere. He expects the turn to come when banks cannot sell a data center loan on to other investors. Some early signs are already visible, such as pickier banks, a stuck loan on Oracle's New Mexico project, and an Australian data center firm pulling its IPO. This packet steelmans the arguments Ed Zitron makes in this episode, in their best light, for analysis; it is not an endorsement of his conclusions.

Conclusion

Continued lending to AI data centers is weak evidence that the projects are sound, because banks and private credit funds face pressure to deploy capital, are drawn by the promise of factory-like returns, sometimes commit with little diligence, and follow one another; lending is likely to continue until a visible failure, most plausibly a data center or neocloud debt sale that banks cannot fully distribute and must hold or sell at a discount, and early signs of that turn have appeared.

Premises

1. Banks and private credit funds hold large pools of capital seeking yield, and AI infrastructure is presented as a factory-like asset with steady payments.
2. Fast commitments suggest thin diligence; Blue Owl executives reportedly agreed to invest up to $10 billion in future projects with Primary Digital Infrastructure within 15 minutes of their first in-person meeting.
3. Banks that arrange data center loans typically syndicate most of the debt to other investors, so their willingness to lend depends on expecting to sell it on.
4. Some major lenders, including Sumitomo Mitsui Banking Corp., Mitsubishi UFJ Financial Group, and Societe Generale, were reported in October 2026 to be more selective about data center loans, and a Meta-linked CleanSpark bond required concessions including a 98.5 issue price and a 7.875 percent coupon.
5. The $18 billion of Project Jupiter construction loans has not been fully distributed and was quoted at about 89 to 91 cents and later about 84 cents on the dollar.
6. Australian neocloud Firmus shelved its planned IPO in October 2026 for lack of demand and turned to a private raise.
7. When an arranging bank is stuck holding debt it meant to sell, other lenders become more cautious, which can tighten credit across the sector.

Assumptions

• 'More selective' does not mean these banks have stopped lending; the CleanSpark bond drew more than $10 billion of orders.
• In the Paramount financing the banks completed the $52 billion debt sale, and investors rather than the banks took the initial losses when the bonds fell to about 95 to 96 cents, so the comparison illustrates syndication risk rather than a deal the banks could not sell.
• Banks' direct exposure to AI-related lending is estimated at under 1 percent of assets, with most risk held by private credit funds and the insurers and pension funds that invest in them.
• The Blue Owl account describes a framework commitment to future projects, not a single funded loan.

Analysis

Argument type: Inductive.

Potential Fallacies

• Hasty generalization / anecdotal evidence (P2 and the 'sometimes commit with little diligence' clause of the conclusion) — A single reported 15-minute meeting is used to suggest thin diligence among lenders generally. Even granting the fact, speed of a final meeting is a poor proxy for the total diligence done, and the commitment is a framework for future projects rather than a funded loan. The conclusion's hedge ('sometimes') limits but does not remove the problem.
• Confirmation bias / selective evidence (P4–P6 and the 'early signs of that turn have appeared' clause) — Stress indicators (Jupiter discount, CleanSpark concessions, Firmus, bank selectivity) are assembled into a pattern, while counter-signals (over $10 billion of CleanSpark orders, the completed Paramount sale, continued lending) are treated as qualifiers. Without base rates for how often healthy credit markets show such events, ambiguous signals are easily read as confirming the thesis.
• Unfalsifiable or flexible prediction (Conclusion: 'lending is likely to continue until a visible failure') — Continued lending is explained as herding, while pullbacks and discounts are read as the turn arriving. With no time horizon or defined threshold for 'visible failure', almost any outcome fits. Jupiter already looks like a partial failure to place, yet lending continues, which blurs the line between a precursor and the trigger itself.
• Non sequitur (mechanism to prediction) (Inference from P1–P3, P7 to the forecast in the conclusion) — P1–P3 and P7 explain why lenders might keep lending and how a stuck deal could spread caution, but they do not establish that lending persists until a specific trigger, that the trigger will be a failed syndication, or that the contagion will actually occur. P7 is stated only as a possibility ('can tighten credit').
• Equivocation across event types (P4–P6 with A2, as support for 'early signs') — Jupiter (partly undistributed loans), CleanSpark (a bond that cleared with concessions), Firmus (an equity IPO withdrawn) and Paramount (a completed sale) are different kinds of events, but are treated as variants of one signal. Only some involve an arranger holding unsold debt.
• Inferring unsoundness from pricing or placement difficulty (P5 and the conclusion) — Discounts and incomplete distribution can reflect deal size, rates, supply overhang or thin liquidity in loan quotes rather than poor underlying project quality. The argument needs a bridge from market-clearing difficulty to the claim about project soundness.

Counterarguments

These challenges are not ranked.

• Conclusion / Premise 1 (High impact) — Much data center debt is backed by long-dated leases with creditworthy tenants such as Meta, so lenders may be correctly pricing contracted cash flows rather than herding. If so, convergence among lenders reflects independent, accurate assessments and continued lending is meaningful evidence of soundness. The herding explanation fits the same facts without being shown to be better.
• Premise 2 (Medium impact) — A 15-minute meeting for a framework commitment reveals nothing about total diligence. Months of prior work, repeat relationships and project-level conditions can precede a short final meeting, so speed is not evidence of thinness.
• Premises 4–6 and the 'early signs' claim (High impact) — Concessions, selective banks and partially placed deals are what price discovery looks like in a heavy-issuance market. The over-$10 billion CleanSpark order book and the completed Paramount sale show market depth. Jupiter's size and structure may make it idiosyncratic, and Firmus is an equity event.
• Premise 7 and the forecast (High impact) — With bank exposure under 1% of assets and risk held by private credit, insurers and pensions, a stuck sale may be absorbed at higher yields or by new buyers without sector-wide tightening. Banks can also hold to maturity or use risk-transfer tools, which dampens forced-selling cascades.
• Conclusion (trigger identification) (Medium impact) — The turn could arrive through other channels: a hyperscaler capex cut, tenant or offtaker problems, GPU obsolescence, power constraints or a funding-cost shock. Stress in opaque private credit marks may also surface slowly rather than as a single visible failed sale, so the named trigger may miss the real turning point.
• Conclusion (overall) (Medium impact) — Soundness depends on cash flows and AI demand, not lender behavior. If AI revenue and tenant payments hold up, the herd will have been right and the thesis fails without any failed placement.

Suggested Improvements

• Falsifiability and timing — Define the 'turn' and the failure threshold in advance, for example a number of large deals hung or priced below par by a stated margin within a stated window, and state what observations would count against the thesis. Without a threshold or horizon, both continued lending and tightening can be absorbed into the narrative, which makes the prediction hard to test or act on.
• Distinguish credit types — Separate hyperscaler-leased data center loans from speculative neocloud and GPU-backed borrowing, and analyze how tenant credit and contract structure affect lender behavior and risk. The herd claim is much stronger for credits that depend on refinancing and GPU demand than for contracted investment-grade leases, and aggregation overstates the effect.
• Base rates and systematic data — Provide data on how often AI data center deals clear at par versus with concessions, how often discounted syndications in other cycles preceded a turn, and how Jupiter-type deals compare with prior hung deals. A handful of salient cases cannot establish a trend without knowing how frequently healthy markets produce such events.
• Bridge from banks to the real risk holders — Trace how stress at arranging banks transmits to private credit funds, insurers and pensions, including bank lending to funds, back-leverage and opaque marks. The contagion mechanism runs through banks while the stated exposure sits elsewhere, and this gap is unaddressed.
• Strengthen the diligence evidence — Replace or supplement the Blue Owl anecdote with evidence on covenant strength, underwriting terms, or post-deal performance across lenders. Speed of a commitment is a weak proxy for diligence, and direct evidence would support the 'thin diligence' claim far better.
• Engage the strongest opposing view — Address the case that contracted cash flows and tenant credit explain lending, and explain why herding is a better account than informed underwriting. This would move the argument from a framing that is hard to disconfirm to a comparison of competing explanations.

Scenario Tests

• A further large data center or neocloud syndication fails, and the arranging banks must hold the loans or sell at a deep discount. Other arrangers visibly cut commitments and spreads widen across the sector. (Supports) — This would confirm the predicted trigger and the P7 contagion mechanism, and would suggest lending had been herd-driven rather than discriminating.
• Jupiter-type discounts are absorbed by private credit, insurers or sovereign funds at higher yields, and new deals continue to clear with larger concessions. (Challenges) — The 'visible failure' occurs but produces repricing rather than a sector-wide turn, in line with the Paramount precedent. Lending would continue, and the thesis would be reduced to a claim about higher prices, not a break.
• Hyperscaler tenants keep paying, AI revenue grows, and loans perform while order books stay strong. (Challenges) — Lender convergence would turn out to reflect real fundamentals. Continued lending would have been better evidence of soundness than the argument allows.
• Stress emerges first through a hyperscaler capex cut, tenant default or GPU rental price collapse, with no prior failed syndication. (Challenges) — The identified indicator would miss the actual turn. The broader point about herding may hold, but the specific early-warning signal would be unreliable.
• Private credit funds holding AI-related loans face redemption pressure or gate withdrawals as marks fall, while banks remain largely unscathed. (Neutral) — This would confirm the stress and the assumption that risk sits outside banks, but it would not match the 'debt sale banks cannot place' trigger. The thesis would be partly right about the outcome and wrong about the transmission.
• Banks that were more selective, such as the Japanese megabanks, later return to lend at the same terms, and the Jupiter loans recover toward par. (Challenges) — The early signs would have been temporary price discovery and idiosyncratic factors, not the start of a turn.

Coherence & Relevance

The argument is coherent as a cumulative inductive case. Its modest core claim, that continued lending is weak evidence of soundness, is well supported by the structural premises and the candid assumptions. The forecast and diagnostic parts are less secure. They depend on unstated links: that herding persists until a salient trigger, that a failed syndication is the likeliest trigger, and that the reported events are leading indicators rather than routine market noise. The stated assumptions act as built-in defeaters (continued lending, a completed Paramount sale, low direct bank exposure, a framework-only Blue Owl commitment), which keep the thesis hedged but also leave tension between the 'weak evidence' framing and stress signals that themselves show lenders discriminating. The argument's strongest datum is Jupiter; its weakest elements are the Blue Owl anecdote and the Firmus event. It would be more persuasive with defined indicators, base rates and treatment of contract-backed credits.

• Banks and private credit funds hold large pools of capital seeking yield, and AI infrastructure is presented as a factory-like asset with steady payments. (Strong) — Explains why lending continues, but does not distinguish herding from well-founded lending, so its link to 'weak evidence of soundness' relies on unstated assumptions about lender quality.
• Fast commitments suggest thin diligence; Blue Owl executives reportedly agreed to invest up to $10 billion in future projects with Primary Digital Infrastructure within 15 minutes of their first in-person meeting. (Moderate) — Speed is a poor proxy for diligence, and the framework commitment is not a funded loan. A single anecdote does not establish a sector-wide pattern.
• Banks that arrange data center loans typically syndicate most of the debt to other investors, so their willingness to lend depends on expecting to sell it on. (Strong) — Establishes why distribution is the pressure point, but not that distribution will fail or that end investors are less discerning than arrangers.
• Some major lenders, including Sumitomo Mitsui Banking Corp., Mitsubishi UFJ Financial Group, and Societe Generale, were reported in October 2026 to be more selective about data center loans, and a Meta-linked CleanSpark bond required concessions including a 98.5 issue price and a 7.875 percent coupon. (Moderate) — Motives for selectivity are unstated, and a strong order book and continued lending equally support a normal-repricing reading. The CleanSpark bond cleared, so it is not an instance of failure to place.
• The $18 billion of Project Jupiter construction loans has not been fully distributed and was quoted at about 89 to 91 cents and later about 84 cents on the dollar. (Strong) — A single case with thin quotes. It does not show whether the cause is sector-wide or deal-specific, and lending continued after it.
• Australian neocloud Firmus shelved its planned IPO in October 2026 for lack of demand and turned to a private raise. (Weak) — An equity-market event in another jurisdiction, not a debt placement. The attribution to 'lack of demand' and its relevance to bank syndication are not established.
• When an arranging bank is stuck holding debt it meant to sell, other lenders become more cautious, which can tighten credit across the sector. (Strong) — Offers the contagion mechanism but only as a possibility. Given the stated low direct bank exposure and the Paramount precedent, it does not show that a stuck deal would cause a sector-wide turn, nor how stress reaches private credit and its funders.

View this argument on LogicFirst.ai