ED ZITRON AND DEAN BAKER: The sophistication and wealth of AI's backers is weak evidence that the build-out is sound, given fee incentives and past costly misjudgments

The Gist

People say the AI build-out must be sound because the smartest, richest firms are behind it. But banks get paid fees whether deals work out or not, and Meta lost around $80 billion on the metaverse. Smart money has been wrong before, including in the housing bubble. 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 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.

Premises

1. A common defense of the build-out is that the firms funding it are the best-informed and most experienced investors.
2. Banks earn arrangement and underwriting fees when deals close, such as the $15 billion revolving credit facility a large bank syndicate arranged for Anthropic, so their incentives can favor volume over long-run outcomes.
3. Meta's Reality Labs division accumulated roughly $77 billion of losses by late 2025, and Meta then cut metaverse spending, showing that a leading technology firm can misallocate tens of billions of dollars.
4. In the housing bubble, sophisticated institutions also funded loans that failed, so sophistication did not prevent large losses.
5. Reliance on the reputation of participants substitutes for examining whether the underlying cash flows can support the commitments.

Assumptions

• Sophisticated investors also price risk, as rising credit spreads on Oracle show, so markets are not ignoring the concerns entirely.
• Reality Labs losses have continued to grow, to roughly $88 billion after the second quarter of 2026, which is in line with the figure cited in the conversation.
• Fee incentives do not prove that lenders misjudge risk; they show only that deal volume and soundness can diverge.

Analysis

Argument type: Inductive.

Potential Fallacies

• Hasty generalization / unrepresentative examples (P3 and P4) — Meta's metaverse and the housing bubble show that sophisticated actors can err, so sophistication is not sufficient for soundness. Without a base rate that includes cases where informed capital got it right (cloud, mobile networks), they do not show how much evidential weight participation deserves. The hedged 'weak evidence' wording softens this, but the gap between 'not sufficient' and 'weak' remains.
• Weak or imperfect analogy (P3 and P4) — Meta's loss came from its own cash flow on a speculative product bet, while AI financing involves external creditors with contracts, collateral, and covenants. Housing losses depended on securitization and risk transfer, which may not map onto a revolving facility held on bank balance sheets. The cases support the general claim that large firms can err, but not that AI financing shares their structure.
• Incentive-based discounting (circumstantial motive) (P2) — Pointing to fee incentives gives a legitimate reason to discount the lenders' signal, but incentives alone do not show a signal is uninformative. The argument concedes this in A3. The effect also depends on how much risk arrangers retain and how large the fees are, which are not specified.
• Unsupported generalization / possible straw man (P1 and P5) — P5 asserts that reputation substitutes for cash-flow analysis without evidence that funders skip diligence or that defenders rely only on reputation. Defenders more plausibly cite revenue growth, contracted backlog, and capital at risk. A1 also shows that markets do some independent risk pricing. P5 describes a real heuristic, but it is not shown to be the operative one here.
• Aggregation of heterogeneous agents (Conclusion and P2) — Banks (fee-earning arrangers), asset managers, and hyperscalers (spending their own cash and equity) face different incentives. The fee argument applies most clearly to arrangers and much less to principals with skin in the game.

Counterarguments

These challenges are not ranked.

• Premise 2 (High impact) — Arrangers of revolving credit facilities usually hold commitments on their own books and face drawdown risk, capital requirements, reputational damage, and post-2008 risk-retention rules. 'Fees regardless of outcome' fits a pure originate-to-distribute model, which has not been shown for AI lending. Funders' private diligence, covenants, and collateral may also make their participation substantively informative.
• Premise 3 (High impact) — Meta spent its own cash flow under founder control on a speculative option bet, so its loss says little about arm's-length lenders. Meta's cut in spending can also be read as a sophisticated actor correcting its error. The relevant AI spending is mostly hyperscaler capex, which is neither fee-driven nor lender-funded.
• Conclusion (Medium impact) — Without a base rate, the reasoning can discredit any institutional consensus, including cases where informed capital was right. Large infrastructure build-outs by sophisticated funders (cloud, mobile, even fiber with delayed payoff) often proved productive. Participation at a price, such as credit spreads, is also more informative than mere participation, so the signal may be stronger than 'weak'.
• Premise 5 (Medium impact) — Funders do examine cash flows, and defenders point mainly to rapid revenue growth at leading AI labs and contracted compute backlog. Rebutting reputation-based reasoning does not touch this stronger case, and P5's logic would equally apply to readers deferring to the speakers' own reputations.
• Premise 4 (Medium impact) — Housing losses depended on opaque securitization, rating-agency failures, and regulatory arbitrage, and AI lending involves different collateral, counterparties, and visible revenue. The analogy may overstate the parallel.

Suggested Improvements

• Base rates — Add a reference class of large technology and infrastructure build-outs (railroads, fiber, cloud, mobile, housing) and estimate how often sophisticated financing preceded success versus failure. This would justify the move from 'sophistication is not sufficient' to 'it is only weak evidence', which the current examples alone cannot support.
• Risk retention and fee magnitude — Show how much exposure arrangers of AI facilities retain or syndicate, who ultimately holds the risk, and how fees compare to potential losses. Fee incentives matter only in proportion to risk transferred away, so this would turn an asserted mechanism into a demonstrated one.
• Differentiate backer types — Separate fee-earning arrangers from principals (hyperscalers, equity sponsors, asset managers) and address skin-in-the-game signals directly. It would answer the strongest objection, that the fee critique does not reach the largest and best-capitalized funders.
• Engage the fundamentals-based defense — Present AI revenue, backlog, customer concentration, utilization, and debt-service coverage alongside the critique of the smart-money heuristic. This would avoid the straw-man charge and provide the cash-flow test the argument itself calls for.
• Analogy fit — Make explicit which structural features of the metaverse and housing cases carry over to AI financing (leverage, opacity, circular vendor financing) and which do not. It would turn illustrative anecdotes into relevant precedents.

Scenario Tests

• AI lab revenues grow fast enough to comfortably service the debt and the build-out proves profitable. (Neutral) — The conclusion stays logically intact because it claims only that the funders' sophistication was weak evidence, not that the build-out would fail. It would show that the cash-flow evidence, not the backers' reputations, was what mattered, but critics could use the outcome to portray the argument as mistaken skepticism.
• A wave of AI-related defaults occurs, and it emerges that arrangers sold down nearly all of their exposure to pension funds, insurers, and retail vehicles. (Supports) — This would match the housing pattern of fee-driven volume with risk passed to others, vindicating P2 and strengthening the discount on lender participation.
• Banks are shown to hold most of their AI revolvers on balance sheet, with strict covenants and collateral, and credit spreads widen sharply on weaker borrowers. (Challenges) — The fee-regardless-of-outcome mechanism would lose force, and priced, risk-bearing participation would look more informative than 'weak'. Spread widening would show that markets are actively scrutinizing the build-out, consistent with A1.
• The build-out resembles the fiber overbuild: investors lose money, but the infrastructure proves durable and socially valuable. (Supports) — Sophisticated funding would have coincided with financial losses even though the technology was sound, which supports the point that backers' sophistication does not guarantee returns. It also shows that 'sound' is ambiguous between technological, equity, and lender outcomes.

Coherence & Relevance

The argument is coherent as a defeater-style inductive case. P2 weakens the signal value of lender participation, P3 and P4 show that sophistication has not been sufficient to prevent large losses, and P5 redirects attention to fundamentals. A1 and A3 limit the claim, and the conclusion is hedged to match. The remaining gap is between 'sophistication is not sufficient' and 'sophistication is weak evidence', which needs base-rate or comparative support. The analogies are also looser than the argument implies, and the fee critique does not reach all types of backers. As a rebuttal to a simple appeal to the smart money, the argument is reasonable. As a basis for judging whether the build-out is sound, it is incomplete without revenue and debt-service analysis.

• A common defense of the build-out is that the firms funding it are the best-informed and most experienced investors. (Strong) — It defines the claim being rebutted, but it is a thin version of the defense and no proponent is identified. Fundamentals-based defenses are not addressed.
• Banks earn arrangement and underwriting fees when deals close, such as the $15 billion revolving credit facility a large bank syndicate arranged for Anthropic, so their incentives can favor volume over long-run outcomes. (Strong) — It undercuts the reliability of lender participation as a signal, but only in proportion to risk retained. It applies to fee-earning arrangers, not to hyperscalers or equity sponsors.
• Meta's Reality Labs division accumulated roughly $77 billion of losses by late 2025, and Meta then cut metaverse spending, showing that a leading technology firm can misallocate tens of billions of dollars. (Moderate) — It shows that large, sophisticated firms can err, but its relevance to lenders is limited because it was self-funded corporate spending. No base rate is offered.
• In the housing bubble, sophisticated institutions also funded loans that failed, so sophistication did not prevent large losses. (Moderate) — Its incentive parallel is useful, but the structural differences (securitization, ratings, risk transfer) are not examined and no successful counter-cases are weighed.
• Reliance on the reputation of participants substitutes for examining whether the underlying cash flows can support the commitments. (Moderate) — It points to the right evidence (cash flows) but asserts rather than shows that reputation-reliance is occurring, and it offers no cash-flow analysis of its own.

View this argument on LogicFirst.ai