Which institutions, credit structures, and counterparties are quietly absorbing the risk of AI capex, and what would a repricing event look like across chipmakers, hyperscalers, private credit, and equity markets?

Advocate A’s weakest point is the leap from “unconventional and opaque” to “systemically dangerous.” That inference is not established by the evidence cited. The Nvidia/OpenAI backstop story, even if reported accurately, shows price discovery and risk governance occurring before final commitment: the guarantee was reportedly scaled down, and Nvidia then announced a broader structure in which Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR would mobilize more than $500 billion of third-party capital through independent compute-financing platforms. That is not risk disappearing into a single vendor balance sheet; it is risk being syndicated to institutions whose business is underwriting long-duration infrastructure and credit exposure. Nvidia’s own language emphasizes independent underwriting and usage-linked revenue, not a blank guarantee of customer demand.

A’s “shadow credit” framing is also incomplete. Private credit, or nonbank direct lending, can be risky because it is less publicly transparent than syndicated bank credit. But the BIS description of AI data-center financing is not equivalent to a 2007-style hidden household leverage chain. BIS says these special-purpose vehicles are often backed by data-center assets, long-term leases or offtake agreements, and sometimes contractual guarantees from investment-grade hyperscalers. That means lenders are not simply lending against speculative model progress; they are lending against physical infrastructure and contracted cloud/compute cash flows. BIS does identify transmission channels, but its later bulletin on private-credit technology exposure notes low leverage and secured lending may limit spillovers. The relevant risk is repricing of infrastructure yield and hyperscaler equity multiples, not automatic systemic breakage.

The strongest affirmative evidence is Anthropic’s revenue base. Its May 2026 Series H valued the company at $965 billion post-money after run-rate revenue crossed $47 billion, implying roughly 20.5x run-rate revenue. That is aggressive, but not facially irrational for a frontier software platform with extraordinary growth, enterprise deployment, and cloud distribution across AWS, Google Cloud and Microsoft Azure. Anthropic also disclosed major infrastructure relationships with Amazon, Google/Broadcom and SpaceX, plus strategic memory partners, which supports the view that valuation reflects real enterprise and developer demand rather than purely circular financing.

The Jane Street example, as framed by A, actually cuts both ways. A private market-maker or trading firm taking a large AI-linked trading loss is evidence that sophisticated counterparties can absorb mark-to-market losses without immediately transmitting them through insured deposits or payment systems. Market-making firms routinely warehouse risk as the cost of providing liquidity; absent evidence of forced selling, unpaid margin, fund gating, or bank impairment, one large monthly loss is a stress event, not proof of systemic fragility.

Finally, Zuckerberg’s agent comments are a real caution but not a crux. An AI agent is software that autonomously performs multi-step tasks; Meta’s four-month disappointment does not prove enterprise AI demand is plateauing. The financing case should be judged by contracted compute usage, revenue retention, and credit performance, not by one CEO’s short-term product-timing miss.

A’s specific weakness is that he concedes the two facts most damaging to his thesis, then treats them as if they had no analytical consequence. He concedes that the Nvidia structure is no longer a bilateral vendor guarantee and that many AI data-center SPVs are backed by leases, offtake contracts, physical assets, and sometimes investment-grade hyperscaler support. That concession implies the core issue is not “hidden danger” but whether long-duration infrastructure lenders are mispricing contract, utilization, residual-value, and refinancing risk. A has not shown that they are. His rebuttal substitutes correlation vocabulary for loss-transmission evidence.

Concentration among Apollo, Blackstone, BlackRock, Brookfield, Goldman, KKR, PIMCO, and Blue Owl is not automatically systemic concentration. These are multi-strategy institutions managing hundreds of billions to over a trillion dollars each, and the Nvidia platform is explicitly designed to mobilize third-party capital through independently underwritten projects, not to warehouse all risk on Nvidia’s balance sheet. The fact that the same firms recur is what one would expect in infrastructure finance: the relevant expertise is scarce, and deal size filters out smaller lenders. The adverse case would require evidence that these exposures are large relative to each firm’s capital, cross-collateralized, or funded with runnable short-term liabilities. A does not provide that evidence.

A’s GPU-depreciation argument also overstates the collateral channel. In project finance, collateral matters, but the primary repayment source is usually contracted cash flow. BIS describes exactly that structure: long-term operating leases, capacity offtake agreements, asset backing, and contractual guarantees supporting debt held by private credit funds and institutional investors. If a hyperscaler or AI lab is paying under a take-or-pay or lease contract, the lender is underwriting counterparty credit and utilization economics, not simply guessing resale value for used H100s. Residual-value risk is real, but it is not equivalent to a mortgage CDO backed by millions of borrowers with correlated default risk.

CoreWeave’s spread widening is better evidence for B than A. A calls it stress; in credit markets, a 100–125 basis-point repricing plus maintenance covenants is the mechanism by which risk is rationed and transferred to parties demanding compensation. Covenants such as debt-service coverage ratios are not proof of fragility; they are lender protections that force earlier intervention before losses become uncontrolled.

Finally, A still has not connected Jane Street’s alleged trading loss or Zuckerberg’s agent lag to systemic credit impairment. A market-maker absorbing a large mark-to-market loss, without cited evidence of unpaid margin, bank losses, fund gating, or forced deleveraging, supports the proposition that sophisticated counterparties can absorb AI repricing. Meta’s four-month agent disappointment is execution-risk evidence at one company; it does not rebut Anthropic’s disclosed $47 billion run-rate revenue, enterprise adoption, and cloud distribution. The remaining evidence points to volatile but functioning risk markets, not hidden systemic breakage.

A’s strongest case is no longer a vague “bubble” claim; it is that private credit vehicles are showing actual liquidity pressure while hyperscalers are accumulating very large lease commitments outside conventional debt metrics. If BDC redemption caps, Blue Owl’s OBDC II restructuring, and Moody’s $662 billion of uncommenced hyperscaler leases were all tightly linked to AI data-center cash flows, A would have a plausible mechanism for hidden leverage to become visible through forced sales and credit losses.

My position remains stronger because A has not established that linkage. The core analytical error is treating three distinct things as one balance sheet: software-company private credit exposure, semi-liquid retail BDC redemption mechanics, and AI infrastructure project finance. The Federal Reserve’s May 2026 Financial Stability Report does say redemption requests rose sharply in some private credit vehicles, but it also reports that the largest perpetual BDCs had cash and bank credit sufficient to cover at least three quarters of redemptions at the 5% NAV level. That is consistent with liquidity rationing in semi-liquid funds, not a systemwide run. Blue Owl’s $1.4 billion sale is also weaker evidence than A implies because the reported sale price was 99.7% of par and the assets were senior secured loans diversified across industries; that looks more like planned liquidity management under an awkward retail wrapper than distressed liquidation of AI collateral.

A’s strongest Moody’s point similarly proves disclosure complexity, not insolvency. Signed-but-not-commenced leases are economically important and should be watched, but they are not identical to callable short-term debt. They become payment obligations as facilities are delivered, typically alongside physical data-center capacity and customer demand. A hyperscaler is a cloud-scale operator such as Amazon, Microsoft, Google, Meta, or Oracle; these are not thinly capitalized mortgage conduits. The right question is whether lease-adjusted debt, free cash flow, and contracted utilization deteriorate together. A has not shown that.

The BIS evidence cuts both ways, but ultimately favors B. BIS flags concentration and spread compression in private credit software lending, yet it also notes that low leverage and secured lending may limit spillovers. BIS’s separate work on AI infrastructure describes special-purpose vehicles supported by long-term leases, capacity offtake agreements, physical assets, and sometimes investment-grade guarantees. That is exactly why “shadow credit,” meaning nonbank or off-balance-sheet lending, should not be equated with hidden danger by definition. Private credit can misprice risk, but in this case the observed repricing—wider spreads, covenants, redemption caps, and syndicated platforms—is how risk is being rationed among parties paid to bear it.

Finally, Anthropic’s $965 billion valuation on more than $47 billion of run-rate revenue remains material evidence of real enterprise demand. A can argue the multiple is high, but not that it is valuation without revenue. The most likely repricing event is equity multiple compression and higher infrastructure yields, not cascading systemic breakage.