Vault
research

AI Circular Financing and Banking Exposure: How Close Is a Banking Issue?

Created

AI Circular Financing and Banking Exposure: How Close Is a Banking Issue?

Related: ai-token-economics-and-open-source-competition, hormuz-to-ai-repricing-causal-chain, the-efficiency-counterthesis, ai-crash-portfolio-defense, ai-infrastructure-endgame-indicators, anthropic-unit-economics-and-the-power-user-loss, anthropic-subsidy-stress-test, why-the-market-refuses-to-crash, japan-debt-trap-thesis-audit Builds-on: ai-infrastructure-endgame-indicators, anthropic-subsidy-stress-test

The Question Under Audit

Two video theses, distilled:

Zitron (bear, structural): AI is a circular-financing scam. Hyperscalers fund OpenAI/Anthropic, who spend the money back on hyperscaler compute. Demand outside this loop is minimal. Data centers are announced not built. Nvidia is the sole beneficiary. Three Horsemen will trigger the unwind: a data center under construction is canceled, an announced one is canceled, or an operational one fails to be profitable.

Contrarian bull/bear (balanced): Bull case is real (earnings, contracts, supply-constrained demand) but bear case wins on margin: capex is depreciation-heavy not capacity-additive, growth is narrow, equity risk premium is collapsed, open-source plus efficiency gains undermine the buildout's value proposition.

The user's actual question: How close are we to a banking issue from the money going around in the AI circuit?

This audit treats the equity-correction question and the banking-system-stability question as separate. The videos blur them. They are connected but operate through different transmission channels and have different probabilities. This doc focuses on the banking-stability question, with the empirical audit of the videos serving as substrate.

Empirical Audit — What the Numbers Actually Show

Circular financing: real but mismeasured

Zitron's claim: Hyperscalers invest in OpenAI/Anthropic → labs spend it on hyperscaler compute → revenue inflates without genuine demand.

Verified in mechanism, contested in scale:

The mechanism is real and well-documented. Microsoft's $13B+ in OpenAI is structured partly as Azure compute credits. Amazon's $40B Anthropic commitment runs partly through AWS. Google's $4B+ Anthropic investment includes GCP usage. Nvidia's announced $100B OpenAI investment (now "not in the cards" per Jensen Huang, March 2026) was conditional on OpenAI compute purchases.

The numbers Zitron cites:

These are real disclosed figures. Where Zitron's framing is incomplete: hyperscaler capex is not purely recycled lab money. Microsoft, Amazon, Google, and Meta have substantial external enterprise customer bases for their cloud services. Azure, AWS, and GCP run real customer workloads beyond AI labs. The labs are a meaningful but not exclusive demand source.

The honest read: circular financing is a measurable component of the system, not the whole system. The question is what fraction of the AI-related revenue and capex would survive if the OpenAI/Anthropic node were removed. Plausible answer: ~40–60% of AI-specific demand is genuinely external; ~40–60% is recycled. Zitron treats it as ~95%+ recycled, which is overstated.

Data center "ghost town": substantially confirmed

Zitron's claim: Buildout is hundreds of megawatts, not gigawatts. Many announced data centers are not actually opening.

Largely confirmed by independent data:

Sightline Climate / industry tracking shows:

This is a meaningful gap between announced and operational. Zitron's framing is closer to right than the hyperscaler narrative implies. The reasons are mostly mundane (power-grid interconnect queue, water access, permitting, supply chain on transformers/switchgear, GPU delivery timing) rather than malicious, but the practical effect is the same: announced capacity is not deliverable capacity.

This matters for banking: data center debt instruments are often structured against expected operational capacity and revenue. If the gap between announcement and operation is structural, debt service projections are systematically optimistic.

Nvidia dependency: high concentration, confirmed

Zitron's claim: AI boom is essentially a Nvidia revenue model. If buildout stalls or debt dries up, Nvidia collapses and takes the market with it.

Concentration is real, dependency is somewhat overstated:

Nvidia FY2026 Q3 disclosure:

This is high concentration by any measure. Compare: Apple, often cited for concentration risk in components, has top-customer concentration nowhere near this level. The customer-concentration risk is substantively present.

But Zitron overstates the binary: "Nvidia depends on this" is true; "if data center construction stalls, Nvidia collapses" is too strong. Nvidia could see revenue compression of 30–50% in a sharp pullback while remaining a profitable, cash-generating business. The question is the path of equity multiple compression, not corporate viability.

Equity risk premium: bear video is wrong

Bear video claim: ERP is near zero or negative; investors not compensated for stock risk vs. bonds.

Empirically false as stated:

Per Damodaran's January 2026 update:

The 4.23% ERP is almost exactly the 1960–2025 average. It's lower than the post-2008 average (which was unusually high due to crisis-era risk pricing) but it is not "near zero or negative." That claim was true in late 1999 and briefly in 2021. It is not true in early 2026.

The bear video's framing is right that valuations are not cheap and the ERP is below recent norms. But "negative ERP" is a specific claim that fails the empirical test. This matters for banking because the ERP is one of the inputs that feeds into how investors assess credit-equity relative value, and a negative ERP would be a flashing signal of mispricing. A 4.23% ERP is not.

Three Horsemen: structurally plausible triggers

Zitron's predicted unwind triggers — (1) in-progress data center canceled, (2) announced data center canceled, (3) operational data center fails to be profitable — are all observable and plausible.

(1) and (2) are happening at the rate of 30–50% of pipeline annually per Sightline. The market has not repriced because each individual cancellation is treated as project-specific, not systemic.

(3) has not yet visibly happened to a fully operational hyperscale-class facility. The closest analogue: some smaller AI-cloud players (smaller than CoreWeave) showing margin compression and customer churn. Watching whether a Tier-1 facility (Stargate, CoreWeave deployments, or hyperscaler-owned) reports operational losses would be the strongest signal.

The Actual Banking Transmission Map

This is where the user's question concentrates. The videos discuss "the market" generally; the banking-specific transmission requires a different analytical layer.

Where AI debt actually sits

Holder Approximate exposure Risk profile
Hyperscaler bonds (investment grade) ~$400B issuance projected for 2026, ~10x 2024 AAA–A1 rated; deeply liquid; broadly held by pensions, insurers, sovereign funds, banks. Loss probability low; mark-to-market risk real.
Data center ABS $30–40B annual issuance projected 2026/27; ~7–10% of total ABS Newer asset class; structures vary; held by insurers and credit funds. Loss potential meaningful in a buildout pullback.
Private credit (direct AI infrastructure lending) ~$800B of $1.5T projected external funding gap Held by private credit funds; not deposit-funded. Investor losses possible; systemic transmission limited.
Mid-tier developer debt (CoreWeave-type) CoreWeave alone ~$25B at Q1 2026, projected $38B end-2026; DDTL rates 11–15% High leverage, customer concentration (Microsoft 67% of FY2025 CoreWeave revenue). Vendor-guaranteed (Nvidia is willing to guarantee lease payments). High individual-firm risk.
Vendor financing (Nvidia → developers) Material but undisclosed full size Nvidia $2B CoreWeave equity Jan 2026; lease guarantees; failed $100B OpenAI deal. Nvidia balance sheet stress in tail scenarios.
Bank direct lending to AI infrastructure Smaller share of buildout funding than private credit Investment-grade syndications dominate; bank exposure is mostly to hyperscalers via bonds, not project finance to riskier developers.
Bank exposure to private credit funds (warehouse, fund finance) Underdisclosed The most interesting indirect channel. Banks lend to private credit funds; if those funds take losses, the warehouse lines and fund-finance exposure transmit.
Money market / repo exposure Limited direct AI exposure Money funds hold hyperscaler commercial paper but not AI-specific instruments.

Why this is not 2008-shaped

Several structural features distinguish current AI credit from the 2008 mortgage complex:

  1. Hyperscalers are cash-flow positive operating businesses. Microsoft, Amazon, Google, Meta have hundreds of billions in operating cash flow. Their bonds are not single-purpose vehicles; they are claims on the operating businesses. Compare to MBS which were claims on specific mortgage pools whose collateral could and did fail.
  2. No retail-funded transmission. AI debt is held overwhelmingly by sophisticated investors. There is no equivalent of mortgage securitization → retail money market funds → bank runs that defined 2008.
  3. Bank capital ratios are higher. Post-Dodd-Frank Tier 1 capital ratios for major US banks are roughly double 2007 levels. Stress-test thresholds incorporate severe market shocks.
  4. AI credit is concentrated in private markets, not bank balance sheets. This is a feature for systemic stability (banks insulated) but a cost for transparency (regulators see less).
  5. No mortgage-style underwriting deterioration. Hyperscaler bonds are issued by investment-grade entities with audited financials. Data center ABS structures have meaningful overcollateralization and reserve requirements. Underwriting standards have not deteriorated to subprime levels.

Where the risk is real

Several channels could transmit AI sector stress to broader credit and banking:

  1. Private credit cascade. Private credit funds with concentrated AI infrastructure exposure could see significant losses (5–15% in worst sleeves). Their LPs (pensions, insurers) would absorb losses. Bank exposure to private credit funds via warehouse lines and fund finance becomes material if cascading defaults force fund liquidations.
  2. Mid-tier developer failures. CoreWeave-type entities are highly leveraged with concentrated customer bases. A CoreWeave-style failure (low-probability, but not zero) would impair private credit holdings, ABS structures, and equity values. Possible secondary effects on Nvidia (vendor financing exposure).
  3. Vendor financing chain breakage. If Nvidia pulls back from guaranteeing developer leases (under stress on its own balance sheet), the financing chain that supports the buildout cracks. This is a chain-reaction risk distinct from individual-borrower risk.
  4. Energy infrastructure spillover. Utilities and energy developers are extending grid investments based on projected AI load. If load projections soften, stranded utility capex becomes a credit issue for energy debt held by banks and insurers. The IEEE / utility-sector exposure is material in specific regions (Texas, Virginia, Arizona).
  5. CRE / data center real estate. Some banks hold data center real estate directly or as collateral. If data center economics compress (lower rent rolls, vacancy, technology obsolescence), this turns into bank credit losses. Regional banks with data-center-heavy markets are the watch list.
  6. Hyperscaler equity correction → capex pullback → cascade. If the hyperscalers themselves cut capex sharply (in response to equity multiple compression or earnings pressure), the financing infrastructure built around their continuing spend faces stranded-asset risk. This is the most plausible transmission mechanism for a meaningful credit event.

Calibration: How Close Is a Banking Issue?

Base case (50–60% probability): Equity correction without banking crisis. NVDA pulls back 30–50%, broader equities -15–25%, AI-specific names harder. Private credit takes losses (5–15% in concentrated sleeves). A few mid-tier AI infrastructure failures. Hyperscaler write-downs (Microsoft impairs OpenAI value, etc.). No bank failures. Some pension/insurer loss absorption.

Moderate stress (25–35% probability): Equity correction plus credit-market stress. Several mid-tier failures (CoreWeave-class), private credit fund liquidations, data center ABS spread widening. Banks see indirect losses through fund-finance and warehouse-line exposure to private credit. Regional banks with concentrated CRE/data center exposure see localized stress. No systemic banking crisis.

Severe (10–15% probability): Hyperscaler capex pullback triggers cascading credit revaluation. Nvidia balance sheet stress. Vendor-financing chain breaks. Private credit fund failures. Bank losses on fund finance and direct lending. Possible regional bank failures (similar to SVB pattern but data-center-flavored). Money market / commercial paper pricing dislocation. Still not 2008-scale because the underlying borrowers are mostly investment-grade and the asset class is mostly held in non-deposit vehicles.

Tail (3–7% probability): Combined trigger — Hormuz energy shock + AI capex pullback + Japan trap + recession — produces broader financial stress. AI is a contributing factor, not the cause. Banking issues come from the broader macro stress, with AI credit as one transmission channel.

The current evidence does not support placing high probability on the severe or tail scenarios. The system is structured for equity-correction transmission, not banking crisis transmission. The honest framing: AI is a likely equity correction event in 2026–2027 with private credit losses as a real secondary effect. A banking-system crisis from AI requires either (1) much larger concentrated bank exposure than is currently documented, or (2) a combined-trigger scenario where AI is contributory, not causal.

The IMF's April–May 2026 financial-stability warnings about AI are about AI-powered cyberattacks on financial infrastructure, not about AI-investment debt as systemic risk. Distinct concern, sometimes conflated. The BIS has not flagged AI infrastructure debt as systemic. This is consistent with the calibrated assessment above.

What Would Move the Needle

Specific signals that would shift the probability toward banking-issue territory:

None of these signals are currently active. Most are watch-list items, not trigger events.

Connection to Existing Vault Theses

This audit is the load-bearing complement to several earlier docs.

anthropic-subsidy-stress-test mapped the implicit Trainium/TPU subsidy at $2–4B/year and traced the dissipation timeline. The current audit extends that analysis from "what happens to Anthropic's economics" to "what happens to the bank credit infrastructure when the subsidies normalize." The subsidy's eventual normalization would not by itself trigger banking issues; it would trigger Anthropic margin compression, which the equity market would absorb.

ai-infrastructure-endgame-indicators identified four archetypes for how the AI buildout absorbs debt: Japan-style slow deflation, ratepayer socialization, efficiency cliff, sovereign absorption. The base case is Japan-style + partial ratepayer. The current audit confirms that base case structurally — the credit losses are real but distributed across non-bank holders (private credit, insurers, pensions, sovereigns), which is exactly the Japan-style pattern.

why-the-market-refuses-to-crash mapped Kevin Ting's structural-bid framework. Applied here: AI debt is largely held by structural-bid holders (insurers running ALM, pensions running asset-allocation, sovereigns running indefinite-horizon investment). These holders are price-insensitive in the short run. This is why AI credit spreads have not blown out despite the Zitron-style concerns — the marginal holder is not selling.

hormuz-to-ai-repricing-causal-chain identified the energy-driven AI repricing scenario. The current audit clarifies that the repricing channel runs through equity multiples and hyperscaler capex decisions, not directly through bank balance sheets. Banking-system stress from the Hormuz path requires the recession transmission, not the AI transmission.

the-efficiency-counterthesis identified efficiency gains (Vera Rubin 10x, etc.) as a path that could reduce the buildout's credit risk by lowering the capacity needed. This is the bear video's "Open-source AI plus inference efficiency undermines the buildout" point in different language. Both are real. The credit risk concentrates exactly to the extent that efficiency gains don't arrive on time.

japan-debt-trap-thesis-audit (just landed) is structurally parallel: a credit-system thesis where the trap framing has been right in pieces and wrong in timing for years. The lesson there applies here too: structural risk can be real and not trigger for extended periods, with the trigger when it arrives often coming from a different vector than the one anticipated.

Open Questions

Calibration Notes

Sources

Backlinks