Hook
A number is circulating from the Bank for International Settlements: 55.2% of AI funding comes from other AI firms. The coverage framed it as a warning about financial stability. Almost none of that coverage asked the first question an auditor asks.
What is the denominator?
This is not pedantry. 55.2% of global AI venture capital, 55.2% of total AI capital expenditure, and 55.2% of a curated sample of disclosed transactions are three distinct claims with three distinct risk profiles. The reporting treated them as interchangeable. That is the first structural failure, and it occurs before a single AI company is named.
I do not trust the pitch; I audit the structure. The pitch is a percentage. The structure is a closed loop. I have spent my working life dismantling closed loops โ in ICO token distributions, in DeFi yield farms, in NFT rarity calculators. The mechanism does not change. Capital enters. It is confirmed as revenue. It re-enters as capital. The loop reads as growth until the quarter it reads as contraction. A circuit has no external boundary until it breaks, and the boundary is the only thing that tells you whether you are looking at a market or a mirror.
Context
The BIS is not a trade publication. It is the central bank of central banks, the institution that coordinates prudential standards across national regulators. When it names a financing structure, it is not doing journalism. It is filing a pre-regulatory notice. The historical record is consistent. BIS commentary on shadow banking preceded disclosure regimes. Commentary on crypto-asset exposure preceded bank capital treatment. Commentary on private credit preceded stress-testing frameworks. A central bank does not publish a figure like 55.2% to inform. It publishes to prepare.
The AI capital stack the figure describes has three layers. Compute producers: GPU designers, foundries, high-bandwidth memory suppliers, data-center operators, power utilities. Cloud providers who rent that compute. Model laboratories who consume it. In a functioning market, money flows from end users upward through this stack, and each layer earns its margin from a distinct payer with a distinct budget.
In the structure the BIS is flagging, money flows sideways. Compute producers invest in model labs. Model labs commit to purchase compute. Cloud providers invest in model labs. Model labs commit to consume cloud. Each arrow is individually defensible. Collectively they form a circuit.
The arithmetic of a circuit is different from the arithmetic of a market. A market has an external boundary: end users. A circuit has no external boundary until the loop is broken. This distinction is the entire substance of the BIS warning, and it is the distinction the coverage erased.
There is a second erasure, and it is more technical. The word "funding" collapses three instruments with three different economic meanings. Equity absorbs loss; it is patient capital that can be marked to zero without triggering default. Debt does not absorb loss; it transmits loss to a creditor and can trigger a cascade. A multi-year procurement commitment is neither โ it is a forward obligation that appears as backlog on the vendor's books and as a contingent liability on the buyer's, and it is only as good as the buyer's future ability to raise the cash to draw it. If the 55.2% figure mixes all three, then it is not a measure of financing at all. It is a measure of a category error. The instrument determines the transmission channel, and a number that hides the instrument hides the channel.
To be precise about my epistemic position: the source material is a news brief, not the BIS report. It contains no methodology, no sample definition, no time series, no distinction between equity, debt, and procurement commitments. Every judgment I make about the BIS document itself carries one grade lower confidence. What I can audit is the structure the number implies, not the number.
Core
Start with the mechanism, because the mechanism is not new. It has a name and a failure record.
In the late 1990s, telecom equipment vendors โ Lucent, Nortel, Cisco โ financed their own customers. The structure was called vendor financing. A vendor would extend credit to a competitive local exchange carrier, the carrier would use the credit to buy the vendor's equipment, and the vendor would book the sale as revenue. On the income statement, this looked like demand. On the balance sheet, it looked like receivables. The two entries were the same dollar seen from two directions.
This worked while capital expenditure grew. When the CLECs collapsed in 2000 and 2001, the receivables became impairments. Lucent wrote down billions. Nortel followed. The equipment demand had never been external; it had been a function of the vendors' own lending. When the lending stopped, the demand stopped, and the impairment arrived with a lag because the revenue had been recognized years earlier. The lag is the trap. By the time the impairment is visible, the capital has already been deployed and the capacity already built.
The AI structure is isomorphic. Compute producer invests in model lab. Model lab commits to purchase compute. Producer recognizes the purchase as revenue. The dollar appears twice: once as an investment outflow, once as a revenue inflow. Net of the two, the economic substance is a transfer within a single balance sheet, dressed as a transaction between two. The dressing is not cosmetic. It changes the reported growth rate, the reported margin, and the valuation multiple that a third party applies to both.
I audited this pattern in 2017, before I had the vocabulary for it. I was a security consultant on three Ethereum ICOs. One, a vehicle claiming fifty million dollars in pre-sale, had a reentrancy vulnerability in its token distribution logic. The distribution contract called an external function before updating internal balances โ the classic ordering error, the one that later drained the DAO. I spent six weeks reverse-engineering the Solidity. I refused to sign off until it was patched. The patch cost the project two months, which killed its momentum, which cost me the client relationship.
The lesson I took was not about reentrancy. It was about ordering. In any system that moves value, the sequence of state updates is the security model. The AI financing loop has the same property. Investment precedes revenue recognition, and revenue recognition precedes the next investment. If the ordering holds, the loop is stable. If any leg reorders โ if revenue is recognized before the investment is committed, or the investment is recalled before the revenue clears โ the loop unwinds, and it unwinds faster than it built because the reordering is simultaneous across participants.
Now the revenue quality question, which is the only question that matters.
Every operating company has a customer mix. A software firm selling to enterprises has a customer mix of enterprises. A firm selling to other firms in its own sector has a customer mix of its own sector. The second is not automatically fraudulent, but it is automatically lower quality, because the demand is endogenous. When the sector contracts, the demand contracts with it, and the contraction is correlated across every counterparty. Correlation is the variable that turns a manageable loss into a systemic one.
The BIS figure implies that more than half of AI funding is endogenous. If funding is a leading indicator of revenue, then more than half of AI revenue is endogenous. The terminal consumer penetration rate โ the fraction of AI revenue ultimately paid by a party outside the AI industry โ is the number that determines whether the industry is a market or a circuit.
No one publishes this number. That absence is itself a finding. In the telecom era, the equivalent metric was available: you could trace CLEC revenue to actual subscribers and count them. In the AI era, the metric is obscured by three layers of intermediation โ cloud credits, API resale, model licensing โ that make end-user attribution structurally difficult. The obscurity is not necessarily intentional. It is a byproduct of the stack. But an auditor does not care about intent. An auditor cares about traceability, and traceability here is absent. A revenue line you cannot trace to a payer outside the system is not a revenue line. It is an internal mark.
Move to the commitment-revenue mismatch, which is where the leverage hides.
A model laboratory raises capital at a private valuation. It then commits to multi-year compute procurement. The commitment is a contractual obligation, not an expense, so it does not appear on the income statement as a cost until it is drawn. But it is a liability in substance. If the commitment exceeds the lab's current annual revenue by a multiple โ and in several disclosed cases it does โ then the lab has pre-committed future cash flows it has not yet earned and may never earn at the assumed margin.
The vendor recognizes the commitment as backlog. Backlog is not revenue, but it is used to justify capacity expansion. Capacity expansion requires capital expenditure. Capital expenditure is financed by debt. The debt is collateralized by the capacity. The capacity is only valuable if the commitment is drawn. The commitment is only drawable if the lab raises more capital. The lab raises more capital from the vendor.
This is a loop with a lever. Each turn of the loop increases the notional size of the system without increasing the external demand that would justify it. That is the definition of a procyclical structure. It expands fastest at the top and contracts fastest after, and the contraction is non-linear because the lever works in both directions. A one-dollar reduction in circular investment does not produce a one-dollar reduction in recognized revenue. It produces a multiple, because the same dollar was being counted at several points in the chain.
Emotion is a variable I exclude from the equation. I am not describing a fraud. I am describing a topology. The topology has a known failure mode, and the failure mode is not visible in any single participant's financial statements. It is visible only in the aggregate, which is exactly why a central bank is the institution that noticed.
Consider the collateral. In the DeFi liquidity mining cycle of 2020, I spent three months simulating impermanent loss under volatile conditions for a protocol advertising five thousand percent APY. The yield was mathematically unsustainable and structurally equivalent to a rug-pull disguised as innovation. I wrote a forty-page memo. The firm ignored it and lost sixty percent of the portfolio when the protocol collapsed.
The parallel here is the collateral asset. In DeFi, the collateral was a governance token whose price was a function of the yield it paid โ circular by construction. In AI, the collateral is the GPU and the data center. Both have a market price that is a function of the demand they serve. When demand falls, the collateral value falls, which forces deleveraging, which reduces demand further. The depreciation schedule compounds the problem: GPUs depreciate on a three-to-five-year cycle, while the debt secured against them may run longer. An asset worth less than the debt it secures is a stranded asset, and stranded assets are how a liquidity event becomes a solvency event.
Liquidity is a mirage; solvency is the only truth. A data-center operator with high utilization can service its debt. A data-center operator with nominal capacity and low utilization cannot, regardless of the headline contract value. Utilization is the tell. It is also, conveniently, the number that is least often disclosed. A rental price is observable; an occupancy rate is not. The industry publishes the price and hides the occupancy, which is the reverse of what an auditor needs.
Now the opacity, which is the layer the BIS actually named.
The core AI players are private companies. Private companies disclose what they choose. Their valuations are set by financing rounds, not by market transactions. A valuation set by the last round is a stale valuation โ it reflects the price a marginal investor paid, not the price at which the whole equity could clear. When a sector is expanding, stale valuations lag upward and look conservative. When a sector contracts, stale valuations lag downward and look fraudulent. The same accounting produces both impressions, and neither impression is information.
Layer on top of this GPU-collateralized debt and data-center securitization. These instruments are priced infrequently and traded thinly. Their marks are model-driven, and the models assume continuity of demand. In a downturn, the marks reprice all at once, and the loss lands in whichever balance sheet holds the instrument. The BIS, as the coordinator of bank supervision, is asking a specific question: which balance sheet is that? If the answer is bank and private credit, the transmission channel runs into the core financial system, and the 2008 comparison becomes literal rather than rhetorical. If the answer is a ring-fenced special-purpose vehicle, the loss is contained but the opacity is total. Either answer is a finding.
The 2008 lesson was never about the assets. It was about the nesting. The losses were concentrated not in the mortgages but in the opaque structures that referenced them. AI financing is building an equivalent nesting: private valuations referencing private commitments referencing private debt referencing depreciating hardware. Each layer is defensible in isolation. The stack is not.
There is a second opacity, and it is the one I am living inside right now. For the last three months I have been auditing the data input pipelines of an AI-agent protocol that claims to use decentralized models for real-time financial modeling. The smart contracts are clean. The oracle layer is not. The training data fed into the inference pipeline carries distributional bias that the contract cannot see and the user cannot inspect. When an AI agent consumes a biased feed and writes a price or a credit decision on-chain, the bias becomes immutable. The chain does not add trust to the computation; it adds permanence to the error.
This is where my two disciplines converge, and it is the insight the AI-financing debate is missing. Verifiability is not a feature of a model. It is a property of the computation, and it must be enforced structurally, not asserted in a whitepaper. I spent six months in 2022 studying Plonk and Spartan proof systems and contributing to an open-source library for efficient proof verification, precisely because I had realized my earlier critiques lacked mathematical grounding. A zero-knowledge proof can attest that a computation was executed correctly without revealing the inputs. Applied to AI inference, it can attest that a model produced an output according to a committed set of weights and a committed input, which is the only way to make an AI-driven financial decision auditable after the fact.
None of the current AI-financing structures has this property. The revenue is unverifiable, the collateral is unverifiable, and the inference is unverifiable. Three layers of the same defect. The circular financing is not the disease; it is the symptom of an industry that has scaled faster than its capacity to prove anything about itself.
Contrarian
Now the part the skeptics get wrong, because a one-sided teardown is not an audit. An audit reports both the finding and the mitigating control.
The bulls are correct about one thing: vertical integration has genuine economic logic. When compute is the binding constraint on model development, a producer that invests in a consumer secures demand for its own output and gains visibility into its customer's roadmap. That is not fraud. That is supply-chain strategy, and it exists in every capital-intensive industry โ automotive, aerospace, semiconductors. The question is never whether integration exists. The question is whether the integration is priced as risk or as certainty. In a bull market it is priced as certainty, and certainty is the one input no model should ever assume.
The bulls are also correct that a circular structure can be self-sustaining for a long time. A loop does not require external validation to keep turning. It requires only that each participant prefers to continue rather than to stop. As long as capital expenditure growth continues, every participant's incentive is to continue, because stopping means ceding share to a competitor who does not stop. This is a coordination problem, not a solvency problem, and coordination problems can persist for years. The 1999 telecom loop ran for at least three years past the point where a careful analyst could see the receivables building.
This is why the BIS signal is a yellow warning, not a red alert. A central bank naming a structure does not time a collapse. It flags a fragility. The fragility may take years to express, or it may never express if external demand arrives in time to convert the circuit into a market. Sovereign AI programs, enterprise deployment, and regulated-industry adoption are all plausible sources of that external demand. If they scale faster than the loop unwinds, the loop becomes a market and the skeptics are wrong. The mitigating control is real. It is simply not yet large enough to measure.
I have been wrong in exactly this way before. In 2021 I audited an NFT collection called PixelFlux. I found that forty percent of its rare traits were algorithmically impossible due to an error in the rarity calculator. I published the finding. The floor value fell ninety percent within a week. But I also underestimated how long the surrounding market would sustain itself despite equivalent flaws. The code was wrong. The market did not care, until it did. Timing is the variable I have learned to hold most loosely.
There is also a regulatory asymmetry the bulls should note, and it cuts against the complacent reading. Most project-level compliance in this industry is theater. I have said for years that KYC on a token sale is bypassed by buying a few wallets' worth of holdings, and that the compliance cost is passed entirely to honest users who file the paperwork. The AI-financing loop has the same asymmetry at a larger scale. The disclosure burden falls on public-market participants and regulated banks, while the circular transactions occur in private markets and special-purpose vehicles that face no equivalent requirement. Regulation that binds the transparent and exempts the opaque does not reduce systemic risk. It relocates it. That is the structural reason a BIS warning matters more than a private analyst's warning: only the BIS can change the asymmetry.
So the honest position is not "AI is a bubble." The honest position is "AI's revenue quality is unverified, its collateral is procyclical, its marks are stale, and its funding is more than half endogenous โ and none of those four facts is dispositive on its own." What makes them dispositive in aggregate is their correlation. They fail together, in the same quarter, for the same reason: the loop stops turning. A skeptic who predicts the timing is guessing. A skeptic who measures the correlation is auditing.
Takeaway
The number is 55.2%, and the number is not the point. The point is the denominator that was never published, the revenue that was never traced to an external payer, and the utilization that was never disclosed. Accountability in this cycle does not look like a prediction. It looks like a disclosure standard: require every AI firm above a threshold to report the fraction of revenue paid by non-AI customers, require every lender holding GPU-collateralized debt to mark it to observable rental prices rather than to model assumptions, and require every AI agent writing financial decisions on-chain to commit to a verifiable computation.
That standard does not exist yet. When it does, it will have been drafted because a central bank noticed a loop before the loop noticed itself. The question for the next quarter is not whether AI is real. It is whether the money paying for AI comes from outside AI. Until that number is published, every valuation in the sector is an internal mark, and an internal mark is a hypothesis, not a price.