The Circle Tightens

On 24 June the Bank for International Settlements published its Annual Economic Report for 2026 and, in the section on non-bank finance, retired a euphemism. What analysts and central bankers had been calling "vendor financing" or "strategic partnerships" the BIS now calls a circular investment structure. Chip suppliers take equity in model labs. Model labs sign multi-hundred-billion-dollar cloud contracts with the same hyperscalers who own equity in them. Hyperscalers guarantee revenue for chip suppliers whose products they are also buying. The BIS conclusion is unambiguous. When each participant is materially exposed to every other, a shock to any one of them propagates through the ring, not around it.

What the BIS actually said

The report frames the concern in three parts. First, capital expenditure by the largest AI-adjacent firms in 2026 is on track to exceed the annual capex of the entire global oil and gas industry at its peak, with a materially larger share funded by debt than a decade ago. Second, private credit exposure to AI model labs, data-centre operators, and specialist chip supply has moved from a niche allocation to more than two hundred billion dollars of committed capital, much of it at fixed spreads set before the last two rate cycles. Third, the fastest-growing category of counterparty concentration in bank and non-bank books is intra-AI. Facilities where the borrower's revenue is contractually linked to another borrower in the same book.

The BIS does not call this a bubble. It calls it a topology problem. Bubbles pop; topologies unwind. The distinction matters because the policy response is not the same. A price correction is absorbed by whoever holds the paper. A topology unwind is absorbed by whichever counterparty is furthest from the cash and closest to a covenant. That is a different set of names, and the report gently observes that the regulators covering them are not the same regulators who cover the rest of the AI capex stack.

Each participant is materially exposed to every other. A shock to any one of them propagates through the ring, not around it.

The numbers, plainly

The single most cited data point is the Nvidia-OpenAI letter of intent, reported at up to one hundred billion dollars of chip supply and capital commitment, structured across multiple years and conditional on capacity milestones. That figure sits alongside Oracle's roughly three hundred billion dollar multi-year cloud arrangement with the same model lab, a Microsoft-Azure commit of similar order, and smaller but still ten-figure supply and equity arrangements with AMD and Amazon. The counterparties collectively pledge more than five hundred billion dollars of forward compute and capital to an entity whose disclosed annualised revenue run-rate is a small fraction of that number.

Anthropic's picture is directionally similar but smaller. Its two largest cloud counterparties are also two of its largest equity holders. The private credit stack sitting behind both firms is not disclosed in comparable detail, but the reported growth in AI-specific direct-lending funds over the last two years is consistent with the BIS estimate.

Why circularity is a real risk, not a rhetorical one

Circular exposure is not new. Vendor financing has been part of technology cycles since the late 1990s. Two features distinguish this cycle. The first is the concentration of counterparty. Five names sit on the receiving end of a large majority of the commitment book. In a distressed scenario, a single downgrade at any of them repriced every facility that references them. The second is the leverage sitting behind the ring. Private credit funds have written multi-year facilities against revenue projections that assume the ring holds together. If the ring narrows, the collateral value narrows with it.

The technical term for what worries the BIS is "wrong-way risk". The exposure grows precisely in the scenarios in which the counterparty's ability to pay shrinks. It is the same feature that made counterparty exposure to monoline insurers ruinous in 2008. Not because monolines were unusually large but because their exposure grew as the underlying they were insuring deteriorated. Circular AI financing has the same signature. Chip supply is guaranteed by cloud revenue which is guaranteed by chip supply.

Where the loop is thin

The thinnest point in the loop is outside revenue. A hyperscaler with a hundred billion dollars of contracted spend to a model lab can absorb a delay if enterprise cloud demand outside the AI stack keeps growing. A model lab with tens of billions of dollars of committed compute can service that commit if end-customer subscriptions grow into it. The systemic question is whether that outside revenue is scaling fast enough to make the ring self-liquidating over any reasonable horizon. On the numbers currently disclosed, it is not.

We built a small tool to make that question concrete. Enter each named commitment, mark whether the counterparty sits inside the ring, and set an estimate of what share of the entity's revenue comes from outside it. The output is a coverage ratio, a years-to-cover figure, and a zero-to-one-hundred circularity index that responds to rate shocks and outside-revenue drops. The presets are the OpenAI and Anthropic pictures as of writing. The tool is not a forecast. It is an arithmetic that forces the ratio into the open.

What breaks it

Two things break a circular financing structure. A rate shock that revalues the debt, and a demand miss that revalues the underlying. In this cycle, either can happen without the other. Rates have already surprised twice in the last three years. Demand for enterprise AI is real but its willingness to pay at the price points assumed by the capex plans is not yet demonstrated at the scale required. A five percentage point miss on either variable, applied to the ring, does not produce a linear result. It produces a repricing at every node.

The BIS report does not recommend intervention. It recommends disclosure. Specifically, it asks banks and non-bank lenders to identify and report intra-AI counterparty concentration in a consistent format, and it asks the SEC, ESMA, and the FCA to consider whether existing disclosure rules capture the structure of the commitments. The Financial Stability Board is expected to take up the topic at its September meeting.

What retail should take from this

Two things. First, the language has changed. When the BIS names a structure, that name enters the vocabulary that every prudential regulator will use for the rest of the cycle. Every rating action, every stress test, every capital surcharge decision from here forward will reference circularity. That will show up in cost of capital, not immediately, but on the schedule that regulatory language always moves.

Second, the structure implies a specific pattern of drawdown. Circular exposures do not fail alone. They fail together. Portfolio risk that is built on the assumption that AI-adjacent names diversify each other is likely mispricing the covariance. The correction, if and when it comes, will look correlated in a way the historical data does not fully price.

Portfolio risk built on the assumption that AI-adjacent names diversify each other is likely mispricing the covariance.

The filing race, and why it matters here

The other feature of this quarter is the confidential IPO filings from OpenAI and Anthropic. On the reported terms, both filings ask public markets to accept valuations that only make sense if the outside revenue side of the coverage ratio grows into the commitment side over the next several years. That is the same question the BIS is asking, translated into price. Our companion piece takes the filings apart on their own terms.

The through-line is that the circle and the filing are the same story. The commitment book is the input. The listing valuation is the output. If the output prices in a version of the input that the disclosures do not support, that gap is not narrative premium. It is systemic risk being distributed to retail balance sheets.

What we will be watching

Three markers. First, whether the Financial Stability Board in September echoes the BIS language on circularity or dilutes it. Second, whether any of the named cloud or chip counterparties disclose the commitments in comparable form in their next quarterly filing, or continue to reference them in prepared remarks only. Third, whether the private credit funds most exposed to the AI stack begin to mark facilities to reflect intra-ring counterparty concentration, or wait for a rating action to force them to.

The circle is not new. What is new is that the largest global prudential authority has now named it. That naming has consequences. The rest is arithmetic, and the arithmetic is now available for anyone who wants to do it.