The Money Goes in a Circle. Should You Care?

Circular financing, in the plain sense, is any arrangement in which the same capital appears to originate demand, fund the supply that meets the demand, and return to its origin. In the current cycle, the pattern presents in a few recurrent forms. A chip maker invests in, or takes a commercial commitment with, a model laboratory. The laboratory uses the commitment to sign a multi-year compute contract with a cloud provider. The cloud provider issues bonds or takes on private credit to fund the data centre build. The proceeds pay the chip maker for the accelerators that populate the data centre. On any single leg, the transaction is arm's-length, disclosed and defensible. Taken as a system, the loop closes.

The BIS position is that this arrangement is not fraudulent, and is not comparable to the accounting fictions that preceded earlier bubbles. What the BIS says instead is that the loop amplifies. If the ultimate customer demand for the compute is smaller than the contracted quantity, the shortfall does not present at a single balance sheet. It presents at all of them at once, because each participant is depending on the next participant's ability to pay. The failure mode is not concealment. The failure mode is correlation. That is a different kind of risk, and it is the risk that the BIS believes the market is not pricing.

The practical question for the ordinary investor is not whether the artificial intelligence trade is a Ponzi. It is not. The practical question is what fraction of the artificial intelligence exposure inside a diversified portfolio is exposed to the loop, and what would happen to that exposure if the loop tightens. On the published index composition, a passive S&P 500 investor now holds a materially larger share of the portfolio in the seven names most directly connected to the loop than at any point since the composition began to be measured in the current form. The concentration is not evidence of malpractice. It is a fact about the shape of the market. The reader who understands the shape can make a personal decision about the size of the exposure. The reader who does not is holding a position they have not chosen.

The practical question is what fraction of the artificial intelligence exposure inside a diversified portfolio is exposed to the loop, and what would happen to that exposure if the loop tightens.

The second practical question is the private credit exposure. Most retail investors do not hold private credit directly, but many hold it indirectly through pension funds, insurance policies and interval funds. On the BIS figure, roughly eight per cent of the total private credit stock is now exposed to artificial intelligence infrastructure. That number is not, on any reasonable base rate, a systemic figure in isolation. It becomes a systemic figure if the correlation described in the BIS bulletin is realised and the losses arrive together rather than sequentially. The disciplined reading is to know the exposure, not to exit it.

The honest counter to the sceptical reading is that the revenue at the endpoint is real, growing and increasingly diversified. Enterprise contracts with the Fortune 500, seat licences at the professional services firms, the application programming interface usage across the software industry, and the federal contracts on both sides of the North Atlantic together produce a recurring revenue base that is measured in the tens of billions and growing at rates that few large-cap technology businesses have posted in the past decade. The revenue is not fictitious. The question the BIS is asking is not whether the revenue exists, but whether the pace and scale of the infrastructure build is calibrated to the pace and scale at which that revenue can compound.

The pace of the build, on the current guidance, requires the enterprise revenue to compound at rates that, in aggregate, have historically been seen only during the earliest phase of a new general-purpose technology. That is a plausible base case. It is not the only base case. A slower diffusion, in which enterprises adopt in stages and demand for compute grows more linearly than the guidance implies, would leave a meaningful share of the contracted supply unabsorbed for a period. The absorption gap, not fraud, is the mechanism through which the loop tightens.

A tightening does not present as a crash. It presents as a repricing. The observable early signals, on the historical pattern of previous capital cycles, are three. The first is a modest widening of the credit spreads on the hyperscaler bond issuance relative to the broader investment-grade index. The second is a slowdown in the pace at which new private credit vehicles allocate to artificial intelligence infrastructure. The third is the pattern of contract amendments, principally the extension of delivery windows and the softening of minimum-take commitments, at the interface between the model laboratories and the cloud providers. None of these signals is a headline event. Each is visible in the disclosures of the participants and in the pricing of the traded instruments. A reader who follows the three signals through the next four quarters will have a better view of the cycle than a reader who follows only the headline valuation of the listings.

The historical comparators are instructive without being determinative. The telecommunications capital cycle of the late 1990s presented a similar pattern, in which vendor financing by the equipment makers of purchases by the network builders inflated the apparent demand for the equipment. The tightening arrived in stages, first at the private credit layer, then at the equipment maker margins, then at the listed equity prices. The eventual recovery was uneven. Some of the equipment makers regained their previous highs on a decade horizon, and some did not return at all. The lesson is not that the current cycle will end the same way. The lesson is that the shape of the tightening, if it arrives, will not resemble a single-day event.

The BIS is not a supervisor. Its verdict is a signal to the supervisors, and the supervisors that matter here are the Federal Reserve, the Office of the Comptroller of the Currency, the European Central Bank, the Bank of England and the Financial Stability Board. On the pattern of previous BIS interventions, a directional shift in the language of these supervisors typically follows a BIS bulletin of this weight within one to two supervisory cycles. The most likely form is a targeted request for enhanced disclosure at the interface between the private credit vehicles and the artificial intelligence infrastructure borrowers, together with a modest tightening of the capital treatment of the direct bank exposure. Neither step is disruptive on its own. Together, they compress the appetite of the marginal balance sheet in the loop, which is the mechanism by which supervisors slow a cycle without breaking it.

For the reader, the regulatory read is a signal about the pace of the tightening, not about its direction. A supervisory response that arrives quickly reduces the amplitude of the eventual repricing. A supervisory response that arrives late increases it. The reader is not required to predict which of the two paths the supervisors take. The reader is required to notice which path is being taken as the disclosures accumulate, and to size the position accordingly.

LUMINAIRE does not publish price targets and does not recommend the purchase or sale of individual securities. Three observations, however, are worth carrying into the second half of 2026.

First, know the exposure. The concentration inside the standard passive vehicles is not, in itself, evidence of imprudence, but it is a fact about the shape of the portfolio that the reader should be able to state aloud. The number is easy to look up and easy to hold in mind.

Second, separate the two questions. The question of whether the revenue at the endpoint exists is a question about the diffusion of a new general-purpose technology, and the honest answer is that it does, at a pace that is genuinely without a modern precedent. The question of whether the infrastructure build is calibrated to the pace of the diffusion is a separate question, and the honest answer is that the BIS believes it is not. Both questions can be true at once.

Third, watch the interface, not the headline. The private credit disclosures, the corporate bond spreads and the contract amendments are the signals that will tell the reader, months in advance, whether the cycle is tightening or extending.

Readers who want the full institutional treatment can read Cabier's companion release, The Circle Tightens, which sets out the BIS taxonomy in the form that the allocators are reading it. The commitment-coverage model, which decomposes a portfolio's exposure to the loop by counting revenue from customers outside the loop against contracted supply inside it, lives on CalculatorIQ. Both are linked at the foot of this article. The educational and illustrative purpose of all three pieces is the same, to support a reasoned personal assessment. Nothing here is investment advice. The desk used Anthropic in preparation for research synthesis. Editorial judgement, structure and conclusions are the desk's own.