The Circle Tightens.

Executive summary

In the AI capex cycle that has defined 2024–2026, chip vendors, hyperscalers and the two largest model labs are bound together by an unusual density of equity stakes, prepayments, capacity contracts and vendor financing. Each link in the loop is defensible on its own terms. Read together, they leave the sector partially self-funding, and they leave a meaningful share of recognised revenue dependent on commitments made by counterparties that are themselves customers.

The institutional question is not whether the loop is improper. It is what happens when the cost of money rises, when a single node renegotiates, or when the first audited filings force the relationships into the open. This article maps the loop and isolates the three channels through which a re-rating would propagate.

An institutional turn

For most of the cycle, circular financing has been treated as a feature of an early market — the natural consequence of three or four firms holding most of the world's frontier compute, model and distribution capacity. The turn, in 2026, is that the same firms are now preparing audited public filings, and that allocator committees are being asked to underwrite the next round of capex on the assumption that revenue growth continues at its current pace. Both conditions raise the evidentiary bar.

What circular financing is — and what it is not

Circular financing, in the precise sense, is the recognition of revenue by one party that depends materially on capital, credit or commitments extended by a counterparty that is also a customer or supplier. It is not a synonym for "round-trip" revenue, and it is emphatically not a synonym for "Ponzi". Each of those framings has been raised in the press; each deserves to be dismantled before it is borrowed.

The honest description is narrower and harder to dismiss. A chip vendor takes an equity stake in a model lab; the lab signs a multi-year capacity contract with a cloud platform; the cloud platform, in turn, is one of the chip vendor's largest customers. Revenue at every node is real. The question is whether the loop, taken as a whole, can continue to clear without sustained external demand growing at the pace the marks assume.

Where the strain shows first

Three places, in order. First, the cash-to-debt ratio of the model labs: operating losses measured in the tens of billions are funded by a mixture of equity, prepayments and, increasingly, vendor and bank credit. Second, the private-credit exposure to data-centre and GPU-leasing structures, much of it warehoused outside the regulated banking perimeter. Third, the off-balance-sheet commitments — purchase obligations, take-or-pay capacity, equity earn-ins — that the first audited S-1s will be required to disclose in their full magnitude.

The institutional discipline is to read the three together. A re-rating of any one will pull the other two with it; supervisors, in turn, will read the package and draw conclusions about systemic exposure that no single issuer's S-1 will frame.

The rate channel

The loop was assembled, and the marks were struck, in a falling-rate regime. The cost of the vendor financing, the discount rate applied to the capex programmes, and the implied option value of the frontier-capability bet are all rate-sensitive. A meaningful and sustained move higher in policy rates — or a widening of credit spreads in the private-credit channel — does not, on its own, break the loop. It does, however, force the renegotiation of the marginal contract, and renegotiation at scale is how the loop transmits stress from the financing layer into the operating layer.

The institutional takeaway

Boards, CROs and allocator committees should be testing three things now. First, the share of issuer-recognised revenue that is sourced from counterparties inside the loop, versus revenue sourced from outside it. Second, the rate sensitivity of the financing layer, modelled on a sustained higher-for-longer base case, not on the roadshow case. Third, the disclosure package each issuer is preparing for its S-1, and whether it permits a disciplined sum-of-the-parts walk between fundamentals, discounted optionality and named narrative premium.

A forthcoming Cabier tool will help institutions express the first of those tests quantitatively. The other two are governance work, and they cannot be outsourced to a model.