How the Race Finances Itself—and Front-Loads the Risk
The AI infrastructure race changes more than the amount of capacity firms want to build. It also changes the contracts and timing through which they build it. This chapter examines circular financing, debt, and the early-mover premium so the reader can see why a financing arrangement may solve a real coordination problem while still making the eventual boom more exposed to uncertainty.
Circular financing is coordination with a failure mode
A conventional equity investment gives one party a claim on another party’s future value. The arrangement examined in the AI build-out is more operational. A hyperscaler takes an equity stake in an AI laboratory, and the lab commits to purchasing compute from the hyperscaler over time. The stake and the purchase commitment reinforce each other. The builder has a reason to reserve and expand capacity for the lab. The lab has a reason to keep using that capacity rather than redirecting its demand to a rival.
The BIS model calls this circular financing because capital and future purchases run around the same relationship. It is not “circular” in the narrow sense that no real service is delivered. The lab can actually use the compute, and the hyperscaler can actually build the infrastructure. The arrangement addresses a hold-up problem: after one party has made a specialised investment, the other party may have an incentive to renegotiate, delay, or walk away. An equity stake gives the hyperscaler a claim on the lab’s success and makes the lab’s future purchase commitment more credible. In that limited sense, circular financing can improve productivity by preserving a partnership that would otherwise be too fragile to support a large build-out.
The contract changes the incentives on both sides, but it does not make demand guaranteed. The lab’s commitment is valuable only if the lab continues to need the capacity and can pay for it. The hyperscaler’s equity stake is valuable only if the lab’s future business and valuation remain strong. In a favourable state, the two claims reinforce one another: the lab has a reason to use the capacity, and the hyperscaler has a reason to keep expanding it. In a weak state, they can deteriorate together. The lab may scale back purchases, reducing the expected cash flow behind the hyperscaler’s investment, while the hyperscaler’s retrenchment makes it harder for the lab to obtain the compute it expected. A relationship that reduced coordination risk in the boom can therefore concentrate exposure in the downturn.
This is also why an equity stake in a partner is not the same as diversification. The stake may look like an asset on one line of a balance sheet and the purchase commitment like revenue on another, but both depend on the same underlying AI success. Their risks are positively related. If the arrangement ties a large amount of capacity to a small number of labs, a common disappointment can impair the asset and the expected demand at once. The paper’s calibration uses an equity-weighted circular share of 20% as an anchor from disclosed deal data. That number is not a claim that every firm has the same exposure; it illustrates that circular positions can be material enough to affect the sector’s financing mix.
That benefit is not the end of the story. The coalition that owns both sides of the relationship also receives the contest benefit described earlier. More committed compute can strengthen its position relative to competing coalitions, and a larger circular share can help it secure demand for the capacity it builds. A social planner values the genuine coordination benefit. It does not value the part that merely shifts contest revenue from one coalition to another. The coalition counts both.
This is why the model predicts weakly more circular financing by a coalition than by the planner when circular financing has a positive social benefit. The result is conditional, not moral. If the partnership benefit were negative at the margin, the contest multiplier would not make circular financing desirable. The potential excess arises when a useful arrangement is also a tool for winning a relative-position contest. The same contract can be productive at one scale and excessive at another.
The composition of funding matters because not all capital has the same fate in a downturn. Internal equity is a claim the coalition supplies from its own resources. Circular capital is tied to the continuation of the lab–hyperscaler relationship. Debt-financed capital carries a repayment obligation and depends on the ability to sell or continue operating the underlying infrastructure. If the partnership unwinds, some circular claims lose their productive connection. If all firms retrench at once, debt-financed infrastructure may have to be sold into a market in which the natural buyers are also short of cash.
A useful way to view the arrangement is therefore as both an insurance mechanism and a concentration mechanism. The purchase commitment gives the builder confidence that capacity will be used. The equity stake gives the builder exposure to the lab’s upside. But the same exposure means that a lab’s disappointment can reduce the value of the backer’s asset and weaken the demand that justified the original construction. Diversification is not automatic merely because the contract is labelled an investment and a purchase separately.
Early commitment moves uncertainty onto the balance sheet
The timing of the model matters as much as the contract. Investment is split into an early period and a later period. In the early period, a coalition chooses how much capacity to commit before it knows the productivity draw for the technology. In the later period, it observes that draw and decides whether additional deployment is worthwhile. The distinction creates an option: waiting preserves information. But the option has a competitive cost if early capacity attracts users, data, developers, researchers, and the market position that compounds those advantages.
The BIS model represents this with a head-start advantage. A unit of capacity deployed later is less valuable in the contest than an otherwise similar unit deployed earlier. The later unit has less time to accumulate training and usage, and it arrives after rivals may have already locked in customers and complementary resources. The paper calibrates the late-deployment factor to about 0.85 in its baseline, an illustration that later capacity is useful but not equivalent to early capacity.
A coalition therefore faces two opposing considerations. Early deployment earns a strategic premium, but it is exposed to the productivity draw. Waiting reduces the chance that capital is sunk before a disappointment, but it may surrender the position that makes the investment pay. The planner faces the same technological reason to deploy some capital early. The coalition faces an additional zero-sum reason: being early can improve its share of the market even when early deployment does not increase total sector output by the same amount.
Consider the decision from the point of view of a lab that expects model quality to improve over time. Waiting might allow it to buy cheaper or more capable compute after the uncertainty is resolved. But if a rival uses the waiting period to attract developers and build a durable user base, the lab may lose more than the price of the delayed hardware. It can also lose access to the partner, data, or distribution channel that would have made its later investment valuable. The head-start premium converts that strategic possibility into a financial incentive to commit early. The coalition is not simply choosing between “invest” and “do not invest”; it is choosing how much uncertainty to accept in order to preserve its place in the contest.
The trade-off is especially important when the physical capacity is difficult to redeploy. A later software improvement may make an old model less competitive, but it does not necessarily make the associated data-centre equipment worthless. The risk comes from having to sell or repurpose capacity at the same time that every rival is trying to do so. Early commitment can thus be rational as a competitive defence and still increase the sector’s common exposure to a weak productivity draw.
That additional reason leads to front-loading. The coalition commits more in the first period than the planner would choose, all else equal. The word “front-loading” is important. The distortion is not only that the sector builds too much in total. It builds too much before uncertainty is resolved. That timing turns an uncertain revenue opportunity into a fixed balance-sheet exposure.
Debt makes the choice possible and consequential. In the model, a coalition can borrow against pledgeable future income. Period-one borrowing is based on expected payoff before the productivity draw. Period-two borrowing can be marked to market against the realised state. A profitable build-out may therefore support substantial borrowing while the outlook is favourable, but the value of the collateral can change when the draw is poor.
The symmetric model deliberately keeps the boom-side borrowing constraint slack for the representative, revenue-rich coalition. That does not mean debt is irrelevant. It means the paper is isolating the contest and fire-sale mechanisms rather than assuming that the initial boom is stopped by a lender’s limit. The coalition still chooses leverage because it weighs the cost of debt against the expected payoff and the recovery value of capital in a bust. The constraint becomes more central when the model introduces heterogeneous backers and an idiosyncratic loss in the financing network.
The risk is distributed unevenly across the parties. Equity absorbs a fall in the value of the lab or the infrastructure, but debt creates a fixed obligation that does not shrink simply because the asset has become less useful. A lender may agree to finance a capacity build because projected revenue leaves a comfortable cushion above the repayment amount. That cushion can disappear when prices fall, when a purchase commitment is cancelled, or when a specialised asset has few buyers. The borrower then has to cut future investment, sell assets, or withdraw from other commitments. In a common downturn, those choices are made by many borrowers simultaneously, which is why the eventual recovery price can be lower than the recovery each borrower would have expected in isolation.
The timing of collateral matters as well. Before the productivity draw, a coalition can borrow against expected revenue across possible states. After a weak draw, the same capacity may support less borrowing because its mark-to-market value and expected cash flow are lower. This creates a one-way pressure in the bad state: the coalition cannot rely on the financing capacity that helped it commit early. The model does not require lenders to make a sudden irrational change. A lower borrowing limit follows from the lower value of the pledgeable income and the specialised asset.
The funding cycle can be represented as a sequence:
The sequence highlights what persists and what is lost. The equity stake and purchase commitment exist before the draw. Early capacity is sunk before the coalition knows whether later deployment will earn its cost. In a strong state, the initial commitment creates a platform for further expansion. In a weak state, the same commitment can become a stranded claim: the partnership no longer supports all of the expected use, while the debt remains payable.
The structure also explains why the phrase “self-financed” can be misleading. A hyperscaler may hold substantial cash, yet a rapid capex increase can still involve bonds, leases, special-purpose vehicles, or non-bank lenders. The source model does not attempt to map every such channel. Its point is narrower: once the investment race rewards early scale, firms have a reason to select funding that allows them to commit before the outcome is known. Funding is part of the competitive strategy, not a neutral wrapper around an already-determined investment level.
There is an important qualification. Opaque circular flows could, in principle, create book revenue that is mistaken for genuine income and inflate the collateral available for later borrowing. The BIS analysis explicitly leaves that book-inflation channel for future work. Its main results treat the disclosed commitments as real and focus on the strategic benefit, the timing choice, and the impairment that occurs when a partnership or specialised asset cannot continue. The omission makes the mechanism easier to isolate, but it also means the analysis is not a complete account of shadow finance.
The central trade-off can now be stated precisely. Circular financing can make a lab and a hyperscaler more willing to invest in a shared future. Early deployment can make capacity more valuable in a contest. Debt can bridge the period between commitment and revenue. Together, those choices can increase productive capacity in a good state. They also ensure that more capital has been committed, financed, and connected before the productivity draw is known. The next question is what happens when the draw is not strong enough to support the capacity already on the balance sheet.