Everybody is financing everybody

12 / The circular economy, but for conviction

A supplier invests in a customer. The customer uses some of its resources to buy the supplier's product. Another company provides infrastructure, financing, or distribution. These relationships can support real deployment. They can also make it harder to distinguish independent end demand from an ecosystem financing its own expansion. The structure alone is not evidence of fraud. It is an invitation to examine the contracts and cash flows.

A July 2026 BIS working paper models the AI investment race as competition for a small number of dominant positions. In the model, debt and circular equity links can amplify fragility and transmit losses through connected firms. These are model-based findings under specified assumptions, not a measured guarantee that actual AI companies will fail, and not an official date for a crash.¹⁵

Imagine three fictional businesses: a chip supplier, a model developer, and an application company. The supplier invests in the developer. The developer commits to infrastructure. The application company receives funding partly because it can build on the model and commits to use it. There may be excellent reasons for every step. The task is to look through the corporate boundaries and ask where money from customers outside the financing circle enters.

The analogy should not be pushed too far. Strategic customers often help fund new capacity precisely because demand is expected to be large. Long-term purchase commitments can reduce uncertainty and make investment possible. Vertical relationships can solve coordination problems. A critic who treats every partnership as suspicious has substituted a conspiracy theory for supply-chain analysis.

What would make the arrangement more convincing? Independent customer receipts. Transparent contractual obligations. Evidence that customers renew when subsidies fall. Financing that can tolerate delays. A clear account of who bears demand risk and who can walk away. None is as exciting as a diagram of interconnected logos. That is why the logos are generally larger.

In a September 10, 2026 speech, the BIS's Pablo Hernández de Cos connected ambitious AI earnings expectations with increasing use of debt and private credit, and described circular financing relationships among suppliers and AI firms. He warned about a potential reversal if returns disappoint, while explicitly saying that a bust was not inevitable. The caution is about a financing structure and an investment race, not a verdict that the technology lacks value.¹⁶

13 / A data center cannot run on a change of narrative

Some costs are unusually resistant to rhetoric. Electricity is one. Cooling is another. So are land, equipment, installation, maintenance, and the time required to connect physical infrastructure. The word cloud has performed an extraordinary public-relations service for a collection of facilities that must obey local engineering constraints.

The IEA's 2025 Energy and AI report projected that global data-center electricity consumption would roughly double to about 945 terawatt-hours by 2030 in its base case. That is a scenario projection for data centers overall, not a meter reading of AI alone and not an assured outcome. Its value for this discussion is to make the physical scale of the buildout visible.¹⁷

An imaginary infrastructure project has two schedules. The first says when the servers arrive. The second says when sufficient power becomes available. If the first runs ahead of the second, expensive equipment may wait. If the second requires commitments made before demand is certain, someone is underwriting that uncertainty. The problem may be solved through sensible contracts, phased development, or better planning. It is not solved by calling the project inevitable.

This is another way real progress can inconvenience the people financing it. Faster technical improvement benefits users while threatening the economics of existing capacity. The next model may make the service more valuable, or reduce the need for a particular kind of computation, or intensify competition. A company can be exposed to AI succeeding differently from the way its investment plan expected.

The power engineer has joined the imaginary board meeting. She has a list of constraints and an unremarkable pair of shoes. Nobody has asked her about the end of work. Several people have asked whether she can accelerate a transformer.

14 / The second spreadsheet

Let us give the finance director a second fictional example. A facility earns 30 in annual revenue and spends 18 on operating costs, leaving 12 before interest, taxes, replacement investment, and other charges. It owes 8 in annual interest. On this deliberately simplified basis, there is a cushion of 4 before those other demands. Now let revenue fall by 20 percent while operating costs stay unchanged. Revenue becomes 24, the operating surplus becomes 6, and interest still costs 8. A business that looked profitable under one assumption now cannot cover interest from that surplus.

This is not a model of an actual data center. Real contracts, variable costs, hedges, taxes, depreciation, and financing terms would change the result. The example demonstrates why a modest disappointment in revenue can produce a large change in financial flexibility when costs and obligations are relatively fixed. It also demonstrates why the original cushion should not have been casually described as free cash flow.

Now change the assumptions in a more favorable direction. Suppose costs fall with revenue, the project has more equity, or customers have creditworthy take-or-pay commitments. The stress may be manageable. Suppose the owner has other profitable operations and can fund the shortfall without jeopardizing them. The situation changes again. An analysis that omits these possibilities is not tough-minded. It is incomplete.

In the imaginary boardroom, someone asks whether the downside slide can be moved to the appendix. It makes the opportunity look complicated. The finance director explains that the opportunity is complicated. This is why she is not in charge of the narrative.

15 / What turns a disappointment into a crisis?

A failed product is not a recession. A falling stock market is not automatically a banking crisis. A credit loss is not necessarily contagious. To get from an investment disappointment to a systemic event, something must connect the original problem to the ability of other people and institutions to function. The path is more important than the adjective.

One possible path begins with returns below expectations. New financing becomes harder to obtain. Companies cut spending, suppliers lose business, and borrowers struggle to refinance. Lenders and investors become more cautious. If leverage, concentrated exposures, or forced sales amplify the pressure, losses can spread. If institutions have adequate capital, stable funding, and time to adjust, the same initial disappointment may remain contained. This is a conditional mechanism, not a forecast for a particular year.

The practical question is how many promises depend on the same favorable conditions. A company may appear diversified across several products while all its customers depend on the same funding market. A lender may have many borrowers whose revenues depend on one sector. A household may have a job, a home, and a portfolio that all benefit from the same local boom. Counting the names does not count the independent sources of resilience.

This is why a map of exposure can be more informative than a list of villains. Who owes whom? Which obligations are short-term? Which collateral values move together? Which contracts permit an early exit? Which institutions can absorb losses without reducing essential services? A system's strength is not the sum of its most confident presentations.