When Everyone Is Everyone Else’s Customer: The Circular Economics of the AI Boom

Editorial image generated by the author.

Artificial intelligence is usually described as a supply chain. Nvidia sells accelerators, cloud companies install them in data centres, AI laboratories rent the computing capacity, and businesses and consumers eventually pay for AI services. Increasingly, however, the money does not move through that chain in only one direction.

Suppliers are becoming investors. Customers help finance suppliers. Infrastructure providers borrow against contracts signed by companies that may themselves depend on continued external financing. A chip manufacturer can sell the hardware, invest in the company buying it and commit to purchasing some of the capacity subsequently built around its equipment.

The transactions can all be genuine and still be less independent than they appear. Valuations, infrastructure backlogs, semiconductor sales and financing decisions may ultimately rest on the same expectation: that future customers will pay enough for AI services to justify the extraordinary amount of capital now being committed. Several apparent signals of demand can therefore be different financial expressions of one underlying assumption.

That changes the nature of the risk. The AI boom increasingly resembles an interconnected financial network as well as a technology supply chain. Such networks are exceptionally good at moving capital quickly. They are also good at transmitting mistakes.

Why Circular Financing Exists

AI infrastructure presents an awkward financing problem. Advanced computing capacity has to be constructed before the customers who will ultimately pay for it have necessarily generated enough revenue to finance it. A data-centre operator cannot wait until demand has fully materialised before securing land, power, cooling equipment and thousands of accelerators. Nor can an AI laboratory necessarily fund the next generation of models from cash flow produced by services that are still being developed.

The industry therefore has to pull future capacity into the present. Long-term compute contracts, strategic investments, guarantees, debt financing and equity partnerships all help accomplish that. Imagine a specialist cloud provider planning several billion dollars of new capacity. Lenders want evidence that somebody intends to use it, so a major AI company signs a long-term compute contract. That commitment improves the cloud provider's ability to borrow; the borrowed money finances data centres and accelerators; and the AI company gains access to computing capacity that its current operating cash flow might not support.

There is nothing inherently suspicious about this. Aircraft manufacturers help customers finance aircraft. Equipment manufacturers lease machinery. Property developers build against commitments from future tenants. Capital markets exist in part because productive investment often has to precede the revenue that eventually pays for it.

AI is unusual less because these arrangements exist than because they have become so dense. The same company can occupy several places in the financing structure at once, making the boundary between supplier, investor and customer increasingly porous.

Nvidia and CoreWeave: Supplier, Investor and Backstop Buyer

Nvidia and CoreWeave provide an unusually clear example. In January 2026, Nvidia invested another $2 billion in CoreWeave as the companies expanded a partnership aimed at building more than five gigawatts of AI infrastructure by 2030. CoreWeave's platform is built heavily around Nvidia technology, and the expanded agreement further aligned their hardware, software and infrastructure plans.

The relationship goes further than an equity investment. Under a separate $6.3 billion agreement disclosed to the SEC in September 2025, Nvidia gained access to residual CoreWeave computing capacity and, subject to the agreement's conditions, is obligated to purchase capacity that CoreWeave cannot sell to other customers through April 2032.

Nvidia therefore occupies several positions in the same economic network. It supplies crucial equipment, invests in the infrastructure company buying that equipment and acts as a backstop buyer for some of the resulting computing capacity. None of those roles is irrational. Nvidia benefits if a fast-growing cloud provider can expand more quickly; CoreWeave gains capital and some protection against unsold capacity; and customers gain access to infrastructure that might otherwise take longer to build.

The arrangement nevertheless shows why company-by-company analysis can miss something important. The supplier helps finance the customer, the customer's expansion generates demand for the supplier, and the supplier provides support against some of the risk that the resulting capacity cannot immediately find another buyer. The transactions are real. Their economic independence is much less clear.

When One Assumption Appears Several Times

Consider the same mechanism in more general form. An AI laboratory expects enormous future demand, and investors provide capital on the strength of that expectation. The laboratory signs a long-term infrastructure contract. The infrastructure provider can now show a large backlog of future revenue, making lenders more comfortable financing construction. It buys accelerators, semiconductor companies report higher sales, and investors interpret those sales as further evidence that AI demand remains exceptional. Higher valuations make another round of investment easier.

Nothing in this sequence requires accounting manipulation. The contracts can be enforceable, the data centres can be built, the chips delivered and the semiconductor revenue entirely genuine. The difficulty is that the AI laboratory's valuation, the cloud company's backlog, the lender's confidence and the semiconductor company's sales may all depend partly on the same proposition: future customers will pay enough for AI services to support the infrastructure being constructed today.

A backlog therefore provides evidence of demand, but not necessarily demand independent of the financing environment that helped create it. GPU sales are real economic activity, but they do not by themselves establish that the ultimate applications running on those GPUs will earn enough to justify their cost. Rising valuations may reflect genuine progress while simultaneously making continued expansion easier to finance.

This is the more useful meaning of circularity. The problem is not that money literally travels around a closed loop and somehow becomes profit. It is that one expectation can appear several times in different financial forms and gradually begin to look like several independent pieces of evidence.

Real Revenue Can Still Produce Bad Returns

This is why simple comparisons with the dot-com bubble are unsatisfying. Today's leading technology companies generate enormous revenues and profits. AI services have real users, businesses pay for them, and computing equipment is being purchased because somebody genuinely wants computing power. The industry is not merely selling stories.

Real demand, however, says little by itself about the return on investment. Imagine a railway between two growing cities. It can carry genuine passengers, generate rising ticket revenue and transform the regional economy. None of that would prove that constructing twenty parallel railway lines was sensible.

Technology and investment are separate questions.

AI could become one of the most important technologies developed this century while a substantial portion of today's infrastructure investment still earns disappointing returns. The internet was transformative and telecommunications networks became indispensable, yet investors in the late 1990s still demonstrated that valuable infrastructure could be built too quickly and financed at prices that assumed more profitable demand than eventually appeared. AI does not need to repeat that episode closely for the underlying lesson to hold. Whether the capacity proves useful and whether its owners earn an adequate return are different questions.

How the Loop Reverses

The network becomes more revealing when expectations weaken. Assume AI remains useful and adoption keeps growing. There is no technological collapse. Enterprise deployment simply develops more slowly than anticipated, competition pushes model prices down, or customers discover that many applications are useful without being valuable enough to support the margins assumed when the infrastructure was financed.

An AI laboratory reduces planned compute purchases. Infrastructure providers lower utilisation forecasts. Financing becomes more expensive and some expansion projects are postponed. Accelerator orders slow, semiconductor growth expectations weaken, and companies that had been willing to finance customers become more selective. Capital becomes more expensive for firms that were relying on those relationships, encouraging further reductions in infrastructure commitments. The connections that accelerated construction now transmit disappointment in the opposite direction.

This outcome is not automatic. AI compute is much more flexible than a single-purpose industrial plant. Capacity may be redirected from one model developer to another, from training towards inference, or towards enterprise and scientific workloads. Losing one customer does not necessarily render a data centre useless. The financial question is whether replacement demand exists at a price high enough to support the assumptions under which the capacity was financed. A GPU cluster can remain technically productive while becoming a poor investment if rental prices fall far enough.

Efficiency creates another ambiguity. If models suddenly require much less computing power for the same work, existing infrastructure could face pressure. Yet lower costs may stimulate so much additional usage that overall demand for compute rises. The vulnerability therefore lies less in any single technological outcome than in the distance between the outcome that actually occurs and the one already embedded in contracts, valuations and financing structures.

The Weak Links Do Not Have to Break

Looking at Nvidia, Microsoft, Alphabet, Amazon or Meta can create an impression of formidable financial resilience, and rightly so. These companies possess balance sheets that make crude comparisons with earlier technology booms misleading. A moderate downturn in AI investment would not suddenly leave Nvidia unable to operate or the major hyperscalers insolvent.

But the system does not have to fail at its strongest nodes. A specialist infrastructure provider whose expansion depends on repeated financing faces a different set of constraints from a hyperscaler funding data centres from diversified operating cash flow. An AI laboratory making large future compute commitments while still raising outside capital occupies another position again. Smaller participants can matter disproportionately when other companies have planned capacity around their promised purchases.

CoreWeave illustrates the tension. At the end of June 2026 it reported approximately $104 billion of revenue backlog, excluding more than $25 billion of additional customer commitments signed early in the following quarter. At the same time, the company raised its expected 2026 capital spending to $35–39 billion.

Both figures can be signs of extraordinary opportunity. Together, they also show how much future performance has already been pulled into present investment decisions. CoreWeave does not have to fail for this to matter. If an infrastructure provider plans ten facilities and ultimately needs seven, the missing three represent orders that equipment manufacturers, utilities and construction companies expected but will not receive. Across an entire industry, such revisions can produce a sharp investment downturn while underlying AI use continues to grow.

The critical variable is therefore not simply whether demand exists. It is whether demand arrives at something close to the price, scale and speed assumed when the contracts were signed and the assets financed.

Where the Money Has to Come From

The AI economy is filled with impressive numbers: gigawatts of proposed data centres, hundreds of billions of dollars of capital spending, huge compute commitments, rapidly growing semiconductor sales and private-company valuations that would have seemed implausible only a few years ago.

The less spectacular number may matter more in the long run: how much cash enters the system from customers and businesses whose ability to pay does not itself depend on another round of AI financing?

Consumers paying for subscriptions contribute to that pool. So do businesses paying for AI because it lowers their costs or increases the value of their output, advertisers paying to reach valuable audiences, pharmaceutical companies purchasing useful research tools, and diversified technology companies funding AI infrastructure from profits earned elsewhere. New investment can bridge the period between construction and mature demand, but eventually the assets have to produce economic value sufficient to compensate whoever supplied the capital.

This does not require every individual AI company to become independently self-financing. Modern economies routinely support enormous capital-intensive industries through debt, equity markets and cross-subsidisation. It does require enough economic surplus to emerge outside the financing loop. The more expansion can eventually be supported by operating cash flows rather than repeated revaluation and fresh capital, the less dependent the system becomes on continued confidence in the same forecast.

Circular Confidence

None of this demonstrates that AI is a bubble. That framing is too blunt for the problem. A technology can be transformative, demand can grow for decades, and investors can still pay too much for the infrastructure required to serve it.

The more interesting question is how many apparently separate investment decisions depend on the same underlying forecast. Semiconductor companies, cloud providers, model developers and data-centre operators occupy different parts of the AI economy, but their fortunes are not fully independent if all of them require profitable AI demand to expand at roughly the pace now assumed.

The financial architecture developing around AI has an obvious virtue. It can move enormous quantities of capital towards computing infrastructure much faster than waiting for each generation of technology to finance the next from operating profits. If current expectations prove broadly correct, that mechanism may be one reason the industry develops so rapidly.

Acceleration also magnifies forecasting errors. When suppliers invest in customers, customers make commitments to infrastructure providers, infrastructure providers borrow against those commitments, and investors interpret the resulting equipment purchases as fresh confirmation of demand, the network can contain more risk than any individual transaction appears to create. Circular financing is neither inherently fraudulent nor inherently foolish. It is an amplifier.

AI could succeed technologically. Usage could continue rising. Productivity could improve. Much of the infrastructure being constructed today could remain useful for decades.

Investors could still discover that they built too much of it, too quickly, at prices that assumed a future more profitable than the one that arrived.

The problem is not circular money. It is circular confidence.