Money & infrastructure

The Demand Behind the Demand

October 2026
11 min

A chip supplier is lending its future largest customer money to lease computing capacity built around its technology. The arrangement may be rational and still make the boom harder to read.

Broadcom has agreed to lend Anthropic as much as $42 billion.

Anthropic can use the facility to finance part of a five-year lease for computing capacity built around tensor processing units, or TPUs, developed through the relationship between Broadcom and Google. The total lease commitment is worth approximately $125.2 billion. The debt may eventually be converted into shares in Anthropic.

By 2027, Anthropic is expected to become Broadcom’s largest compute customer. Broadcom will therefore help finance a company that is expected to buy more of its computing technology than anyone else.

There is nothing inherently improper about this. Suppliers have financed customers for a long time. Aircraft manufacturers support airline purchases. Industrial companies arrange credit for expensive machinery. Car companies operate financing divisions because most customers cannot purchase vehicles outright.

Financing can make a market possible before its customers have accumulated enough capital to fund themselves. It can also signal confidence: a supplier willing to place its own balance sheet behind a customer may understand the technology, equipment and commercial opportunity better than a distant lender.

But supplier financing changes the meaning of a sale.

If a customer independently raises money and chooses to buy a product, the order contains information about demand. If the supplier provides the money, guarantees the financing or invests in the customer while receiving a purchase commitment in return, the same order becomes harder to interpret. The product may still be needed. The customer may still succeed. But the financial capacity to buy and the decision to supply are no longer separate events.

The market begins generating some of its own demand.

The roles are collapsing

The Anthropic–Broadcom arrangement is unusually large, but its underlying structure is becoming common across AI.

AI developers need access to chips, data centres, electricity and cloud platforms. The companies controlling those resources want to secure large, dependable customers. Both sides are making decisions years before anyone can know precisely how much AI consumers and businesses will use.

The result is a network in which companies occupy several roles simultaneously.

Amazon and Google are investors in Anthropic. They supply it with computing infrastructure and distribute Claude to customers through their cloud marketplaces. They also develop competing AI systems. In 2025, 47% of Anthropic’s sales passed through Amazon and Google, according to its confidential prospectus, while Anthropic paid the platforms approximately $351 million in distribution fees.

Broadcom supplies computing technology, supports equipment leases and now proposes to lend money that may convert into equity. Anthropic buys capacity, helps justify the construction of more capacity and may provide Broadcom with a substantial financial return if its valuation rises.

The categories remain visible in legal agreements. In economic terms, they are becoming entangled.

This matters because each role normally provides a partial check on the others. A lender asks whether the borrower can repay. An investor asks whether the company can grow. A supplier asks whether the customer will pay. A customer compares competing suppliers. A distributor decides whether there is genuine demand from end users.

When one company occupies several of these positions, the relationship can become more efficient. It can also become less independent.

The supplier wants the customer to expand because expansion creates equipment sales. The investor wants the valuation to rise. The lender wants the borrower to remain solvent. The customer needs the supplier to keep providing scarce infrastructure. Decisions that appear to come from different parts of the market may therefore rest on the same commercial belief.

That belief is that future demand for AI will be large enough to pay for the infrastructure being committed today.

Circular does not mean fictitious

The phrase circular financing can sound accusatory. It suggests money being passed around to create the appearance of business.

That is too simple for what is happening.

Anthropic has substantial external revenue. In April, it said its annualised revenue run-rate had exceeded $30 billion, compared with approximately $9 billion at the end of 2025. It reported that the number of business customers spending more than $1 million a year had doubled to more than 1,000 in less than two months. These are company-provided figures rather than audited proof of long-term profitability, but they indicate that the infrastructure is not being built in the absence of customers.

Nor is vertical coordination irrational when an important input is scarce. Anthropic believes access to computing power, rather than customer demand, may become the principal limit on its growth. Waiting until every future customer has arrived before reserving capacity would leave the company unable to serve them.

Broadcom faces a related problem. Developing custom chips and networking systems requires large commitments before equipment is installed. A long-term customer allows it to plan manufacturing and attract financial partners.

Circular relationships can therefore perform useful work. They align investment schedules across companies that would otherwise hesitate to move first. The chipmaker gains confidence that someone will use the equipment. The AI developer gains confidence that the equipment will exist. Investors receive long contracts against which they can lend.

The danger is not that the circle contains no real economic activity. It is that the circle can obscure where independent demand ends and mutually supported expectation begins.

The Bank for International Settlements examined investment involving 1,246 AI-producing companies. It found that 55.2% of incoming investment in AI firms between 2021 and 2025 came from other AI companies. Among AI-to-AI investments, transactions representing 46.4% of value also connected companies with a commercial supply relationship. The BIS concluded that these links can help secure critical inputs and address information gaps, while also increasing opacity and macroeconomic risk.

That is the more useful interpretation. Circularity is neither proof of fraud nor a harmless accounting curiosity. It changes the evidential value of the market’s own signals.

An order is no longer just an order

Imagine that a chipmaker announces a large increase in orders.

An investor might reasonably conclude that customers expect strong demand. The orders suggest confidence further down the supply chain.

But suppose the chipmaker has also invested in those customers, guaranteed their lease payments or lent them money for purchases. The orders now tell us several things at once. Customers want capacity, but the supplier wants them to have the capacity. The supplier believes future sales justify taking credit risk. Outside financiers may have participated because the supplier agreed to absorb part of the loss if the customer fails.

None of those facts cancels the order. They change the questions that should follow it.

Who provided the purchasing power? Would an independent lender have offered the same terms? Who carries the loss if utilisation is lower than expected? Can the equipment be sold to another customer? Does the borrower generate enough cash from outside the AI investment network to service the obligation?

Broadcom’s public filings show how risk can travel through such a structure. In June it arranged for Apollo to acquire agreements involving Broadcom-designed AI racks and the related customer leases. Broadcom then provided a backstop for the customer’s five-year lease obligations, with maximum exposure of $29 billion. If the customer defaults, Broadcom can assume the lease or arrange a sale of the equipment.

The financing sits with an investment partner, but some of the economic risk returns to the supplier.

This is one reason balance sheets alone can provide an incomplete picture. A company may not directly own all the debt financing infrastructure built around its products, yet guarantees and backstops can reconnect it to losses. The legal entity holding an asset is not always the entity carrying the final risk.

A chain built on future usage

Anthropic’s infrastructure plan makes the scale of the commitment visible. The company expects to spend at least $518 billion with six partners over approximately a decade. About 80% of those commitments are non-cancellable or payable regardless of how much capacity Anthropic eventually uses.

Its disclosed obligations include at least $111.1 billion to Google, $110 billion to Amazon, $31.4 billion to Microsoft and approximately $161.2 billion in Broadcom-related equipment leases.

These commitments could prove farsighted. Scarce capacity reserved early may become enormously valuable if AI usage continues growing. The same contracts could become burdensome if models become more efficient, competing infrastructure becomes cheaper or demand grows more slowly than expected.

The financial structure therefore brings the future into the present. Suppliers build against long-term commitments. Investors lend against expected lease payments. AI companies secure the resources needed to pursue future revenue. Present valuations incorporate the belief that all of this capacity will eventually find productive use.

The problem is not simply leverage. It is correlated belief.

A chip supplier, data-centre developer, private-credit fund and AI laboratory may appear to represent different sectors and investment decisions. Yet each may depend on the same forecast of AI usage. Diversification by corporate name does not help much when the cash flows share one underlying assumption.

BIS researchers modelled the AI build-out as a race for a small number of dominant market positions. Their study suggests that companies may rationally overcommit because none can afford to wait while a competitor secures the scarce capacity first. Debt and circular ownership can then deepen losses if expectations fail, particularly because specialised equipment may have to be sold into a falling market at the same time.

This remains a model, not a forecast that a crash is imminent. Its useful insight is about incentives. Every company can make a defensible decision individually while the system collectively builds too much.

Financing is becoming a competitive capability

The arrangements also affect who can compete.

A frontier AI company no longer needs only excellent researchers, models and products. It needs access to vast amounts of capital and the confidence of infrastructure providers willing to reserve equipment years in advance.

A chipmaker’s balance sheet consequently becomes part of what it sells. It can offer not only processors but the financial means to acquire them. That gives established suppliers a way to strengthen selected customers and make their technology more difficult to displace.

The same is true in reverse. A large cloud company can invest in an AI developer, distribute its model and offer the computing infrastructure on which it runs. The AI company gains access to customers and capacity. The cloud provider receives revenue at several points and reinforces demand for its own platform.

Smaller participants may struggle to enter this system even if their technology is strong. They lack the long-term commitments needed to reserve capacity, while infrastructure providers may prefer companies in which they already have a financial interest. Competition gradually becomes a contest not only between products but between financed ecosystems.

That does not automatically harm customers. Coordinated ecosystems can lower costs and accelerate deployment. But they can also reduce practical choice. An enterprise may believe it is choosing among AI models while each option arrives bundled with a cloud platform, chip architecture, distribution agreement and network of financial commitments.

The visible product is the model. The operating system beneath it is capital.

The public signal problem

This eventually becomes a problem for institutions beyond the companies themselves.

Investors need to decide whether rising equipment orders reflect durable end-user demand. Lenders need to understand whether the cash flows supporting their debt come from external customers or other companies inside the same network. Competition authorities need to consider whether financing arrangements are securing necessary supply or enclosing the market around a few preferred participants.

Governments face a similar difficulty. AI capital expenditure is contributing to economic growth, electricity demand and infrastructure construction. If policymakers treat every investment commitment as evidence of proven future demand, they may overbuild supporting infrastructure. If they dismiss it all as speculative, they may fail to prepare for a genuine technological transition.

The distinction cannot be made by looking at the headline size of the investment. It requires tracing the money through the system.

Who invested in whom? What purchases were promised in return? Which commitments remain payable regardless of use? Who guaranteed the debt? Where does revenue from people and businesses outside the network enter?

These questions are less dramatic than asking whether AI is a bubble. They are also more answerable.

What the loan really tells us

Broadcom’s proposed $42 billion facility is evidence of conviction. The company is willing to place enormous financial resources behind the proposition that Anthropic will need its computing technology and generate enough value to support the commitment.

It is also evidence of dependency. Anthropic needs the capacity. Broadcom needs the demand. Investors need the leases. Each participant’s confidence helps justify the confidence of the others.

The arrangement may work extremely well. Anthropic’s revenue may grow quickly enough that today’s commitments appear conservative. Scarce computing capacity may retain strong value even if one customer stumbles. Supplier support may prove to have accelerated infrastructure that conventional finance was too cautious to fund.

But the deal cannot be read like an ordinary customer order. Part of the demand and part of the financing come from the same relationship.

That is the deeper development behind the $42 billion figure. AI companies are not merely competing within a market. The largest of them are building the market’s supply, credit and distribution structures around one another.

Broadcom is helping Anthropic buy access to Broadcom-linked technology because both companies believe future customers will ultimately validate the investment.

The next useful number is therefore not the value of the lease.

It is how much money eventually enters the circle from outside it.