The Indispensability Trap Is Coming Due

First came the promises. Hundreds of billions of dollars. Gigawatts of capacity. Enough chips to power an AI future that everyone assured us was inevitable.

Now comes the business of paying for it.

The Wall Street Journal reports that Broadcom is discussing more than $50 billion in financing for its custom-chip program with OpenAI. Oracle is talking to financiers about funding a major chip purchase. SpaceX reportedly wants $40 billion for Nvidia chips. Apollo, Blackstone and Goldman Sachs are among the firms involved in the discussions.

These deals have not closed. But the search for financing brings an argument I made in The Indispensability Trap into sharper focus. Oracle and Broadcom have committed themselves to enormous projects with OpenAI. The question is whether the economics of those commitments will reward them as handsomely as AI’s growing importance suggests.

AI can fulfill its promise while the companies funding its infrastructure struggle to collect the payoff. That possibility is the trap.

Growth Has a Financing Requirement

Oracle’s business is growing. Its latest quarterly results show cloud infrastructure revenue more than doubling. The company is delivering capacity, signing contracts and collecting money.

It is also spending at a rate that requires a considerable financing effort.

For its first fiscal quarter, ended August 31, Oracle reported $23.1 billion in operating cash flow against $28.5 billion in capital expenditures. Free cash flow was negative by approximately $5.4 billion. Its balance sheet carried approximately $125.3 billion in borrowings.

Even that strong operating cash flow deserves a closer look. It included $11.4 billion from customer prepayments with a significant financing component. Those dollars help fund construction today, but Oracle owes the services those customers purchased. The company also raised $20 billion through an equity program.

This is a profitable enterprise mobilizing multiple sources of capital to support an enormous expansion. It is also evidence of how demanding that expansion has become.

The proposed chip-financing arrangement would reportedly put the hardware in a separate company financed by investors, with Oracle leasing it over time. That can help align payments with revenue. It still leaves Oracle needing enough income to meet its obligations.

A different owner on the paperwork does not make the bill disappear.

Even Deep Pockets Have Limits

Broadcom enters this discussion from a different position. Its latest quarter produced approximately $13.7 billion in free cash flow, and it ended the period with roughly $24 billion in cash. This is a company with substantial financial resources.

Yet facilitating AI infrastructure can involve more than designing and selling chips.

Broadcom’s latest 10-Q describes an existing arrangement in which a financial partner acquired AI-rack purchase agreements and related customer leases. Broadcom provided a backstop for the customer’s lease obligations. Maximum potential liability, once all relevant racks are deployed, was approximately $29 billion.

That is contingent exposure, not a loss already incurred. Broadcom reported no payments under the backstop and assessed its fair value as immaterial. The filing also does not identify that customer as OpenAI, so we should keep this arrangement separate from the newly reported OpenAI financing.

Nevertheless, the mechanism matters. If the customer defaults, Broadcom’s exposure depends partly on what the hardware brings when sold. The supplier has a stake in the customer’s ability to pay and the equipment’s resale value.

That connects technology economics directly to credit exposure.

OpenAI’s announced programs already establish the scale: 10 gigawatts with Broadcom and 4.5 gigawatts of additional Stargate capacity with Oracle. Capacity announcements do not tell us each company’s net liability. They do tell us these relationships require exceptional amounts of infrastructure and money.

OpenAI wants control over its compute and lower operating costs. Its partners want growing sales and durable returns. Those objectives can align. They can also become harder to reconcile as the cost of delivering intelligence changes.

The Trap Tightens as AI Improves

Here is the part I keep coming back to. The more indispensable AI becomes, the greater the incentive to make it cheaper, more efficient and available wherever people need it.

Every business that starts depending on intelligence has a reason to reduce its cost. Every supplier has a reason to offer a less expensive alternative. Every device manufacturer has a reason to put useful AI directly into its products.

That creates pressure on the infrastructure economics underpinning today’s commitments.

Moving AI onto phones and devices at the edge can itself make intelligence cheaper and more accessible. Local execution can reduce the need to send every task to a distant server, while improving responsiveness and keeping some data close to its source. Those advantages can encourage more use.

This matters when we invoke Jevons paradox. Better efficiency can expand consumption. But AI has a deployment flexibility that the steam engines associated with that argument did not possess. A steam engine was not going into your watch. Useful AI capability can.

The growth in consumption can therefore occur across a changing mix of infrastructure. More AI tasks do not automatically translate into proportionately more revenue for the centralized facilities being financed today.

Custom silicon adds another pressure. It can improve the economics of a particular deployment and reward the companies supplying it. It can also raise the competitive standard, forcing other operators to deliver more intelligence for less money. Better models and software can push in the same direction.

The financial consequence need not be an empty data center. A busy facility can still earn disappointing returns if prices, utilization and replacement costs fail to support the investment.

Meanwhile, businesses using AI can capture value through better services, proprietary data, customer relationships and workflows. That is what moving up the stack means in practice. Intelligence becomes a more widely available input. The payoff increasingly depends on what someone does with it.

Oracle has applications and enterprise relationships. Broadcom has software and a diversified semiconductor business. Both can adapt. But adaptation does not cancel commitments made to a particular buildout.

There is a strong defense of these investments. Training advanced models requires centralized compute. Demanding inference workloads do too. Growing usage can absorb efficiency gains, and contracts can provide meaningful protection. Cheaper custom chips can make facilities more competitive.

I accept those arguments. They still leave the financial question unresolved: will these assets generate sufficient returns, on the required timetable, to support the obligations attached to them?

Financing cannot settle that question. It can provide time, allocate risk and make deployment possible. The underlying business still has to produce the money.

The reported SpaceX effort suggests this funding challenge extends beyond Oracle and Broadcom. But we should resist treating every negotiation as proof of distress. Borrowing can be sensible. Outside financing can support valuable infrastructure. The issue is what assumptions make those commitments work, and how much room exists if the assumptions change.

The indispensability trap does not require AI demand to collapse. It can unfold while adoption grows, capabilities improve and intelligence becomes part of everyday life.

Those developments can strengthen the businesses applying AI while squeezing the returns of companies that committed enormous capital to supplying it. Financial payments follow their schedules. Technology follows its own.

AI can become indispensable while the infrastructure built to deliver it struggles to pay its bills.

That is the trap in action.