The AI money loop

Look closely at the venture deals financing the artificial intelligence boom and a circular pattern appears, one that recalls the worst speculative habits of the dot-com era. Call it the vendor equity money loop. It is spreading fast.

The deals of the past two months show its scale. In September Nvidia said it would invest up to $100 billion in OpenAI, in stages, as the lab deploys Nvidia systems in new data centres. In early October AMD agreed to supply OpenAI with chips and granted it warrants to buy up to 160 million AMD shares for a nominal price, vesting as purchases and AMD's share price hit agreed milestones. OpenAI has also signed a large, multi-year computing contract with Oracle and a deal with Broadcom to develop its own chips. The customer at the centre of all these commitments is a company that is still losing money.

Chipmakers fund their customers

It works like this. A dominant hardware supplier such as Nvidia puts hundreds of millions into a richly valued funding round for a private frontier lab such as OpenAI. The lab then signs binding multi-billion-dollar contracts to buy its accelerators only from the investor. Chip designers also hand large stock warrants to enterprise customers in exchange for commitments to buy compute in future.

The loop extends beyond the labs. Nvidia holds a stake in CoreWeave, a cloud provider that rents out Nvidia chips, and has agreed to buy capacity CoreWeave cannot sell to others. The chipmaker thus stands on several sides of the same transaction: supplier of the hardware, shareholder in the buyer and, in some cases, customer of last resort for the output. Each role is defensible on its own. Together they make it hard for outside investors to tell how much demand is independent of Nvidia's own balance sheet.

Lucent and the telecom bust

The closest precedent is the telecoms boom of the late 1990s. Equipment makers such as Lucent and Nortel lent billions of dollars to young telecoms carriers so they could buy the vendors' own switches and fibre-optic gear. Reported sales soared. When the carriers' business plans failed in 2000 and 2001, many of those loans went bad, the equipment was dumped on the second-hand market, and the vendors had to write off both the loans and the revenue growth they had supported. Some of the carriers also swapped network capacity with one another and booked the swaps as revenue, which flattered sales further until regulators and auditors caught up.

The differences matter. Nvidia's customers today are led by the largest and most profitable companies in the world, and its deals take the form of equity rather than loans. Equity cannot default in the way a loan can. It can, however, lose most of its value, and a vendor holding shares in customers whose valuations depend on its own chip sales is exposed to the same reflexive dynamic.

Revenue of lower quality

The arrangement speeds up reported revenue growth and makes demand look stronger than it is. It also lowers the quality of earnings, because a meaningful share of reported chip sales is paid for with the vendor's own investment capital rather than with cash its customers generated on their own.

The accounting does not flag the problem. A chip sale to a customer in which the vendor has invested is recorded as revenue like any other, and the investment sits separately on the balance sheet. An analyst has to reconcile the two to see how much of the growth was self-financed. Warrants create a similar effect in reverse: the customer's reward for buying is a claim on the vendor's future share price, which rises as the sales it enables are reported.

Who carries the risk

The chipmakers' shareholders carry the first layer, through the investments on their balance sheets and the risk to future orders. Lenders financing data-centre construction carry the next, because many projects are funded with debt secured on long-term leases to AI labs whose ability to pay depends on continued fund-raising. The labs' own investors carry the rest. Electricity utilities and builders that are adding capacity on the strength of these commitments are exposed if the projects are delayed or cancelled.

The case for the loop

The strongest argument in its defence is that demand for AI computing is real and exceeds supply. Most of Nvidia's revenue comes from large cloud companies such as Microsoft, Alphabet, Amazon and Meta, which pay from enormous operating cash flows, and the vendor-financed deals are a smaller share. Investing in customers is also a rational way for a supplier to secure demand in a market where capacity takes years to build. All of that is true. The concern is at the margin, where the fastest-growing commitments rely on customers whose spending plans far exceed their revenues.

The signals to watch are the capital-spending guidance of the big cloud companies in their results at the end of this month, Nvidia's own results in November, the terms of OpenAI's next fund-raising, and borrowing costs for data-centre projects. When chipmakers finance their own orders, their reported sales rest on their own balance sheets. The loop will hold while the labs' valuations keep rising. When they stop, the orders and the earnings built on them will probably shrink together. The first sign of trouble is likely to be a delayed data-centre project or a down-round at a leading lab, and chipmakers with the largest investments in their own customers will be the first to feel it.