Halloween: The scariest balance sheets in AI

Nvidia's market capitalization passed $5 trillion on 29 October, making it the world's most valuable company. By Halloween the celebrations were in full swing. Underneath the milestone sits a web of circular customer financing and debt-funded spending commitments by the hyperscalers, and it makes for the most frightening balance sheet in technology.

The climb has been steep even by the standards of this cycle. Nvidia first touched a valuation of $1 trillion in May 2023, passed $4 trillion in July of this year and added the last trillion in less than four months. The final push came after Jensen Huang, its chief executive, told a conference in Washington on 28 October that the company had $500 billion of bookings for its Blackwell and Rubin chips through the end of 2026. Investors took that as proof that demand is locked in for years.

Customers borrowing to buy

A $5 trillion valuation assumes that hyperscaler capital spending keeps growing at fifty per cent a year for the foreseeable future. The customers paying for it, Microsoft, Meta, Alphabet and the private AI labs, increasingly fund their chip purchases with large bond issues and circular vendor equity deals. If AI software keeps earning little on the capital invested, the build-out could stop abruptly.

This week's results showed the bill rising. Microsoft, Alphabet and Meta all reported on 29 October and all pointed to heavier capital spending ahead. Meta told shareholders that spending would grow notably faster next year, its shares fell sharply the following day, and on 30 October it sold $30 billion of bonds, one of the largest corporate bond deals on record. It had already moved the cost of its Hyperion data centre in Louisiana into a joint venture with Blue Owl, funded with private debt that sits off its own balance sheet. Oracle sold $18 billion of bonds in September for the same purpose.

Money that goes round in circles

The more unsettling financing runs through the suppliers themselves. In September Nvidia signed a letter of intent to invest up to $100 billion in OpenAI, which in turn plans to fill its data centres with Nvidia systems. In October OpenAI agreed to buy chips from AMD, which granted it warrants over up to a tenth of AMD's shares. Nvidia also holds a stake in CoreWeave, a cloud operator that buys its chips, and has agreed to take any capacity CoreWeave cannot sell. OpenAI has committed to buy computing power from Oracle on a scale that dwarfs its own revenue.

The loop works like this. A supplier invests cash or shares in a customer, the customer uses that money and its borrowing capacity to buy the supplier's products, and the supplier books the sale as revenue. Each sale is real, but part of it is funded by the seller's own balance sheet, and the customer's ability to keep ordering depends on raising fresh money. When that money stops, orders stop with it, and the supplier's revenue and its investment lose value together.

What the telecom boom taught

This has happened before. In the late 1990s equipment makers such as Lucent and Nortel lent heavily to young telecom carriers so that they could buy switches and fibre, and Cisco briefly became the most valuable company in the world in March 2000. When the carriers' access to credit dried up, orders collapsed, the vendor loans went bad and Cisco lost most of its market value over the following two years. Internet traffic kept growing throughout. The technology was sound; the financing that pulled demand forward was the weak point.

A tenth of the index

Nvidia now accounts for ten per cent of total US equity index capitalization, which leaves the whole market exposed to a single company. If the hyperscalers pause spending to absorb the capacity they already have, its valuation multiple could shrink on a historic scale.

That concentration also reaches savers who never chose to own a chipmaker. Index funds and pension schemes that track the broad market hold Nvidia in proportion to its size, so every rise in its price increases their exposure automatically. The hyperscalers that buy its chips are among the next-largest companies in the same indices, which means a downturn in AI spending would hit the buyer and the seller in the same portfolio at the same moment.

The case for Nvidia

The defence deserves a fair hearing. Nvidia is highly profitable, carries little debt and has kept growing earnings fast enough that its valuation multiple is far below the levels Cisco reached in 2000. Its largest customers generate enormous cash from advertising and cloud software and could keep spending for some time before their own finances came under strain. None of the big four hyperscalers looks close to financial distress.

The weakness lies at the margin. The fastest-growing sources of demand are the buyers with the thinnest finances: OpenAI and other loss-making labs, specialist clouds such as CoreWeave that borrow in the high-yield market, and Oracle, which came into the boom with heavy debt. These are the customers most dependent on circular deals and borrowed money, and they are the first that would have to cut back if credit markets turned.

Signals to watch

Nvidia's own quarterly results later in November will almost certainly look strong, so they will reveal little. More telling will be the hyperscalers' capital spending guidance with their fourth-quarter results early next year, the spreads on new technology bonds, the cost of insuring Oracle's debt against default, and the terms of OpenAI's next fundraising. A slowdown in any of these would reach Nvidia's order book within a few quarters. Nvidia at $5 trillion marks the high point of the AI boom. When the hyperscalers eventually take a breather, the circular financing beneath it means the pain will spread well beyond one chipmaker to equity markets everywhere.