Nvidia's $57bn quarter doesn't end the bubble debate

Nvidia reported $57 billion of revenue for the third quarter of fiscal 2026, another huge number and comfortably ahead of Wall Street's consensus forecast. The market shrugged. The shares ended the session roughly flat, and the argument among institutional investors about a bubble in AI capital spending carried on as before.

The figures themselves left little to criticise. Revenue rose by 62 per cent on a year earlier, data-centre sales supplied the great bulk of it, and the company guided to about $65 billion for the current quarter. At the end of October Nvidia became the first company to be valued at $5 trillion, and a report of this quality was what that valuation required rather than a reason to raise it.

A beat already in the price

Nvidia's flawless execution was already built into its multi-trillion-dollar valuation. What the report could not supply, and what portfolio managers most want to see, is hard evidence that Nvidia's hyperscaler customers are making money from AI software. Selling $57 billion of hardware in a quarter shows that the hyperscalers are spending heavily. It says nothing about whether their enterprise AI models earn lasting profits.

Management repeated that it has visibility into $500 billion of revenue from its Blackwell and Rubin chips through the end of 2026, a figure first given at a conference in Washington in October. Gross margins stayed above 70 per cent, which shows that Nvidia still sets its own prices. Neither point addresses the doubt, because both describe what customers have agreed to buy, and the question is what those purchases will earn.

Growth without China

The guidance also assumes almost nothing from China. American export controls and Beijing's discouragement of purchases have shut Nvidia out of what was once a large market for its data-centre chips, and the company has said its outlook includes no such sales there. That cuts both ways. It shows that demand elsewhere is strong enough to absorb the loss, and it leaves a source of upside if Washington and Beijing ever agree terms. It also concentrates Nvidia's dependence on a handful of American buyers.

Who is paying for the chips

The source of the money matters as much as the amount. Nvidia's largest buyers are the cloud giants, which fund their purchases from advertising and software profits, and increasingly from bond issues. A growing share of demand, however, comes from AI laboratories and specialist cloud firms that are themselves losing money and relying on outside capital. Nvidia has also signed a letter of intent to invest up to $100 billion in OpenAI, one of the largest end users of its chips, which has led sceptics to ask how much of the demand is financed by the supplier itself.

That kind of circularity is not proof of a bubble, since vendor financing is common in capital-goods industries. It does mean that Nvidia's order book depends on credit and equity markets staying open to its customers. If borrowing costs rise or equity investors lose patience with loss-making AI firms, orders could slow even while end demand for AI services keeps growing.

Credit markets have already started to charge more. Oracle, which has promised OpenAI enormous amounts of computing capacity, has seen the cost of insuring its debt against default rise sharply this month, and Meta's $30 billion bond sale at the end of October was followed by heavy issuance from other technology companies. Each of those borrowers is, directly or through its tenants, a buyer of Nvidia's chips.

Depreciation comes due

As the hyperscalers expand their server fleets, the depreciation charges on their income statements will climb fast over the next twenty-four months. If enterprise software revenue does not arrive, capital spending budgets will be cut sharply.

The accounting choices make this more sensitive than it looks. Over the past few years most of the big cloud operators lengthened the assumed useful lives of their servers, which spread the cost over more years and lifted reported profits. Amazon has already shortened those lives again for part of its fleet, citing the faster pace of AI hardware. Each new Nvidia generation arrives roughly a year after the last, so if older chips lose value faster than the accounts assume, the hyperscalers will face write-downs on top of rising depreciation, and pressure from their own shareholders to spend less.

This has become the bears' favourite argument. Michael Burry, the investor known for betting against mortgage bonds before 2008, argued publicly this month that the hyperscalers are understating depreciation by stretching server lives. Whether or not he is right about the size of the gap, the argument has moved the debate from Nvidia's revenue, which is plainly strong, to its customers' profits, which are harder to judge.

The Cisco comparison

Bears reach for Cisco in 2000. Its sales kept rising for months after its shares peaked in March of that year, and the collapse came only when its telecom customers, many funded by debt, stopped ordering. The comparison is imperfect: Nvidia's biggest customers are among the most profitable companies in the world, and its own margins are far higher than Cisco's were. What the episode does show is that a supplier's results can look strongest just before its customers' spending turns.

The evidence will come from the customers rather than from Nvidia. Hyperscalers' fourth-quarter results early next year, their capital spending guidance, and any disclosure of revenue earned from AI services will matter more than Nvidia's own guidance. Nvidia's $57 billion quarter shows its own business firing on all cylinders. Whether its customers will earn a return on all that compute is another matter, and the answer will probably not come until depreciation starts to bite over the next two years. Until then the bubble debate will go on.