Nvidia's boom: peak or plateau?

Nvidia's data centre revenue reached $22.6 billion in the first quarter, up 427 per cent from a year earlier. The company also announced a ten-for-one stock split, and the shares hit a record. Investors reading the customer list, though, found a great many eggs in very few baskets.

A quarter without blemish

On the numbers alone the quarter was hard to fault. Total revenue came to about $26 billion, and the company guided to around $28 billion for the current quarter, again ahead of what analysts had pencilled in. It raised its dividend by 150 per cent as well, though the payout remains token for a company of its size. The day after the results the shares closed above $1,000 for the first time, before the split takes effect in June.

The profitability is as striking as the growth. Gross margins are running close to 80 per cent, the kind of figure normally associated with software rather than hardware. That reflects pricing power: demand for Nvidia's H100 accelerators has outstripped supply for more than a year, and customers have paid what was asked rather than wait in the queue. The shares more than tripled in 2023 on the strength of that position, and they have kept climbing this year.

Four big buyers

Close to half of Nvidia's revenue comes from a handful of hyperscalers: Microsoft, Meta, Alphabet and Amazon. They are spending billions on GPU clusters before the commercial uses are clear. If their shareholders push back, or they pause to digest the capacity they have already bought, Nvidia's growth could slow abruptly.

Nvidia's own finance chief put the large cloud providers' share of data centre revenue in the mid-40s per cent on the results call. Each of the four signalled higher spending when it reported in late April, and the money flows almost directly from their capital budgets into Nvidia's income statement. A chip bought this quarter is revenue for Nvidia now, while the cloud company depreciates it over several years and has to find paying customers for the computing time it sells.

Shareholders are already twitchy

The pushback has a precedent from only last month. When Meta raised its capital spending guidance for the year to $35-40 billion and gave little detail on when artificial intelligence would pay for itself, its shares fell by more than a tenth the next day. Alphabet and Microsoft, which could point to faster cloud revenue growth to set against the spending, got a kinder reception. The hyperscalers will keep buying as long as their investors believe the payback is coming, and that belief is what Nvidia's valuation depends on.

The concentration runs through to ordinary investors too. With a market value above $2 trillion, Nvidia is now one of the three largest companies in the S&P 500, so index funds and pension schemes that never chose to bet on AI hardware own a large slice of it anyway. A sharp fall in the shares would travel well beyond technology specialists, and the stock's results-day swings already move the whole index.

Nvidia's big customers are also its would-be rivals. Alphabet has its TPU chips, Amazon its Trainium and Inferentia, Microsoft its Maia and Meta its own accelerator. None yet threatens Nvidia's lead in training the largest models, but every in-house chip that handles routine inference is a sale Nvidia does not make.

Bottlenecks and borders

Part of Nvidia's pricing power rests on scarcity it does not control. Its chips are manufactured by TSMC in Taiwan, and the advanced packaging that joins the processor to its high-bandwidth memory has been the tightest constraint in the supply chain. SK Hynix, the main supplier of that memory, has become one of the boom's big winners. As TSMC adds packaging capacity, supply will catch up with demand, and the queue that sustains today's margins will shorten.

Geography adds another concentration. American export controls introduced in October 2022 and tightened in October 2023 bar Nvidia from selling its most advanced chips to China, once a large market for its data centre products. The company has designed lower-specification versions to stay within the rules, but Chinese buyers have alternatives from domestic suppliers, and Washington can tighten the restrictions again. That leaves Nvidia leaning even harder on a few American customers.

Booms that stopped

Nvidia has lived through sudden stops before. In 2018 demand from cryptocurrency miners evaporated, the company was left with a glut of gaming chips, and its shares roughly halved within a few months. The larger precedent is Cisco, briefly the world's most valuable company in March 2000, whose customers among telecoms carriers and internet start-ups built capacity far ahead of demand and then stopped ordering. Cisco's shares have never regained their 2000 peak.

The case against the comparison is strong. Cisco's customers financed their build-out with debt, and many went bust; Nvidia's biggest customers are among the most profitable companies in history and pay for their chips out of cash flow. They are also locked in a race in which no one wants to be the first to blink, which argues for spending to continue for some time. The risk is a pause rather than a collapse, but at Nvidia's valuation a pause would be enough to hurt.

Signals to watch

The hyperscalers' next round of results, due in late July, will show whether their capital budgets are still rising. The ramp of Nvidia's next-generation Blackwell chips later this year will show whether customers upgrade or wait. Any sign of shorter lead times for its current chips would suggest that supply is catching up with demand. TSMC's monthly sales figures, published around the tenth of each month, offer an earlier and rougher read on how hard the supply chain is running, and a sustained slowdown there would be the first warning sign worth heeding.

Nvidia's fortunes now rest on the capex budgets of a few Big Tech companies. When one of them signals a pause, the stock is likely to fall hard, however good the latest quarter looked, and the late-July results season is the first point at which such a signal could arrive.