Nvidia beats and still falls

Nvidia's second-quarter revenue rose 122 per cent to $30 billion, ahead of consensus forecasts. Almost any other company would have been celebrated. Nvidia's shares slipped six per cent in after-hours trading. The chipmaker has reached the point where beating the official consensus is not enough to satisfy the whisper numbers.

A year of escalating bets

The bar has risen with the share price. Nvidia briefly became the world's most valuable listed company in June, shortly after a ten-for-one stock split, and it went into these results valued at around $3 trillion. Its revenue has more than doubled year-on-year for several consecutive quarters, driven by the cloud giants' appetite for chips to train and run large AI models. Each beat has taught shareholders to expect the next one to be bigger.

Good is no longer enough

Investors used to enormous upside surprises noticed that sequential revenue growth was slowing, and that gross margins dipped slightly because of production changes for the coming Blackwell chips. When a company trades at thirty times forward sales, anything short of perfection brings a lower multiple. Even the leader in AI chips is subject to the law of diminishing surprise.

The detail explains the reaction. Revenue grew about 15 per cent on the previous quarter, down from 18 per cent in the first quarter and more than 20 per cent in each of the two quarters before that. Guidance for the third quarter of $32.5 billion, plus or minus two per cent, topped the official consensus but not the most bullish hopes. The company also said it had changed the mask used to make Blackwell to improve production yields, and gross margin slipped from the high seventies to the mid-seventies, partly because of provisions against low-yielding Blackwell stock.

Who is paying for the boom

Nvidia's revenue is somebody else's capital spending. Almost all the growth comes from its data-centre division, and a small number of very large buyers, chiefly the cloud platforms run by Microsoft, Amazon and Alphabet, together with Meta, account for much of it. Those companies have told investors this summer that they will keep raising investment in AI infrastructure, arguing that the risk of spending too little outweighs the risk of spending too much. Their shareholders have started to ask when that spending will produce matching revenue.

That question is the main risk to Nvidia. The chips are bought ahead of demand for AI services, on the expectation that customers will pay for them. If the payback takes longer than hoped, the cloud giants can slow their orders quickly, and a supplier whose sales have doubled every year would then face a sharp deceleration. Export controls add a second constraint: American rules introduced in 2022 and tightened in 2023 bar Nvidia from selling its most powerful chips to China, which once accounted for a much larger share of its data-centre sales.

What thirty times sales implies

A multiple of thirty times forward sales assumes more than continued growth. It assumes that margins stay near today's levels as rivals and customers design their own chips, and that the biggest buyers keep raising capital spending at the current pace. Alphabet, Amazon and Microsoft each design in-house AI chips, and AMD is selling rival accelerators. None of these threatens Nvidia's lead this year, but each limits how long margins in the mid-seventies can last.

The stakes reach beyond Nvidia's own shareholders. The company is now one of the three largest weights in the S&P 500, so index funds and pension schemes that never chose to bet on AI chips carry a big exposure to them anyway. Suppliers such as TSMC, which manufactures Nvidia's chips, trade on the same expectations. A de-rating would spread through benchmarks well beyond the semiconductor sector.

A precedent from 2000

The closest parallel is Cisco. In March 2000 the network-equipment maker briefly became the world's most valuable company, selling the routers the internet was built on. Its annual revenue still rose in the following fiscal year, but the shares had peaked and have never regained that level. Cisco survived as a business. Its valuation had simply assumed more growth than any business could deliver.

The internet did go on to transform the economy, as Cisco's bulls had said it would. The profits, however, flowed largely to companies that were barely visible in 2000, while the suppliers of the first wave of equipment saw growth normalise and their multiples collapse. Investors paid for a correct forecast about technology and an incorrect one about who would capture its value.

Where the bears may be wrong

The comparison has limits. Nvidia's earnings have risen with its share price, so its price-earnings ratio is far lower than Cisco's was at the peak, and its main customers are among the most profitable companies in the world rather than cash-hungry start-ups. The company also authorised a further $50 billion of share buybacks. If Blackwell ships on schedule and the cloud giants keep lifting capital spending, the multiple could compress through rising earnings rather than a falling price.

The next tests are Blackwell shipments, which Nvidia expects to start contributing revenue in its fiscal fourth quarter, and the spending plans of its biggest customers when they report in late October. At thirty times sales, Nvidia's shares need a perfect quarter every quarter. Triple-digit growth will not stop the multiple shrinking from here, and the stock will probably be choppier for it. Expect the shares to trade sideways or lower into the October capital-spending updates, with passive index investors, rather than the AI enthusiasts who drove the rally, carrying much of the risk of any de-rating.