Nvidia's 'Wow' Factor Is Fading - Heard on the Street — WSJ

By Asa Fitch

Nvidia shocked the world two years ago in May when it posted earnings that included a forecast of $11 billion of quarterly revenue, some 53% more than analysts were predicting.

A string of surging sales and profits soon followed. But such barnburner quarters aren't likely happening again any time soon.

Nvidia's latest quarter put the deceleration on display. While sales came in above analysts' expectations, it was the narrowest beat in the nine quarters since the artificial-intelligence boom propelled the company to unprecedented heights. Its shares have fallen about 6% since the results.

One reason for Nvidia's slower growth is obvious: The law of large numbers dictates that the bigger you get, the more difficult it is to keep getting bigger at the same breakneck pace.

But there is another, more structural reason Nvidia's growth looks less exponential: supply-chain constraints that no amount of financial success can cure.

Nvidia's AI chips are almost all produced by Taiwan Semiconductor Manufacturing Co., then are packaged up into elaborate configurations that are inserted into increasingly elaborate computing systems. Those systems then go into the data centers where AI number-crunching takes place.

Nvidia has made significant strides over the past two years in both producing more chips and getting them into the systems its customers crave. But it is a complex dance: There are about 600,000 components in each of the company's biggest AI computing systems, Chief Executive Jensen Huang said in March, and some of its future systems are set to have 2.5 million parts.

At a company conference in June, Huang said supplies for Nvidia's current systems weren't "horribly difficult to get" as long as the company could forecast to suppliers what it needed. But there were challenges. "It's constrained, but we're still growing fairly fast," he said, adding that it currently takes about a year between when an Nvidia chip starts being made and when an AI supercomputer ships to a customer.

There are other signs that Nvidia is on a more measured trajectory. In Nvidia's latest quarter, UBS analysts said their metric of "supply" — the sum of inventory items and supply commitments that attempts to gauge future revenue trends — rose to about $45 billion from about $41 billion in the previous quarter. Juxtaposing those figures with Nvidia's revenue guidance, this analysis "tells us that supply looks sufficient for revenue to continue growing ratably, but does not foreshadow a big 'hockey stick' in revenue," the analysts said.

Already, Nvidia's 56% annual revenue growth rate in its latest quarter was its slowest in more than two years. If analyst projections hold, growth will slow further in the current quarter.

Huang thinks that many of the supply-chain issues will be worked out soon — part of the reason he has confidence that the build-out of AI data centers will become a $3 trillion to $4 trillion endeavor by the end of the decade. Nvidia's next-generation Rubin chips, which are expected to be widely available early next year, will coincide with a far more mature and scaled up supply chain, Huang said on a call with analysts last month.

It is possible that Huang is correct and that supply-chain fixes could propel Nvidia even further. But it's also reasonable to suppose that some constraints will persist — any number of which could hold the company back. Hangups in chip production have already had an impact, including limiting availability of an advanced manufacturing technique from TSMC for knitting chips together that is essential in producing Nvidia's AI hardware.

Even if Nvidia does work supply-chain magic, though, there is another structural problem coming up that it has less control over: The world's electricity grids aren't growing fast enough to handle AI computing demand.

The energy problem hasn't slowed AI's march much yet, but it's easy to see a collision coming. U.S. utilities are reluctant to develop energy infrastructure for big AI projects because they aren't convinced the AI boom will last long enough for them to recoup huge investments that could add to costs for other customers.

Expanding energy generation and transmission lines also takes years to do — a pace that's far slower than AI-demand growth.

Huang has fended off the power challenge by pointing to how Nvidia's newest chips have become far more energy-efficient — thus putting less strain on electrical grids. While the efficiency gains may be real, they don't necessarily mean customers will use less electricity. Like adding lanes to a roadway, greater efficiency invites greater usage.

Nvidia is still delighting customers with its chips and making boatloads of money. But given its supply and infrastructure constraints — as well as recent signs that AI progress is slowing — there is a good chance its future earnings surprises will be a little less delightful for investors.

Write to Asa Fitch at asa.fitch@wsj.com