AI Spending Will Cloud Chip Slowdown

During a two-hour long presentation at its I/O developers conference on Wednesday, Google took pains to point out its use of Nvidia’s H100, a high-end data center processor that the chip maker began shipping only earlier this year. No other chip makers or data-center component suppliers got such a shout-out. And it wasn’t just Google; Microsoft Chief Executive Satya Nadella name-dropped Nvidia at the top of his company’s earnings call last month as a key supporting point to buttress his claim that “we have the most powerful AI infrastructure” for training large data models. Again, no other chip maker was mentioned.

Those are good signs for Nvidia, the most valuable chip company on the market. It has already seen its market value nearly double since the start of the year.

In its own results coming later this month, Nvidia is expected to report data-center revenue growing 4% year over year to $3.9 billion for the fiscal quarter that ended in April. That might seem slight for a business unit that was running high double-digit growth last year, but it contrasts sharply with the 38% plunge that Intel posted for its own data-center segment for the March quarter. Even Advanced Micro Devices, which has been gaining on Intel in this key market, saw basically no growth in data centers for the March period and projected a decline for the June quarter.

Nvidia’s advantage lies in its graphics processors that are specifically designed for AI uses. And with tech giants such as Alphabet’s Google, Microsoft, Amazon and Facebook-parent Meta Platforms now rushing to build up generative AI capabilities similar to the popular ChatGPT technology, they need components such as Nvidia’s H100 chip. But those same companies are also facing slowdowns in their core businesses and are thus under pressure to slow down capital spending that had been on a tear of late.

Combined capital spending by the four aforementioned companies fell 4% year over year during the March quarter—the first such decline in four years. Market-research firm Dell’Oro Group projects that capital expenditures specific to data centers will grow in the mid-single-digit range this year compared with a 36% surge in 2022.

But the pain of that slowdown won’t be spread equally. The need to invest in pricey chips such as Nvidia’s H100 to power generative AI offerings means cuts need to come elsewhere in data-center outlays. In a report Friday, Morgan Stanley analyst Joe Moore wrote that “it’s very clear that customers are consolidating purchases of traditional servers to make room in the budget for GPU servers that have 20x the price.”

Google’s plan to infuse chatbot functions into its huge search business alone is an expensive prospect: Pierre Ferragu of New Street Research says powering all Google queries with Nvidia’s chips would represent an additional $80 billion in capex, though he added that he expects the company to offset some of those computing needs with its own, in-house TPU chips.

Digestion of last year’s surge in chip purchases will likely help other data-center chip makers as the year progresses. Mr. Moore of Morgan Stanley says the AI opportunity for AMD “looks to be multiples of our initial assessment,” which originally projected AI revenue of $100 million for the chip maker in 2024. Mr. Moore now sees that in the $400 million range.

Still, gains from AI will be one of few bright spots for the chip industry that is still mired in a major sales slump. The Semiconductor Industry Association reported earlier this month that industry sales in March fell 21% from a year ago—the worst monthly decline the industry has seen since 2009. The last major slowdown tracked by the group lasted for 13 months, while the current one is only eight months in. Even Big Tech’s deep pockets only go so far.

Write to Dan Gallagher at dan.gallagher@wsj.com