Chipmaker stocks will weather any AI slowdown

ONE fear over the past few days is that an artificial-intelligence slowdown, whatever that looks like, might harm the stocks of chipmakers. If AI development takes a breather, the insatiable appetite that has propelled component makers to unprecedented heights would dissipate.

Will AI companies still need all those data centres? The answer is yes. And then some.

The 6 per cent drop in the Philadelphia Semiconductor Index – which tracks the shares of top chipmakers like Nvidia and Intel – over the past five days, as AI doom fears have captured the news agenda, is an understandable reaction to the uncertainty created by hyperbolic but not entirely unwarranted talk of an AI-created catastrophe.

But missing from any of the pledges from the hyperscalers, or leading AI shops, is any talk of scaling back the infrastructure buildout. On the contrary. In his 4,500-word We Must Pace the Frontier essay, Anthropic co-founder Dario Amodei notes that “pacing does not mean halting model training or technical progress”.

It’s early days, and there is a faint chance Congress could enforce a slowdown of sorts, but so far there’s no indication of any pulling back in the numbers.

Bank of America (BOA) analysts note sky-high memory costs – a key indicator of demand for the AI buildout – are unchanged. The rental cost for Nvidia’s widely deployed B200 system is US$5.72 an hour, BOA said, and has risen steadily over the past two months.

Now there may be a lag between AI slowdown talk and AI slowdown action. But don’t bet on it: “2027 remains much a fully booked/contracted year across all compute/networking/memory vendors,” BOA analysts wrote. “And we expect 2028 to also remain tight led by accelerating demand.”

In fact, were it not for the headlines and dramatically resigning employees, you would think the AI race had actually stepped up a notch. On Tuesday, Meta Platforms announced it still planned to go full steam ahead with putting its own AI chips into data centres by the beginning of 2027.

This comes from a company that claimed to have its own AI-gone-rogue cyberattack moment in August, just like OpenAI’s Hugging Face fiasco.

Elon Musk, with a comment firmly in grain-of-salt territory, outlined his plans to put Nvidia’s Vera Rubin AI stack in space in 2027. I don’t believe him, but that’s beside the point, which is that none of the main players in AI show signs of letting up on investing in capacity to run these models.

A useful exercise is to consider how AI companies describe the path to safer AI and ask whether it indicates less spending.

In August, with an urgency echoing the “pause” talk of the past few days, more than 100 AI companies – OpenAI, Anthropic, Alphabet’s Google and Microsoft among them – put out a joint statement arguing that companies should make it “an immediate leadership priority” to protect from AI harm – whether that be vulnerabilities discovered by supersmart models or models acting in unexpected ways. How might they do that?

“Today’s AI advances are already giving defenders new ways to fix weaknesses that have accumulated for years,” the letter suggests. It appears the solution to the AI problem, handily enough, is more AI.

That doesn’t sound like a slowdown in demand. BLOOMBERG