Wall Street Fears an AI Slowdown. I’m Not Buying It.
It’s a considerable leap from slowing AI development to assuming the industry will need less computing power.
You've reached your free article limit
You've read 0 of 1 free Pro articles.
I suspect some of the anxiety behind Monday’s AI infrastructure selloff goes back a quarter-century. Investors who lived through the telecom bust learned that a technology can change the world while the companies building its infrastructure still destroy shareholder wealth.
Back then, carriers laid fiber far ahead of demand, leaving much of it dark or unused. More people were using the internet, but there still wasn’t enough business to make all that investment pay. By July 2001, Federal Reserve Vice Chairman Roger Ferguson was warning that it could take a long time to work through the excess. And telecom investors were getting killed in the meantime.
I get why that memory makes investors nervous, especially when the people leading the AI race start talking about slowing down.
Dario Amodei’s call for slower advances in AI capabilities, backed by Sam Altman and Elon Musk, gave investors a reason to sell the companies supplying the AI buildout Monday. But I think investors are making a pretty big leap from talk of slowing model development to the assumption that the industry will need less computing power.
If Dario slows down, does anyone believe Sam and Elon will politely wait? They’re not trying to arrange a three-way tie for first place. They’re competing for customers and talent, and their investors expect a return on the enormous checks they’ve written. Two things can be true at the same time. They can care about safety and still spend like hell to win.
President Trump is also resisting calls for a slowdown. The Wall Street Journal reported Monday that he rejected tighter federal AI regulation and argued that restrictions would help China. Speaking with Nvidia (NVDA) CEO Jensen Huang, he acknowledged a need for care but added, “That doesn’t mean we are going to stop an industry.”
Meanwhile, customers still need more computing power, and they’re willing to pay for chips that have been around for years. On Nvidia’s February earnings call, Huang said, “Even our six-year-old GPUs in the cloud are completely consumed, and the pricing is going up.”
CoreWeave (CRWV) told a similar story in August. It said older generations of Nvidia GPUs remained largely sold out and disclosed a contract for A100 capacity extending into 2029. That chip was introduced in 2020, so we’re not talking about new tech.
Customers are finding enough work for an older generation of chips to keep them in production for several more years. That’s a long way from laying fiber and waiting for enough business to show up.
Even Anthropic has been struggling to keep up with demand. In April, the company said heavy use of Claude was making the service less reliable during busy periods. That helps explain its commitment of more than $100 billion over 10 years to Amazon’s (AMZN) cloud services, securing up to five gigawatts of capacity to train and run its models.
And customers don’t need a new version of Claude to need more computing power. A company using it to help 20 employees write code can roll it out to 200 employees. It’s the same model, doing more work for more people. As that happens across more businesses, the chips and data centers have more work to do, regardless of when the next model arrives.
Serving that demand requires chips, racks, cooling, power, and data centers.
And customers are booking more than just time on existing chips. They’re reserving space in data centers that haven’t opened yet.
Applied Digital (APLD) announced a 15-year lease in June for 210 megawatts at its Delta Forge 2 project, with about $5.2 billion in contracted revenue. Operations are expected to begin in early 2028. For the companies developing these facilities, that means customers are already committing to years of business before the doors open.
And Applied Digital is hardly alone. TeraWulf (WULF) and Core Scientific (CORZ) have signed similar deals. Customers want this capacity so badly that they commit to years of payments before the facilities are even ready for use.
Even the proposed safety work needs infrastructure. Gavin Baker, managing partner and chief investment officer at Atreides Management, expects labs that pace development to devote more time and compute to alignment, monitoring, and evaluations. In his view, that could mean slightly higher (not lower!) compute spending and lower margins for the labs. Their suppliers would still have work to do.
Bottom Line
Getting burned in telecom is a good reason to check who’s buying, how the bills get paid, and what we’re paying for the stock. You can still lose money buying a good business at the wrong price. That’ll never change. But a scary essay doesn’t mean its customers have stopped calling.
I don’t expect the labs to give up the race, and their customers already need more capacity. I wouldn’t abandon a sound supplier, neocloud, or data center developer simply because those customers are discussing more safety reviews. I expect these stocks to stay volatile, so use those swings to your advantage without letting a big rally talk you into chasing or a sharp selloff convince you the whole story is over.
At the time of publication, Byrne had no positions in any securities mentioned.
