This summer, OpenAI paused its largest planned frontier training run after its AI agents accidentally hacked Hugging Face and early evidence suggested that its next model could independently find and exploit security flaws. In September, Anthropic CEO Dario Amodei called for slowing advances in the most powerful AI models, warning that they could be used to launch cyberattacks or become harder for their own creators to control.
It was a strange moment for the AI boom. The people building the most powerful models were warning about what those models might be capable of, and investors were suddenly wondering what would happen to chip demand if the labs eased up on new models. US chip stocks bore the brunt of the sell-off.
“Pumping the brakes on frontier training won’t shrink the AI chip market but it will simply reallocate the spend.”
Despite the hubbub it created, whatever slowdown there was appears to have been short-lived. OpenAI had announced only a two-week pause; it restarted the run 10 days later, at the end of August, though it continued to “temporarily hold back some smaller experimental training runs.” Anthropic released upgraded versions of its Fable and Mythos frontier models in November, touting improvements in coding, reasoning, and vision.
And the companies selling the chips powering those advances are doing better than ever.
Nvidia on Tuesday posted a blockbuster quarter, with data center revenue rising 110 percent from a year ago to $107.6 billion in the three months ended October 25. The chip designer began shipping its most advanced AI computing system, the Vera Rubin NVL72, calling it the “fastest product ramp in the company’s history.” The Rubin platform generated nearly $20 billion in revenue in its first quarter of production shipments alone. The racks are already running at CoreWeave, Google, Microsoft, and Oracle. Meta plans to expand Vera Rubin deployments to train and run AI models and agents.
Nvidia name-dropped Anthropic and OpenAI, too. Anthropic has expanded access to Vera Rubin capacity through cloud providers and continues to work with Nvidia on future chip designs tailored to its workloads. OpenAI is among the companies adopting Nvidia’s new Vera CPUs.
“Proprietary models from OpenAI and Anthropic continue to advance the frontier,” Nvidia interim CEO Colette Kress said on the earnings call. She said the company expects major frontier labs to account for approximately 15 percent of its data center revenue this fiscal year. Nvidia also raised its revenue growth forecast for next fiscal year from 70 percent to 75 percent.
Chipmaker AMD, meanwhile, reported its biggest quarter ever earlier this month, with revenue up 45 percent to $13.4 billion. Data center sales drove that growth, nearly doubling to $8.5 billion. OpenAI and Anthropic were among the first customers running workloads on AMD’s new high-end Helios systems, which use its latest MI450 GPUs.
Broadcom, which helps many of the biggest AI companies design their own custom chips, has yet to report earnings. But so far, there’s little sign of a frontier slowdown hurting chip sales.
Of course, even if AI labs were to significantly scale back their training ambitions, it would take a while to show up in chipmakers’ earnings. Neil Shah, vice president of research at Counterpoint Research, told The Verge in September that it would take about six months to see the financial effects of a slowdown, owing to existing computing and memory procurement agreements. The impact could show up sooner in orders placed with chip manufacturers like TSMC.
So far, there’s little evidence of that, either: TSMC’s latest sales figures also show continued strength.
But there’s another reason a frontier slowdown might not be particularly bad for the chip industry: growing demand for inference, the process of running the AI models we already have.
“While one might argue the need for top-end GPUs will reduce with this shift toward inferencing, these workloads will collectively require more compute infrastructure than training,” Shah said. “Overall, it means a net increase in demand for all types of accelerators.”
There’s another reason a frontier slowdown might not be particularly bad for the chip industry: growing demand for inference, the process of running the AI models we already have.
According to our analysis of the latest data from corporate spending platform Ramp, frontier model families accounted for about 60 percent of token usage among businesses using Ramp in early November — a record high and about three times their share at the start of the year. That includes OpenAI’s Astra and Sol models and Anthropic’s Opus and Fable models.
But even as businesses use more AI, their spending is only growing modestly.
“It’s not really a demand issue though,” Ara Kharazian, lead economist at Ramp, tells The Verge. “By all measures, companies are using more AI.”
Rather, competition among AI model providers is driving down token prices so rapidly that it’s calling OpenAI and Anthropic’s revenue projections into question, Kharazian says.
That could prove beneficial for the companies supplying computing power. On Nvidia’s earnings call, Kress said token prices had fallen nearly 60 percent since May, spurring more usage. Meanwhile, rental prices for Nvidia GPUs across generations have gone up.
Nvidia and AMD are positioning their most powerful chips for both training and inference, including the demanding workloads created by AI agents.
That doesn’t necessarily mean Nvidia will maintain its dominance. As AI companies get more serious about the economics of serving their models to millions of people, they’re increasingly turning to custom chips that can do so more cheaply than general-purpose GPUs. Counterpoint expects shipments of custom AI server processors to surpass those of data center GPUs by 2028.
“Pumping the brakes on frontier training won’t shrink the AI chip market but it will simply reallocate the spend,” Shah said.
Indeed, on Nvidia’s earnings call, UBS analyst Timothy Arcuri asked Kress whether the company would be willing to sacrifice some of its hefty profit margins as frontier labs increasingly invest in custom chips and prioritize computing efficiency. Kress maintained that Nvidia’s systems justify their pricing and reiterated the company’s gross margin outlook.
For companies like Nvidia, the bigger threat may not be that AI companies slow down, but that they find cheaper chips on which to keep going.
Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.
- Rani Molla
Source: The Verge · Rani Molla
At a glance
- Topic
- Technology
- Country
- 🇺🇸 US
- Publisher language
- EN
- Added to Allnira