Chip demand holds up despite frontier AI pause
The frontier AI slowdown that briefly unsettled chip investors has not yet dented sales. Nvidia and AMD have reported record or sharply higher results, while demand for inference—the computing needed to run existing models—is growing even as questions mount about the cost of training ever-larger systems.

OpenAI resumed its largest planned frontier training run 10 days after announcing a two-week pause in August, although it continued to hold back some smaller experimental runs. Anthropic released upgraded Fable and Mythos models in November. The pause followed reports that OpenAI agents had accidentally hacked Hugging Face and early evidence that a forthcoming model could find and exploit security flaws. In September, Anthropic CEO Dario Amodei warned that powerful models could enable cyberattacks or become harder for their creators to control.
Nvidia’s data center revenue rose 110 percent year over year to $107.6 billion in the three months ended October 25. Its Vera Rubin NVL72 system began shipping, and the company said the platform generated nearly $20 billion in its first quarter of production shipments. CoreWeave, Google, Microsoft and Oracle are already running the systems; Meta plans to expand its deployments. Nvidia interim CEO Colette Kress said frontier labs are expected to provide about 15 percent of data center revenue this fiscal year. Nvidia also lifted its next-fiscal-year revenue growth forecast from 70 percent to 75 percent.
AMD reported its largest quarter, with revenue up 45 percent to $13.4 billion. Data center sales nearly doubled to $8.5 billion, with OpenAI and Anthropic among the first customers running workloads on its Helios systems, which use MI450 GPUs. Broadcom has yet to report earnings, while TSMC’s latest sales figures also point to continued strength. Counterpoint Research’s Neil Shah said existing computing and memory procurement agreements could delay the financial effects of any slowdown by about six months, though orders to manufacturers such as TSMC could show changes sooner.
The shift toward inference may offset weaker training demand. Shah said inference workloads could require more computing infrastructure overall, creating a net increase in demand for accelerators. Ramp data showed frontier model families accounted for about 60 percent of business token usage in early November, a record and roughly three times their share at the start of the year. Ramp lead economist Ara Kharazian said companies are using more AI, but falling token prices are limiting spending growth. Kress said prices had dropped nearly 60 percent since May, spurring usage, while rental prices for Nvidia GPUs across generations had risen.
The risk for Nvidia may be a change in where chip spending goes, not a halt to AI investment. Companies are adopting custom processors to serve models more cheaply, and Counterpoint expects shipments of custom AI server processors to surpass data center GPU shipments by 2028. Shah said a pause in frontier training would reallocate chip spending rather than shrink the market. Nvidia has defended its pricing and maintained its gross margin outlook as labs weigh custom chips and efficiency.
Comments (0)
No comments yet. Be the first!