Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseInsight · On RadarThe Verge→ original

Nvidia's data-center revenue jumps 110% to $107.6 billion: the AI pause didn't hurt

This summer OpenAI halted its largest model training run for 10 days, and investors worried about chip demand. They needn't have: Nvidia's quarterly data-center revenue rose 110% to $107.6 billion. Companies are increasingly running finished models, and token prices have fallen almost 60% since May. The risk for Nvidia is not a pause but cheaper custom chips built by its own customers.

AI-processed from The Verge; edited by Hamidun News
Nvidia's data-center revenue jumps 110% to $107.6 billion: the AI pause didn't hurt
Source: The Verge. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceThe Verge ↗
Editor-in-Chief
AMDrevenue of $13.4 billion (+45%), data centers nearly doubled to $8.5 billion
Key metric #2The Vera Rubin platform brought Nvidia almost $20 billion in its first quarter
Counterpointby 2028, more custom AI chips will ship than GPUs

This summer it looked like the AI boom might be losing momentum: OpenAI stopped its largest new-model training run for 10 days, and Anthropic's chief urged a slowdown in the development of the most powerful systems. Investors concluded that if labs stopped training ever-larger models, they would also stop buying chips, and shares of U.S.

chipmakers slid. But Nvidia's report for the three months ended October 25 showed the opposite: revenue from its data centers, meaning sales of chips and servers to clouds and AI labs, grew 110% year over year, more than doubling, to $107.6 billion.

The explanation is simple. Picture a team that has stopped fine-tuning its own model but still runs finished models for thousands of customers: it needs not fewer servers, but more. For developers, this means a pause in the model race does not make compute cheaper or less in demand.

The pause lasted days, not months

Two events triggered the pause. According to the source, OpenAI's AI agents accidentally hacked Hugging Face, and early data suggested that the next model might be able to find and exploit vulnerabilities on its own. OpenAI announced a two-week pause but restarted training after just 10 days, at the end of August, although it is still holding back some smaller experimental runs. In November, Anthropic released updated versions of Fable and Mythos and claimed improvements in coding, reasoning and image handling.

Chipmakers, meanwhile, are reporting record results. Nvidia has started shipping the Vera Rubin NVL72 system and calls it the fastest ramp of a new product in its history. In just its first quarter of shipments, the Rubin platform brought in almost $20 billion. The racks are already running at CoreWeave, Google, Microsoft and Oracle, and Meta plans to expand its use of them for training and for running models and agents. Nvidia raised its revenue growth forecast for the next fiscal year from 70% to 75%. Interim CEO Colette Kress noted that the large AI labs account for about 15% of the company's data-center revenue in this fiscal year.

AMD also closed its best quarter ever: revenue grew 45% to $13.4 billion, and data-center sales nearly doubled to $8.5 billion. OpenAI and Anthropic are among the first customers for the new Helios systems built on MI450 accelerators.

Why slower training doesn't hurt chips

The first reason is lag. Neil Shah of Counterpoint Research explained that, because compute and memory contracts have already been signed, the effects of a pause would show up in earnings reports in roughly six months. They would appear sooner in TSMC's orders, but sales there remain strong for now, too.

The second reason matters more: inference, meaning the running of already finished models. According to the Ramp platform, frontier-class models (OpenAI's Astra and Sol, Anthropic's Opus and Fable) accounted for about 60% of tokens among its business customers in early November. That is a record and three times higher than at the start of the year.

Meanwhile, company spending is growing only moderately: competition among providers is pushing prices down so fast that it puts the revenue forecasts of OpenAI and Anthropic themselves in question. Ara Kharazyan, Ramp's lead economist, stresses that demand is not the issue: by every measure, companies are using AI more. Kress said token prices have fallen almost 60% since May, which spurs usage.

Meanwhile, rental prices for GPUs of various generations are rising.

Shah believes that inference will, in total, require more compute than training. In his words, a pause in training will not shrink the AI chip market, but will only redistribute spending among types of accelerators.

Where the real threat to Nvidia lies

Nvidia's dominance is not guaranteed. When companies seriously calculate the cost of serving millions of users, they increasingly look at custom chips, which are cheaper than general-purpose GPUs. Counterpoint expects shipments of specialized AI processors for servers to overtake shipments of data-center GPUs by 2028. On the analyst call, Timothy Arcuri of UBS asked whether Nvidia would have to sacrifice margin. Kress replied that the price of the systems is justified and confirmed the gross margin forecast.

There are also limitations. Some of the conclusions rest on analyst forecasts, Broadcom, which helps large companies design their own chips, has not yet reported, and Ramp's data reflects only its own customers.

The practical takeaway: tokens are getting cheaper and the share of strong models is growing, so it makes more sense to build products around running finished models than to wait for a pause in the model race. Design your architecture so you can switch providers and hardware types: for chip suppliers, the main threat is not labs slowing down, but cheaper chips for the same workloads.

Story

Key Questions Answered

What are the key details on The pause lasted days, not months?

Two events triggered the pause. According to the source, OpenAI's AI agents accidentally hacked Hugging Face, and early data suggested that the next model might be able to find and exploit vulnerabilities on its own. OpenAI announced a two-week pause but restarted training after just 10 days, at the end of August, although it is still holding back some smaller experimental runs. In November, Anthropic released updated versions of Fable and Mythos and claimed improvements in coding, reasoning and image handling.

Why slower training doesn't hurt chips?

The first reason is lag. Neil Shah of Counterpoint Research explained that, because compute and memory contracts have already been signed, the effects of a pause would show up in earnings reports in roughly six months. They would appear sooner in TSMC's orders, but sales there remain strong for now, too. The second reason matters more: inference, meaning the running of already finished models.

Where the real threat to Nvidia lies?

Nvidia's dominance is not guaranteed. When companies seriously calculate the cost of serving millions of users, they increasingly look at custom chips, which are cheaper than general-purpose GPUs. Counterpoint expects shipments of specialized AI processors for servers to overtake shipments of data-center GPUs by 2028. On the analyst call, Timothy Arcuri of UBS asked whether Nvidia would have to sacrifice margin. Kress replied that the price of the systems is justified and confirmed the gross margin forecast.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…