Warnings have emerged that the pace of developing frontier artificial intelligence (AI) models should be slowed, but the AI semiconductor market is instead continuing to grow. Spending by AI companies is shifting from model training to inference, and demand for semiconductors is being reshaped rather than shrinking, according to an analysis.
On Nov. 18 (local time), IT media outlet The Verge reported that OpenAI and Anthropic at one point mentioned the need to moderate the pace of developing high-performance AI models, but there has not yet been a clear hit to results at semiconductor companies such as Nvidia and AMD.
The controversy began this summer when OpenAI temporarily halted large-scale frontier model training. An AI agent inadvertently hacked Hugging Face, and at an early stage there were signs the next model could find and exploit security vulnerabilities on its own. In September, Anthropic CEO Dario Amodei (다리오 아모데이) warned that the development of powerful AI models should be slowed, citing the risk of cyberattacks and loss of control.
At the time, investors worried that if AI labs reduced new model training, semiconductor demand could also weaken. OpenAI, however, ended a pause originally forecast to last two weeks after 10 days and resumed training in late August. It only temporarily held back some small experimental training. Anthropic also unveiled upgraded versions of "Fable" and "Mythos" in November, highlighting improved coding, reasoning and vision performance.
Semiconductor company results remain solid. Nvidia said recent-quarter data centre revenue for the period ended Oct. 25 rose 110 percent from a year earlier to $107.6 billion. It also began shipping its latest AI computing system, "Vera Rubin NVL72". The company described it as the fastest product ramp in its history. The Rubin platform recorded close to $20 billion in revenue in its first quarter of production shipments and is running at CoreWeave, Google, Microsoft and Oracle, among others. Meta also plans to expand adoption for training and operating AI models and agents.
Nvidia also highlighted cooperative ties with OpenAI and Anthropic. Anthropic is expanding its use of Vera Rubin computing resources through cloud service providers and is working with Nvidia to design next-generation chips optimised for its workloads. OpenAI is among the companies adopting Nvidia's new Vera CPU. Nvidia interim CEO Colette Kress (콜레트 크레스) forecast major frontier AI labs would account for about 15 percent of data centre revenue in the current fiscal year. The company also raised its revenue growth outlook for the next fiscal year to 75 percent from 70 percent.
AMD also posted a record quarterly result early this month. Revenue rose 45 percent from a year earlier to $13.4 billion, and data centre revenue climbed 97 percent to $8.5 billion. OpenAI and Anthropic were among early customers using AMD's high-performance "Helios" system equipped with the latest MI450 GPU. There has also been no clear sign of a slowdown so far in TSMC revenue.
The article cites existing procurement contracts and inference demand as reasons semiconductor demand does not immediately decline even if frontier model training slows. Neil Shah (닐 샤), a vice president at Counterpoint Research, said it could take about six months for the financial impact of slower training to appear because of existing computing and memory procurement contracts. Demand for inference, which uses already developed AI models in live services, is also rising quickly. Shah said a shift toward inference could reduce demand for individual high-end GPUs, but overall it could require more computing infrastructure than training, boosting demand for all types of accelerators.
Corporate AI usage is also increasing. Data from corporate spend management platform Ramp showed frontier model families accounted for about 60 percent of companies' token usage in early November, a record high. That is about three times higher than at the start of the year. But token prices are falling quickly due to competition among model providers, limiting the rise in corporate AI spending. Ara Kharazian (아라 카라지안), Ramp's chief economist, said companies are using AI more, but falling prices are weighing on revenue outlooks for OpenAI and Anthropic.
Falling token prices could instead be an opportunity for semiconductor companies. Kress said token prices have fallen about 60 percent since May, lifting usage. By contrast, rental prices for Nvidia GPUs by generation have risen. Nvidia and AMD are supplying high-performance chips not only for training but also for inference and high-load tasks for AI agents.
Still, there is no guarantee Nvidia's dominance will continue. AI companies are adopting customised chips that are cheaper than general-purpose GPUs to reduce the cost of running large-scale model services. Counterpoint forecast shipments of customised AI server processors would exceed data centre GPUs by 2028. The bigger variable for the semiconductor industry is not a slowdown in AI development, but whether AI companies can keep growing using cheaper chips, the article said.