Concerns from global AI leaders about slowing the pace of artificial intelligence (AI) development have shaken semiconductor stocks. However, a closer look at the discussions reveals that they are not advocating for a complete halt to AI development.
Dario Amodei, CEO of Anthropic, argued in a public statement on September 12 titled 'We Must Pace the Frontier' that the speed of performance improvements in cutting-edge AI should be regulated to a level that safety technologies can keep up with.
As AI is increasingly used to develop next-generation AI, the pace of technological advancement is accelerating, while safety measures are struggling to keep pace. Sam Altman, CEO of OpenAI, Elon Musk, CEO of SpaceX, and Demis Hassabis, CEO of Google DeepMind, also expressed agreement on the need for speed regulation.
The market reacted immediately. On September 14, the KOSPI index fell by 3.26%, with Samsung Electronics and SK Hynix dropping 4.05% and 6.35%, respectively. In the global market, AI semiconductor stocks, including Nvidia, also showed weakness. However, the stock market is facing other challenges, including rising international oil prices and interest rate pressures, in addition to the AI controversy.
Amodei clarified that speed regulation does not mean stopping model training or technological advancement itself. Instead, he proposed that external evaluators verify the internal models and training processes of AI companies, ensuring that safety standards are met when a model reaches a certain level of capability. Altman also stated, 'When we talk about speed regulation, we do not mean 'stopping.' Progress has been rapid and will continue to be, but it should be slower than it would be without any intervention.'
The semiconductor industry’s sensitive reaction is not merely a market misunderstanding. Amodei mentioned that speed regulation could involve limiting the computing resources and training methods used for models, as well as the development of AI using AI. He also noted that a comprehensive speed regulation or temporary halt to significantly limit the pace of AI development through international agreements is a long-term option, although he believes such an agreement is unlikely to be realized soon.
Training cutting-edge AI models requires large-scale GPUs and high-bandwidth memory (HBM). If the scale or frequency of training is limited, the previously assumed growth rate of semiconductor demand for AI could also decline. Reuters reported that if speed regulation becomes a reality, it could pose challenges for companies like Nvidia that directly benefit from AI training infrastructure.
However, there are currently no common regulations in place that forcibly limit the training scale of AI companies. Amodei's proposals are still largely in the suggestion phase. He believes that limiting the pace of AI development between countries is difficult due to verification and competitive issues.
Notably, AI computing demand is not solely dependent on training. AI computing demand is divided into training for creating new models and inference for using already developed models. McKinsey forecasts that global data center AI inference demand will increase from 20.9 GW in 2025 to 93.3 GW by 2030, surpassing training demand of 62.2 GW during the same period. This projection pertains to data center power demand and does not directly predict specific GPU or HBM sales figures.
The key question is where AI semiconductor demand will decrease and where it will increase. If AI speed regulation is institutionalized, it may burden the demand for semiconductors used in training frontier models, but the inference demand driven by the expansion of AI services remains a separate growth variable.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.