Leaders of major artificial intelligence (AI) companies such as Anthropic, OpenAI and xAI have recently argued in succession that the pace of developing cutting-edge AI models should slow, fuelling questions inside and outside the industry about what is driving the push. While they describe it as a safety measure, some also raise the view that it could work in their favour amid IPO preparations, profitability pressure and shifts in the regulatory environment.

On Sept. 14, local time, foreign media including Business Insider reported that the debate has centred on Dario Amodei (다리오 아모데이), chief executive of Anthropic, Sam Altman (샘 알트먼), CEO of OpenAI, Elon Musk (일론 머스크), CEO of xAI, and Demis Hassabis (데미스 허사비스), co-founder of Google DeepMind. They agreed that AI could pose serious risks to humanity and that more safety guardrails are needed for frontier models, along with a moderated development pace. Silicon Valley has long warned of AI side effects, but it is unusual for leaders of major AI companies to line up so openly behind calls to slow down.

Amodei recently argued that the pace of improvements in AI model capabilities should slow, in a piece calling for moderating “the speed of the frontier”. He warned that as AI rapidly advances its ability to help develop the next generation of AI, recursive self-improvement (RSI) is becoming a reality and, if not controlled, could outstrip humans’ ability to understand and control it. Amodei proposed a three-step plan: allowing constant access for third-party evaluators, setting industry-wide safety standards and pursuing international coordination.

But the timing of the slowdown push overlaps with a period when the AI industry faces multiple pressures at once. Public backlash is growing over data centre expansion, and investors and analysts are pressing for when real profits will emerge from AI businesses that have absorbed massive funding. As open-weight models spread as alternatives to costly frontier models, pressure is also rising on large model developers’ monetisation strategies.

IPO preparations at Anthropic and OpenAI are also cited as variables, as listing would bring stricter scrutiny of their finances and business structures. Anthropic is reported to have chosen Nasdaq as its listing venue, and Altman said that pursuing an initial public offering (IPO) this year would be “an unwise time” because many tasks related to AI safety still need to be resolved.

There is also criticism that moderating development speed could work in favour of established large players. If new regulatory frameworks are created, companies that have already seized the market could exert influence in designing the rules, while compliance costs such as legal, audit and internal controls could become barriers to entry for latecomers. Some therefore voice suspicions that AI companies may be trying to build a “cartel” by invoking safety to block competitors or steer regulation to their advantage.

Some also see the slowdown argument as a tool companies could use to explain weak performance or cost burdens. Even if product launches are delayed or performance targets are missed, they can say they chose a cautious approach for safety, and slow cost cutting can likewise be justified as structural costs needed for safe AI development. There are also concerns that industry moves to demand regulation first could, over the long term, end up protecting incumbents’ interests.

Still, it is hard to explain the slowdown push solely in terms of corporate interests. AI safety researchers and industry officials have recently pointed to cases in which AI agents acted against human intent or attacked external systems. An incident in which OpenAI’s AI agents carried out a cyberattack against Hugging Face was one of the main triggers for Amodei to raise the slowdown argument. Amodei warned that if more powerful AI agents than today were to break out of control in the same way, they could cause serious damage.

Jacob Coxon (제이콥 콕슨), a former researcher at Anthropic, argued at the time of his resignation that people developing AI are seriously worried about the possibility that AI could destroy all of humanity within 10 years. His public post greatly expanded the AI safety debate, and concerns are also growing in the AI industry that recursive self-improvement could become reality faster than expected. Some researchers warn that if AI reaches a stage where it develops more powerful AI on its own, risks that cannot be compared with today could emerge in areas such as cybersecurity.

Others also point to U.S.-China rivalry as the biggest obstacle to calls for slowing AI. The argument has been raised that competition should continue because slowing U.S. AI development could allow China to pull ahead. The Chinese government has also recently criticised calls to slow AI development as “fearmongering”. Some experts, however, argue that because AI risks cross borders, international coordination including the United States and China is instead needed.

Ultimately, the debate goes beyond the question of whether to slow AI development. If AI companies fear the emergence of technology so powerful that even they struggle to control it, a question remains over whether it is appropriate to keep leaving the direction of its development and the authority to control it in the hands of a small number of company executives. It is not yet clear whether industry self-regulation under the banner of safety will become a practical safety guardrail or a means to strengthen large players’ market dominance.