In the United States, public skepticism over the misuse of AI technology and cybersecurity risks continues to grow. Real-world problems such as controversies surrounding data center construction and vulnerabilities in AI models have become increasingly prominent. Recent cybersecurity incidents, including cases in which OpenAI models reportedly infiltrated external organizations, have also become issues in the midterm elections, prompting U.S. lawmakers to address the immediate and tangible risks posed by artificial intelligence.
In September, the “AI slowdown” became a major point of contention among Silicon Valley’s AI giants, and the debate quickly spread to Wall Street investment circles as well as the White House and Congress, the centers of political power in Washington.
Put simply, the “AI slowdown” argument, built around safety concerns, runs counter to the Trump administration’s narrative that “if AI wins, America wins.” The U.S. political establishment typically frames domestic agendas around the “China threat” before eventually seeking a compromise between competing interests.
With Chinese leader Xi Jinping scheduled to visit Washington this week, U.S. media have seized on the opportunity to elevate the issue from a peripheral aspect of U.S.-China competition to a major agenda item. Public opinion has called for Washington and Beijing to reach an “AI safety agreement” or “AI safety consensus” in order to prevent the development of AGI from “destroying humanity.”
Stay ahead of the geopolitical week.
MD Briefing delivers expert analysis across five global fronts — the Indo-Pacific, energy, geoeconomics, European security, and the Middle East — every Monday morning. Free.
However, U.S. public opinion generally expects that China will not reach a high degree of consensus with the United States on AI, let alone sign an agreement. Meanwhile, Chinese public opinion has largely dismissed the U.S. “AI slowdown” argument, almost unanimously viewing it as a strategy designed by the United States to preserve its technological lead.
Here is how the story unfolded.
Anthropic CEO Dario Amodei was the first to sound the alarm, warning that AGI could acquire the ability to take over the entire internet within the next six to twelve months. He fears that without advanced safety guardrails, the scale of potential destruction could continue to expand.
In simple terms, Amodei believes that the pace of AI self-improvement is exceeding expectations and that AI could soon—or may already—be slipping beyond human control, ultimately producing catastrophic consequences. For the general public, AGI can be thought of as “self-improving AI,” something that once belonged to science fiction.
His remarks quickly received support from several other industry leaders, including Elon Musk of xAI, Sam Altman of OpenAI, and Demis Hassabis of Google DeepMind. At the same time, they drew opposition from other industry figures, including Jensen Huang of Nvidia and Mark Zuckerberg of Meta, whose Llama models are open-weight.
The key phrase in this story is “trip up.” Some AI giants may seek to maintain their lead through “safety regulation,” while their opponents firmly believe that “someone is trying to cheat.”
It is important to clarify that the so-called “AI slowdown” refers to slowing the training of large-scale models. It does not necessarily mean a reduction in demand for computing power, nor does it necessarily imply a reduction in total investment.
Demand for computing power can be divided into three areas: training, inference, and deployment/safety evaluation. Anthropic is advocating a shift away from heavy investment in large-model training toward the application layer—that is, inference and deployment/safety evaluation. Such a shift would benefit its monetization model by giving its existing products a longer period in which to generate revenue.
Given that Wall Street investors are now placing greater emphasis on the profitability of AI companies, while the cash flows of major AI firms are gradually coming under pressure, companies such as Anthropic urgently need to create a “new narrative” to sustain high valuations.
Therefore, slowing large-model training could help closed-source models maintain high margins. At the same time, strict “safety regulation” would significantly raise the barriers to entry for newcomers, creating a natural regulatory moat around incumbent AI giants.
The opposite logic applies to Meta. Its open-source strategy is ultimately designed to strengthen its core platforms—Facebook, Instagram, and WhatsApp—by attracting millions of developers around the world to build on and improve its AI ecosystem. This allows Meta to share the enormous cost of AI development with the global developer community while preventing a small number of competitors from monopolizing the underlying technical standards. The strategy depends on rapid iteration and widespread adoption of large models. Strict “AI safety regulation,” however, could impose high barriers on model development, effectively shutting millions of developers out of the competition and undermining the very ecosystem on which Meta’s strategy depends.
In other words, the dispute can be viewed as a battle between the open-source and closed-source development paths. Strict regulation favors technological leadership based on a “small number of developers” while making it more difficult for a “large developer ecosystem” to participate in the competition because of higher barriers to entry.
As for hardware supplier Nvidia, although an “AI slowdown” would not necessarily reduce total demand for computing power, large-model training consumes more computing power per workload than inference and deployment. Training therefore drives rapid GPU upgrades and helps Nvidia maintain high profitability. In other words, slowing large-model training would reduce customers’ incentives to purchase new GPUs and lengthen the infrastructure refresh cycle, severely hurting Nvidia’s revenue growth rate.
Then there is Elon Musk’s perspective. xAI is heavily focused on massive physical industries involving Tesla, SpaceX, robotics, satellite communications, and other areas. Musk clearly wants Wall Street investment to shift toward applications, and a slowdown in large-model training provides a credible pathway for redirecting capital in that direction.
For venture-capital firms on Wall Street, a situation in which a small number of AI giants use strict regulation to establish monopolies would naturally be unfavorable to their investments in industry newcomers. This explains why David Sacks, a prominent voice for the venture-capital industry and co-chair of the President’s Council of Advisors on Science and Technology (PCAST), has also strongly opposed the “AI slowdown” argument.
Finally, there is Trump’s position.
The “AI boom” is one of the most important achievements claimed by the Trump administration. Its policy priorities include deregulation, expanding infrastructure, and establishing America’s global technological advantage. Massive data-center investments can be presented as manufacturing reshoring; new power-generation and transmission projects can be counted as infrastructure achievements; and rising technology stocks directly reinforce the political narrative that “America is leading again.”
Put simply, America’s AI arms race embodies Trump’s ambitions while also constituting a key arena in which the United States seeks to maintain its lead over China. Any proposal to “slow down” AI conflicts with this narrative. It is therefore unsurprising that Trump has described the AI slowdown argument as a “sick conspiracy” and rejected the idea of “robots taking over the world.”
Unlike the technology right, Trump’s instinctive response to the AI slowdown argument is to invoke the idea of “trip up China”—a message that ordinary voters can quickly understand and identify with.
In fact, Amodei is one of the Silicon Valley CEOs most actively advocating restrictions on the development of China’s AI industry. He has long argued that the United States and its allies should take tough measures to constrain the development of Chinese AI companies, with the core instruments being restrictions on the sale of semiconductor chips and manufacturing equipment.
But the question now is, how can Dario Amodei explain why slowing AI would help contain the development of China’s AI industry?
The answer is “let politics be politics.”
Dario Amodei has advocated forming a “Democratic Entente” to negotiate international treaties. Through “AI safety,” the West could obtain the power to set international standards and then pressure China to join an international AI safety treaty. If China refused, the Western alliance could impose economic and technological isolation on China.
However, the biggest blind spot in this proposal is Europe. As one of the West’s principal economic powers, would Europe be willing to accept American AI safety standards? If Europe is serious about “AI sovereignty,” it is unlikely to accept American dominance over AI safety standards.
The meaningful question, therefore, is this: given Europe’s relatively disadvantaged position in AI today, would an open-source path or a closed-source path be more conducive to achieving “AI independence”?
If this question were posed to any AI chatbot—even Anthropic’s Claude—it would likely answer that an open-source model is more favorable to Europe’s pursuit of “AI independence.” Of course, a hybrid model combining open-source and closed-source approaches should also be considered. But before Europe produces its own AI giants, prioritizing the open-source path would allow a latecomer to catch up at a relatively much lower cost.
In other words, accepting American AI safety standards would not necessarily serve Europe’s development interests. After all, Amodei’s logic is fundamentally unfavorable to latecomers.
The same logic applies to Japan, South Korea, India, and other countries seeking to cultivate their own AI industries.
This is why China is unlikely to accept an American AI safety initiative. Xi Jinping may make vague and ambiguous diplomatic remarks while in the United States, but he is unlikely to fall into a trap that could constrain China’s own development.
China is not unconcerned about AI safety. Rather, its development path and industrial ecosystem differ from those of the United States, meaning that its security priorities and those of Anthropic are unlikely to align. Moreover, the strategic logic behind the initiatives of Anthropic and similar AI giants is so obvious: a small number of AI oligopolies want to further reduce competition in order to strengthen their monopolistic positions. Even the U.S. president does not buy into the argument. Why, then, would Beijing cooperate?
As an AI competitor that stands almost on equal footing with the United States, Beijing’s inevitable response would be to propose a Chinese version of AI safety standards while seeking to accommodate the interests of the Global South and even Europe and other latecomers. It would use open-source models as a major source of competitive strength and promote the narrative of “technology for all and win-win cooperation.”
Therefore, if ideological considerations are set aside, China may face fewer obstacles than the United States in building an AI safety coalition in practical terms, because China would advocate for multiple countries to jointly formulate AI safety standards rather than allowing China itself to dominate the process.
China’s AI industry will not slow down—at least not for the reasons or along the path proposed by the United States. Consequently, whether Anthropic’s slowdown initiative succeeds will ultimately depend on whether Wall Street’s market valuation of AI and Trump’s “political valuation” of AI change.
