Time to ‘pace’ AI development?

On 14 September, global artificial intelligence (AI)-related stocks suffered a sharp selloff. The Philadelphia Semiconductor Index fell as much as 5.2%, while AI-related stocks in Europe and Asia were also broadly sold off. Semiconductor and chip stocks have experienced steep declines before, but the immediate trigger this time was different: several of America’s most prominent AI entrepreneurs publicly called for slowing the pace of AI development.

In his essay “We Must Pace the Frontier”, Anthropic CEO Dario Amodei proposed slowing the development of the most advanced AI systems, a position supported by OpenAI’s Sam Altman and Elon Musk. Altman has even said that a 10% risk of AI causing human extinction by the end of the decade would be unacceptable. Earlier, frontier AI researchers had already warned that AI could potentially pose an existential threat to humanity before 2030.

A fear of losing control over AI

A single day’s stock market movement, of course, does not mean that the AI industry has undergone a fundamental reversal. But this market shock nevertheless sent a signal that had rarely appeared before: capital markets are beginning to seriously consider whether the people building the world’s most advanced AI systems might themselves slow technological development because they fear losing control of AI.

That stands in sharp contrast to the position of the US government.

Trump not only opposes putting the brakes on AI development, but has dismissed some warnings about AI risks as exaggerated alarmism. He worries that excessive restrictions would weaken US competitiveness and benefit China. His administration explicitly treats rapid AI development as part of the strategic competition between the US and China.

Developing too fast, or not fast enough

This has produced an interesting divide in the US: the government worries that American AI is not developing fast enough, while some of the country’s leading AI entrepreneurs worry that it is developing too fast.

China’s response shows a different tension. Beijing has rejected what it calls AI “fearmongering” and opposes using safety concerns to justify slowing China’s technological advance. Yet China’s state security chief Chen Yixin has also warned that advanced AI could threaten political security and critical infrastructure. The distinction matters: Beijing acknowledges AI risks, but not as a reason to slow development. If Washington reads Chen’s warning as evidence that China may also be prepared to decelerate, it would be a misreading.

Taken together, these developments matter because US-China competition in AI may now be entering a particularly critical moment.

So far, America’s advantages in this competition remain substantial. The world’s most advanced AI chips and much of the design ecosystem surrounding them are controlled by US companies. OpenAI, Anthropic, Google and other American firms remain in the first tier of frontier model development. The US also possesses the world’s strongest concentration of AI talent, capital markets, software ecosystems and cloud-computing capabilities.

Chinese models have advanced rapidly in recent years. DeepSeek, Kimi and others can now approach American frontier models in many practical tasks. Overall, however, the US still occupies the technological high ground.

China’s biggest weakness remains semiconductors. US restrictions on advanced chips and semiconductor equipment have not stopped the development of Chinese AI, but they have increased the cost of obtaining top-tier computing power. China’s domestic AI chips are improving, yet substantial gaps remain with Nvidia in advanced manufacturing processes, chip performance and the broader software ecosystem.

Pushback against data centres

However, as AI competition develops, the factors that will determine the outcome are no longer limited to chips and models themselves.

China is also beginning to display several advantages that previously received less attention, including electricity supply, application scenarios and the government’s organisational and planning capacity. At the same time, two factors that were rarely regarded as American disadvantages in AI competition are also beginning to emerge: growing public resistance to data centres, and concern among American AI entrepreneurs, represented by Musk and others, that AI development is moving too fast.

Surveys show that roughly 70% of Americans oppose the construction of AI data centres in their own communities, with nearly half strongly opposed. The reasons are straightforward. Data centres consume large amounts of land, water and electricity. Residents worry about higher electricity bills, noise and damage to their living environment, while such facilities require enormous investment but create relatively few long-term jobs.

Pressure to restrict or even temporarily halt large data centre projects has already appeared in a number of US states and localities.

This creates a problem that received little attention in the early stages of AI competition.

Trump can define winning the AI race as a US national strategic objective, but an American president cannot simply order a community to accept a data centre that local residents do not want. Local politics, public opposition, environmental approval procedures and the capacity of the electricity grid all become part of the decision-making process.

From this perspective, the US may possess the world’s most advanced chips but still be unable to build data centres at the pace AI companies would prefer. China’s chips may be less advanced, but it may find it easier to build data centres, power plants and electricity grids at the same time.

In the US, tech firms hold the cards

Compared with public opposition to data centre construction, the fear among AI entrepreneurs themselves that AI is developing too quickly is an even newer phenomenon, and a particularly American constraint.

The high degree of autonomy enjoyed by private companies in the US has long been regarded as one of the country’s greatest advantages in technological competition. The government does not need to tell companies how to innovate. Companies, driven by markets and profits, race ahead on their own.

AI may be the first technology to reveal another side of that institutional advantage.

Amodei, Altman, Musk and others who previously championed and promoted rapid AI development are now publicly and seriously discussing the possibility that AI could escape human control and even threaten humanity.

Although the US government opposes slowing AI development partly because it fears being overtaken by China, it cannot manage private AI companies as though they were government agencies. It cannot simply require a company to ignore what it believes are serious safety risks and continue training the next generation of models at a pace set by Washington.

If these entrepreneurs conclude that a particular generation of models is too risky, they can extend testing, delay release or voluntarily slow development. The US government has limited ability to interfere directly with such decisions.

China’s national strategy comes first

The same situation would be difficult to imagine in China.

Chinese AI companies also care about model safety, but their position within the country’s national strategy differs from that of American companies. Once Beijing defines AI as a core technology determining China’s national competitiveness, Chinese companies are ultimately more likely to subordinate the pace of development to national strategy than to the personal judgments of entrepreneurs about long-term risks to humanity.

It is difficult to imagine China’s leading AI companies jointly demanding that the entire industry slow down simply because their founders or researchers believe AI development has become too dangerous. This does not mean that the Chinese government is indifferent to AI safety. But it is indeed difficult to imagine Chinese AI companies publicly urging the government to reduce the pace of development.

That difference is rooted in China’s political environment.

A race between two national systems

In these newly emerging variables, China’s relative “advantages” are becoming more visible.

This does not mean that China is about to overtake the US. Technological progress in AI still depends heavily on chips, models, talent and original innovation, and the US continues to lead China in these areas. But the conditions that will determine the outcome of US-China AI competition are clearly no longer limited to these traditional technological factors. The changing structure of the two countries’ respective advantages and constraints is something the US should watch carefully.

US-China AI competition is shifting from a technological race among a handful of companies into a competition between the overall capabilities of two national systems.

From this point forward, the outcome may depend not only on which country can build a more intelligent model, but also on which system is better able to absorb the enormous resource demands, social resistance and risk pressures that will emerge as AI moves out of the laboratory and into the broader economy and society.

Related: China’s AI race: Can it make energy and computing work as one? | [Big read] China’s AI catch-up is forcing Silicon Valley to cut prices