Prior to the September 24 summit meeting in Washington of US President Donald Trump and Chinese President Xi Jinping, a rising cacophony of experts in the field of AI called for the two leaders to declare a slowdown in the development and release of AI products until information technology (IT) leaders and corporations can get a better grip on corralling the cybersecurity, economic, and social risks of the technology.

The leaders did not do that. Instead, the two governments announced the establishment of a vaguely defined AI “dialogue,” which is set to meet again in November. Meanwhile, US Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng agreed to the creation of an “AI incident communication channel” — essentially a promise to pick up the phone if one side reaches out to the other to share information or ask for help in case of a severe AI-created or AI-enhanced cybersecurity problem that involves the two nations.

This agreement is an important and positive first step. But — as predicted by The Economist and many other thoughtful observers of US-China relations — China and the United States are not ready to slow down on AI. In fact, both nations’ leaders view the new technology primarily as an arena for competition rather than cooperation.

Global competition is the main motivation

According to China’s Xinhua News Agency, Xi told Trump that “the two sides can continue conducting dialogue on artificial intelligence, exchange views on its risks and benefits, and work together to prevent the misuse and malicious use of AI [and] maintain human control over artificial intelligence.”

In Washington and elsewhere, President Xi is tipping his hat toward the idea of making sure that AI is used for the common good. But he is not making any promises to slow down. This reflects the fact that, as summarized in the Wall Street Journal, “those building and regulating China’s leading AI models don’t think they have built the 21st century’s version of a nuclear weapon yet — and they aren’t going to slow down while America refines its own.”

Indeed, in his travels to Central Asia, India, and Africa prior to coming to Washington, President Xi’s top theme was enticing the Global South to buy into China’s vision for a global AI economy powered by Chinese technology.

Trump, meanwhile, has vigorously rejected slowing down or more tightly regulating the advancement of AI technology. He has referred to fears about AI as “hoaxes,” and said, about cooperation with China, “When you’re leading, you don’t open it up to each other.”

On Truth Social, President Trump has proclaimed that AI should be called “Super Intelligence,” and said, about his meeting with President Xi, that “Super Intelligence (SI) will be a big topic of discussion, but I want to leave it exactly where it is. That is China’s position also.”

In retrospect, it has become clear that the AI dialogue agreed to by the United States and China was engineered primarily by Bessent, rather than Trump. This appears to be due to the Treasury Secretary’s concerns about the possibility that rogue AI programs could upset the stability of US and global financial markets.

Applying a bit of game theory to the situation

This outcome, in terms of the US-China summit in Washington, and deeper leadership attitudes in general, is highly predictable if viewed from the perspective of basic game theory. The classic “prisoner’s dilemma” analysis predicts that one side will only make a concession if it trusts the other side will follow suit. In other words, the deep distrust prevailing in US-China relations precludes the possibility that either Beijing or Washington will “slow down” on AI if they view the situation mainly as a race for supremacy — and that is exactly how they view things. “Slowing down” could allow the other side to get ahead.

That said, however, government-to-government cooperation should still be possible on setting boundaries and stemming AI externalities that do not ask either side to “slow down.” This is why Bessent and Vice Premier He may be able to make progress in working together to fight cybercrime or malicious hackers using AI (if the hackers are not state actors from either side!).

Three years ago, Xi and President Joe Biden were able to agree that AI systems should not be given control of nuclear weapons.

It would be reasonable to expect that the upcoming US-China official AI dialogue could expand such restraint to other dangerous “use cases” for AI technology, such as election interference or malicious information campaigns. Perhaps the governments could also agree on standards for combating malicious use of deepfakes or other socially damaging exploitations of AI capabilities.

Even so, the only likely cooperation between Beijing and Washington on AI safety will be on “use cases,” rather than encompass more general technological slowdowns or banning certain categories of AI technology.

AI market structure promotes both overlap and competition

If US-China competition will outstrip cooperation, what does that mean for the evolution of the global AI market in the private sector? Will we see a repeat of the competitive pattern of distrust spurring ever-increasing decoupling between the US-led and China-led ecosystems — such as we have seen in semiconductors, computing, telecommunications, and many digital services such as social media and e-commerce?

In IT hardware, US technological superiority and the urge to “stay ahead” of China led to a concerted policy effort to deny the most advanced semiconductors and computing power capabilities from being exported to China. This reinforced China’s incentives to spend enormous resources to create indigenous capabilities as rapidly as possible. This story is well known and well documented, causing many to expect the same competitive dynamic to lead to policy-led decoupling in the AI space as well.

But AI technology may be different from IT hardware. First, it is not clear that either “side” is very far ahead of the other. Second, the extreme fluidity of the technology, and its lightning-fast progression, means that technology denial to the “other side” may be nigh impossible.

Already, Chinese and American AI products are being sold in the same markets and used almost interchangeably by customers. Statistics compiled by OpenRouter suggest that, despite the broader geopolitical backdrop, global users — particularly smaller companies, including US companies — are adopting DeepSeek and other Chinese models while hosting them on US-based servers. AI aggregators make this possible by allowing users to seamlessly switch from one AI model to another depending on the use, purpose, and preferences on cost.

At the same time, there is also significant segmentation in the market. Chinese models are predominantly open-weight, allowing users to piggyback freely and download parameters to their own hardware. They are also lower-cost, using cheaper electricity and leveraging efficient, inexpensive engineering. Roughly speaking, these models seem to be capturing about half the market.

The most famous American models are closed-weight, offering more sophisticated use cases at higher cost, with the underlying codes and training data kept private. They are also generating enormous revenue, while leveraging eye-watering amounts of optimistic investment in the process.

This situation — coexisting technologies that are simultaneously succeeding and improving — reinforces the market dynamics that would make it difficult for either the US or Chinese governments to try to separate the technology into distinct ecosystems. When the US government briefly toyed with blocking international access to advanced closed-weight models made by US companies, the howls of protest came not only from Beijing and Silicon Valley, but from around the world.

The rest of the world also matters

This points to the other reason why US-China technology decoupling may be slow, or impossible, in the realm of AI.

Simply put, building a wall to deny technology from one to the “other” will not be supported by non-US and non-China markets. The rest of the world has an appetite for both competitive sets of AI technologies.

This situation means that unilateral technology denials by government policy — to prevent technology “leakage” to the “other” — must be global in nature and thereby create a wall around one’s own market. But that will cede the rest of the world to the “other.” Such an approach does not seem economically sustainable.

Indeed, it is noteworthy that the Trump administration decided not to implement the “AI diffusion rule” written in the final months of the Biden administration, which aimed to use US superiority in semiconductors to block third countries from providing AI data center access to Chinese users and Chinese AI service providers. Technology denial via third markets will be a hard thing if those markets want to do AI business with both the United States and China.

In the end, it is not clear that either the United States or China can “win” the AI “race,” to the exclusion of the other. But as the technology evolves and creates both good and bad changes in our societies and economies, we can hope that Beijing, Washington, and other governments as well will work together to make sure fewer people “lose” (their lives or livelihoods) from this seemingly inevitable global technology.