The tough AI questions looming over the Trump-Xi summit

[NEW YORK] Artificial intelligence will be high on the agenda when US President Donald Trump sits down with China’s Xi Jinping at the White House on Thursday (Sep 21).

Both superpowers see AI as the new focal point of their economic rivalry. Trump declared on Sep 13 that “whoever wins AI, wins”. 

The question is whether the two nations can put aside their disagreements and cooperate in building a safe and sustainable AI environment. For now, neither trusts the other’s motives, and neither wants to slow down. 

Here are the biggest flashpoints in the US-China AI race: 

Chip smuggling and remote access

The US has restricted exports of the most powerful US-designed AI chips to China since 2022 on national security grounds. There’s a growing awareness that these measures haven’t been working as well as planned. 

Prosecutors in the US have pursued several major investigations into suspected smuggling of advanced AI chips made by Nvidia to customers in China. In the highest-profile case to date, the Justice Department earlier this year charged a co-founder of Super Micro Computer Inc, Yih-Shyan “Wally” Liaw, with helping send US-assembled servers containing Nvidia’s cutting-edge chips to Chinese customers in violation of the US export curbs. Liaw has pleaded not guilty, and Nvidia has said it has strict compliance procedures. 

When Chinese lab Moonshot AI released a large language model, Kimi K3, that almost matched the performance of leading US rivals, a White House official accused the company of illegally obtaining Nvidia chips and accessing them remotely through Thailand. China’s government has pushed back on the allegations, while Moonshot has not responded to them. 

US lawmakers have also claimed that Chinese AI firms are exploiting a loophole in export curbs by renting computing power located in other countries. A bipartisan bill, currently pending in Congress, aims to curb China’s remote access to Nvidia chips via cloud computing. The Commerce Department is also scrutinising ways to possibly use its export control authority to crack down on the practice. 

Distillation

The US government has accused half a dozen Chinese companies including DeepSeek, Moonshot AI, Z.AI and Alibaba Group’s AI unit of using “industrial scale” distillation to reproduce the capabilities of leading US AI models.

SEE ALSO

Bessent hails ‘very successful’ China talks on AI, trade

South-east Asia uniquely positioned to capitalise on AI boom: China-Asean panel 

Distillation is a widely used technique whereby a smaller AI model is created by copying the reasoning and behaviour of a much larger, more complex and capable model as a shortcut to avoid costly, time-consuming training. The issue arises when that is done without permission. China’s government has said the accusations are unfounded. 

Finding a compromise on the issue will be a challenge as the two nations would need to concur on what activities amount to improper distillation and agree to mechanisms for gathering evidence and confirming any unauthorized model extraction.

Catastrophic risks

This year, AI “agents” have emerged that are capable of completing complex tasks independently, and AI systems have begun to improve themselves with minimal human intervention.

This has sparked a wave of concern across the industry that AI is becoming so smart that humans will no longer be able to control it. Leading industry figures are warning that, without new guardrails, AI could do untold damage to the digital ecosystem and even threaten mankind. 

Anthropic CEO Dario Amodei warned of AI’s existential threat in a 3,800-word essay earlier this month that intensified the safety debate. He vowed to slow the company’s development of cutting-edge models in order to better understand the risks they pose, and US competitors including OpenAI pledged to implement similar measures. 

The Anthropic chief also called for more government regulation of AI and indicated that Washington would need to coordinate an international push on AI safety amid concerns that, even if US firms shift to a more cautious pace of development, Silicon Valley’s rivals in China might not. 

Amodei’s message prompted a sharp pushback from Trump, who rejected the Anthropic CEO’s calls for new guardrails on AI and dismissed the existential warnings as “a hoax”. 

Authorities in Beijing also reacted sharply to Amodei’s missive, calling it “fearmongering” and accusing US AI labs of “malicious competition”. Amodei has long antagonised the government in Beijing with his support for tighter US export controls to deny Chinese AI firms access to crucial equipment such as chips. He reiterated that call in his latest essay.

At the same time, China has expressed its own safety-related concerns, including that Anthropic’s Mythos model’s cyber capabilities could be used as a weapon against Chinese interests. 

Significant trust-building would be required for the two nations to craft a framework that would potentially slow the breakneck race to dominate cutting-edge AI. They would need to agree on what an unsafe AI system looks like, or what should be done when its capabilities go beyond controllable limits. 

For now, damage mitigation could be one area of potential cooperation. The US and China might work together to keep powerful AI models beyond the reach of cybercriminals and terrorists who might use them to attack sensitive infrastructure or design new bioweapons. 

Some security specialists have urged the US and Chinese governments to prepare for the possibility that a rogue AI agent might infiltrate a nation’s nuclear command system, or set off a cyberattack on a country that’s misread by the target government as a deliberate attack. One suggestion is that a hotline be established for one government to warn the other of an AI-triggered attack. 

AI humanoids

A race is under way to make commercially viable humanoid robots by endowing them with AI so they can better navigate the physical world. US firms have made many of the biggest advances in robotics, but Chinese companies are now driving ahead in pushing the widespread commercial use of humanoid robots. China wants greater access to Western markets for these products. 

Instead, the US has restricted imports of Chinese humanoid and quadruped robots on the grounds of national security. And in January 2027, procurement rules are due to kick in that would prohibit US defence contractors from sourcing a range of materials that were mined, refined, separated, melted or produced in China, in a bid to spur more production of those materials in other countries. 

China was able to use its dominant role in the supply of minerals that are critical to several US industries as leverage in its tariff showdown with the Trump administration last year.

China controls roughly 80 per cent of several critical layers of the hardware supply chain, including rare-earth magnets and the actuator assemblies required to build humanoids at scale, according to experts including those at the International Energy Agency and Jamestown Foundation. 

Elon Musk’s robot project, Tesla Optimus humanoids, has already been affected by Chinese export restrictions on magnets. If a compromise is eventually struck, the two sides would need to agree on safety standards for robots working in close proximity to humans before bilateral trade in humanoids can take off. 

Military AI 

US technology export controls are aimed partly at stopping the People’s Liberation Army (PLA) from accessing the most advanced US AI chips that might give it an advantage on the battlefield. 

Beijing’s military programs are ultra-secretive, but the priority is clear. At least seven university labs that support the nation’s armed forces have been trying to obtain advanced chips from Nvidia.

Two of the institutions that have expressed interest – Beihang University and Northwestern Polytechnical University – rank among China’s “Seven Sons of National Defence”, an elite group dedicated to aiding the PLA. BLOOMBERG