US-China AI talks: A two-power bargain that could leave the world with no say

Two days before he goes on to welcome Xi Jinping to the White House, Donald Trump stood before the UN general assembly (UNGA), and rejected what he called a 'globalist scheme to control [AI]', and rebranded the tech for US official purposes as 'superintelligence'. While he was at it, he also declared that 'to the victor go the spoils', touting an oil deal with Venezuela after he launched a military operation in January, removing the country's president Nicolas Maduro.

Trump dismissed any warning of catastrophe as the same voices who once predicted doom from the 'hoax' of global warming. It was a revealing prelude. The US president will now be asking Beijing to talk about AI safety in Washington, after having just refused to accept any shared rules for it in New York.

Today's scheduled summit is Xi's reciprocal call after Trump travelled to Beijing in May, and will be their second meeting this year. Managing frontier AI risk and protecting IP top the agenda. The telling detail is 'who asked'. China's state media say the US pressed for the talks because it fears Beijing is gaining the upper hand, a worry sharpened after China signed AI partnerships with 28 mostly developing nations in Shanghai in July to establish the World Artificial Intelligence Cooperation Organisation (Waico).


Washington arrives holding the instruments of restriction: export controls to preserve a widening compute lead; strong models kept behind paid access; and the notion, floated among allies, that US systems should reach only trusted partners. Even Anthropic's frontier models have figured in these curbs.

Beijing's reply is to make containment expensive. Each new restriction, its commentators argue, only speeds domestic capability. This is visible in Huawei's Ascend chips and strong models from DeepSeek - even as China's regulators have opened a probe this week into DeepSeek and Moonshot AI after Anthropic accused both Chinese AI model developers of routing sensitive data through its Claude models.

Rather than gatekeeping, China distributes, offering cheap or free models that download whole, thereby courting the 'global south' through BRICS and a UN governance framework. The US sells a paid ladder; China offers an open door.

The leverage does not run one way. Which is where rare earths enter. China controls roughly 70% of their mining, 85% of refining, and close to 90% of magnet production, a chokehold on materials vital to semiconductor chips and the defence sector. In the run-up to the Washington summit, Beijing turned the dial, with magnet shipments to the US falling 21% in August.

The squeeze is timed to a tariff truce expiring on November 10, and a Pentagon Defence Federal Acquisition Regulation Supplement (DFARS) rule barring rare-earth magnets, tungsten, tantalum and molybdenum sourced from China (and Russia, Iran and North Korea) from January 1, 2027. Before Trump's May trip to China, Beijing reportedly offered a longer moratorium, which Washington rejected in hope of a better deal. This could be this week's summit's hardest bargaining play.

On AI, the achievable ground is narrow: not to collide by accident. The meeting on Sunday between US treasury secretary Scott Bessent and Chinese vice premier He Lifeng floated a notification mechanism for AI incidents that reach national security level. A statement reaffirming the 2024 consensus on keeping AI out of nuclear command is plausible. Both leaders are sceptical of doom, which cools the mood even as it caps ambition.

The contrast with the rest of the world could not be sharper. The same week, Norwegian PM Jonas Gahr Store and Finnish president Alexander Stubb launched 'A Call for Control of Frontier AI Models', endorsed by leaders from Germany, Canada, Singapore, the Netherlands and South Africa, along with the European Commission president and others. It urges shared standards, transparency on serious safety incidents, and an international body to verify limits when capability thresholds are crossed.

In his concluding address at the UNGA, secretary general Antonio Guterres pressed for responsible pacing, and credible, independent oversight. This is precisely the multilateral project Trump has just spurned, even as he seeks a private channel with Xi.

Observers worry more about what won't reach the table: energy demand of AI data centres, chokepoints of lithography and advanced packaging, researcher flows, quiet spread of open-weight models that sets de facto standards, and controlled inputs like gallium and germanium that shape the ecosystem more than any communique. Neither Beijing nor Washington seems keen to raise these real issues.

That silence feeds a wider contest of visions. The June G7 summit in Evian leaned towards a club of access for reliable democracies. The September BRICS summit in New Delhi staked out sovereignty, equity and 'south to south' cooperation, insisting the emerging world help write the rules.

India sits in both camps, as BRICS chair and G7 partner. It also hosted its own AI Impact Summit in February, where New Delhi pushed for a body giving emerging markets a real seat in rule-making. India's concerns are concrete: reliable compute access, exposure to the same mineral pressures, and the fear that a 2-rule-makers deal on standards leaves everyone else a rule-taker. The hope is that whatever channel Washington and Beijing open becomes a floor others can build on, not a perimeter wall around a private club.

The writer is policy adviser ondigital technology issues