Now it is China that wants to restrict the West’s access to its AI

For years, Washington alone drew up the lists: what China could and could not buy, which chips were banned, and which machines from ASML – the Dutch company that manufactures the equipment essential for producing the world’s most advanced chips – were excluded from containers bound for Shenzhen. Now the roles have not been completely reversed, but they are beginning to even out: according to the Financial Times, Beijing is consulting its leading artificial intelligence experts to understand how to prevent the West from getting its hands on the country’s most advanced technologies and start-ups. The Ministry of Commerce has asked Alibaba, ByteDance and Zhipu whether and how to restrict the transfer of training data abroad and limit the ability of foreign users to download model weights. This is a sign that those who were playing catch-up have realised they now hold the advantage, and are therefore beginning to consider how to protect it.

This turnaround concerns the dogma of openness. The leading Chinese models, Moonshot and DeepSeek, are open-weight, downloadable and adaptable on anyone’s servers, whilst those from Anthropic and OpenAI remain closed. Open-source models were the lever Beijing used to erode American dominance by giving away what others were selling; they became a diplomatic talking point with the Global South – a means of commercial penetration and a symbol of supposed technological generosity. At the very moment that this lever proved effective, Beijing is now beginning to consider removing it, allowing foreigners access to the services but no longer permitting them to download the models.

There is also a parallel issue. Among the proposals under consideration are restrictions on foreign acquisitions of strategic groups in autonomous AI (so-called ‘agent-based AI’). The precedent cited is the deal whereby Meta acquired Manus for a sum of around two and a half billion, which was subsequently blocked by the Chinese authorities. Manus was founded in Wuhan and Beijing, before relocating to Singapore and closing its offices in its home country, having done so specifically to evade the cross-surveillance of Washington and Beijing. The Chinese government’s message was that simply moving the registered office is not enough, because the state looks at who created the technology and not where the articles of association are filed. Control transcends physical and geographical boundaries, and extends to those who can acquire national intellectual capital wherever it may be transferred.

However, cracks sometimes appear in this monolithic narrative. The very companies consulted have warned regulators that overly restrictive measures would slow their development and reduce China’s chances of winning the race, in the usual tension between state ambition and corporate interests. For Europe, and for those building products on top of these models, the scenario is one of two powers squabbling over who shuts what down, whilst those caught in the middle risk being left without open-source Chinese software and without American models on their own terms (which is also why Henna Virkkunen, the EU Commissioner for Technology and Cybersecurity, told the Financial Times that artificial intelligence has now become a geopolitical weapon, and urged Europe to develop its own alternatives to American models as quickly as possible, so as not to risk being cut off from access to strategic technologies).

The update under discussion in Beijing would, in any case, be the most significant in years, a sign of Beijing’s renewed confidence that it has achieved leadership in certain areas. After all, a wall is always built when there is something inside worth protecting.