China considers export restrictions on advanced AI models, leaving Europe caught in the middle

China's Ministry of Commerce has held talks in recent weeks with major technology companies about possible restrictions on the export of advanced AI models. Representatives of Alibaba, ByteDance and AI startup Z.ai were present at those meetings, Reuters reported. Companies such as DeepSeek, Manus, Moonshot AI and StepFun are also mentioned in that context.

No final decision has been taken. The scope and the effective date of any measures have yet to be determined. Even so, the move signals a shift: both major technology powers, the US and China, are increasingly treating AI models as strategic assets. For Europe, which over the past year and a half has rapidly adopted Chinese open-source models as a cheaper alternative to American services, this changes the playing field.

What is on the table in Beijing

According to Reuters, China is considering a tiered system. Basic open-source tools would require only a simple registration. More advanced technologies would be subject to a security assessment, while the most sensitive so-called frontier models would remain available for domestic use only.

The models potentially affected include Alibaba's Qwen, ByteDance's Doubao and Z.ai's GLM-5.2. Both closed-source and open-weight variants are involved, the kind of models that users can download and customise themselves. It is precisely those open-weight models that were widely embraced outside China for their low cost and flexibility.

In addition to access restrictions, the agenda also includes tighter rules on foreign investment in Chinese AI startups, and plans to make the leaking or theft of proprietary AI technology a criminal offence under the national security law.

The move fits a broader pattern. In April 2026, China ordered Meta to reverse its acquisition of Chinese startup Manus, valued at 2 billion dollars. In early June 2026, stricter rules for foreign business transactions followed. China has also already introduced export controls on dual-use goods that have affected seven European entities since 24 April 2026, including German defence company Hensoldt AG and Belgian arms manufacturers FN Herstal and FN Browning.

How quickly the use of Chinese models grew

The speed at which Chinese AI models gained ground outside China illustrates the potential impact of export restrictions. Following the launch of DeepSeek's R1 model, the global market share of Chinese large language models rose from 3 percent to 13 percent in just two months. On the routing platform OpenRouter, which tracks language model usage, the share of Chinese open-weight models grew from less than 2 percent at the end of 2024 to around 61 percent by mid-2026.

That growth was largely driven by price. DeepSeek's R1 offered performance comparable to leading Western models, but at significantly lower cost. For developers, research institutions and companies without the budget for the most expensive American APIs, that was attractive. If Beijing restricts access to the most capable models, that option disappears.

Europe caught between two blocs

Europe's position is an uncomfortable one. The continent has no frontier models of its own that can compete with American or Chinese frontrunners in terms of scale and performance. The EU has launched the InvestAI initiative with a budget of 200 billion euros aimed at building domestic AI capacity, but the planned data centres are behind schedule and the total budget looks modest compared with the investments of major American tech companies.

At the same time, Europe has also built up a broader dependency on China, including on rare earth metals needed for chips. The export controls Beijing has already imposed on dual-use items are also affecting European companies outside the AI sector. The combination of potential AI model restrictions and existing export measures increases pressure on European policymakers to gain clarity on their own technological position.

Meanwhile, in June 2026 China published a white paper on global AI governance, in which it called for the establishment of a World AI Cooperation Organization. This highlights a tension: on the one hand Beijing is tightening control over its own technology, while on the other it is positioning itself as an advocate of international cooperation on AI.

For European founders and investors who have built their infrastructure around Chinese open-weight models, the uncertainty is already a factor to reckon with, regardless of how Chinese decision-making ultimately unfolds. The prospect that access to both American and Chinese frontier models could be constrained by geopolitical considerations reinforces the debate within the European ecosystem about how dependent it is willing to be on external parties for core infrastructure.