China eyes export curbs on its top AI models, and Europe is caught in the middle

China is weighing restrictions on foreign access to its most advanced AI models. Both superpowers now treat AI as a strategic asset. Europe's already weak bargaining position could erode further as a result.

Chinese authorities held talks last month with leading tech companies about potentially restricting foreign access to China's most advanced AI models, Reuters reports, citing three people familiar with the discussions. The restrictions would also cover unreleased models. Alibaba, ByteDance, and the startup Z.ai all attended the meetings, according to the report.

The talks, led by the Ministry of Commerce, covered limits on high-performance systems, both closed and open models. Officials considered placing theft or transfer of protected AI technology under China's strict national security law, according to one source. They also discussed tighter controls over who can fund domestic AI startups. The scope of any rules is still being debated and may only apply to future models. Whether and when they take effect remains an open question.

Since DeepSeek's R1 launched, Chinese models have gained ground worldwide thanks to low costs and growing capabilities. Alibaba's Qwen and ByteDance's Doubao rank among the most widely used models in China, and Z.ai recently made waves with GLM-5.2, which approaches US frontier performance at a fraction of the cost. If Beijing limits access, costs for many companies would rise.

Beijing has been shielding its AI for a while

The move fits into a broader pattern of protectionist steps. In April, China's state planning agency ordered Meta to unwind its $2 billion acquisition of the Chinese-founded startup Manus. Tighter rules on overseas business dealings followed in early June, covering Chinese investors, technology, and data. China also investigated Manus and other startups that had moved abroad for possible export control violations.

What access restrictions might look like was sketched out at an expert panel in May, whose summary appeared in a journal of the Supreme People's Court, according to Reuters. The panel proposed a tiered system. Basic open-source tools would only require registration. Advanced technologies would need a security review. The most sensitive frontier models would either not be released publicly at all or be limited to domestic use.

The US set the precedent

China is mirroring an approach Washington already took. The Trump administration worries about foreign military and intelligence agencies misusing American AI. In June, it barred foreign nationals from accessing Anthropic's most advanced models, Fable and Mythos. Because nationality couldn't be verified in real time, Anthropic initially shut the models down worldwide. Export controls on Fable were lifted after new safeguards were put in place. Mythos, designed for cybersecurity professionals, remains restricted to a handful of trusted US organizations and vetted partners.

Chinese officials fear Washington could use Mythos to exploit software vulnerabilities against Chinese interests, according to Reuters. Zhou Hongyi, founder of security firm 360, is calling for a Chinese equivalent to Mythos.

Europe's easy alternative is disappearing

For Europe, this is more than a footnote. Both superpowers now treat their best AI models as strategic assets they'll withhold if needed. Until now, many saw China's open, freely downloadable models as a sovereign alternative to pricier US services. But China's proposal to keep its most sensitive models domestic shows that path is unreliable too. Companies relying on cheap Chinese tech today could find themselves locked out tomorrow. Europe still largely lacks its own competitive offering, with the exception of French provider Mistral. A report by former ECB chief Mario Draghi on European competitiveness spelled out just how large the digital gap is: the EU depends on foreign providers for more than 80 percent of all digital products, services, and infrastructure. A report by Germany's Expert Commission on Research and Innovation (EFI), published in February 2026, painted an equally sobering picture of AI in Germany and Europe.

As long as Europe carries weight as a trade and alliance partner, access will likely hold in most cases. That weight rests on two pillars that remain stable for now: a large single market of roughly 450 million consumers and the power to set global standards through regulation (the Brussels Effect). But the economic prosperity behind those pillars depends heavily on high-value knowledge work, the well-trained engineers, developers, and researchers Europe brings to global trade. And AI threatens to erode the value of exactly that kind of work.

The software industry is already flashing a warning sign. Demand for entry-level developers is dropping because AI tools handle simple coding tasks and companies are filling fewer junior positions. That breaks the traditional pipeline from beginner to experienced developer. What's hitting junior programmers today could reach engineers and researchers tomorrow, once models can routinely do their work. There's no reason to assume AI progress will suddenly stop in the coming years.

An old problem gets a second drain

The brain drain from Europe isn't new. Promising European companies have been snapped up by foreign buyers for years, from AI lab DeepMind to chip designer ARM to Finnish company Silo AI, which AMD acquired in 2024 for $665 million. The root cause is a chronic funding gap. According to the European Investment Bank, more than four out of five major funding rounds in the EU have a foreign lead investor. In San Francisco, the figure is just 14 percent. Whoever writes the check often moves the company abroad or buys it outright. Europe keeps losing the firms that could have become industry leaders.

AI is now opening a second, faster channel for that same leak. Expertise can be bought in small doses and fed directly into a model. Platforms like Mercor connect specialists with AI labs that use their knowledge as high-quality training data. The market is shifting away from massive volumes of simple examples toward smaller, expert-curated datasets. Human expertise flows into training and evaluation data, where it can permanently boost the performance of AI models.

For Europe, both trends add up to a double risk. The continent keeps losing companies and talent to foreign owners, just as it always has. But now the expertise of its specialists also drains away as training material for models Europe barely owns. While manufacturing once moved offshore to cheaper countries, what's migrating now is intellectual capital, flowing into foreign AI models instead.

Without larger public investments, political will, and homegrown capabilities, Europe stays caught between two powers that increasingly treat AI as an instrument of leverage.

Europe is investing now, from a position of weakness

Europe isn't sitting idle anymore, at least. Commission President Ursula von der Leyen announced the InvestAI initiative in early 2025, aiming to mobilize roughly 200 billion euros for AI. That includes 150 billion euros from the private "AI Champions" initiative and 50 billion euros in additional European investment. Twenty billion euros are earmarked for up to five so-called AI Gigafactories, large data centers meant to give Europe its own capacity for training frontier models. The EU Council adopted the legal framework in January 2026.

Implementation is lagging, though. The formal tender for the Gigafactories has been pushed back several times and is now expected in summer 2026. Construction of the first facilities is scheduled to begin in 2027, so any real operational impact won't come before then at the earliest.

Even the full European effort looks modest on a global scale. According to current analyst and media estimates, the four largest US tech companies, Amazon, Alphabet, Microsoft, and Meta, could spend a combined $700 billion on AI-driven investments in 2026 alone. That's nearly double what they spent the year before. Four American companies are putting roughly three times as much into AI in a single year as Europe's entire multi-year initiative envisions.

At the same time, Europe is loosening its own rules. The "Digital Omnibus" package from November 2025 would push back key obligations under the AI Act for high-risk applications to 2027, and in some cases 2028. The official goal is to give domestic companies breathing room against US and Chinese competitors. Critics see it as caving to pressure from big tech and the US government.

Whether money and relaxed rules will be enough to break the double dependency remains an open question. So does whether all the players involved truly grasp the scale of the technology. Some in Europe seem to be banking on an AI bubble bursting or on leapfrogging, the idea that a clever shortcut could skip over the lag in compute and data. Recent developments point the other way.

Raw compute, massive datasets, and hands-on experience produced models like Mythos that upended established cybersecurity practices in short order. Recent studies also suggest most benchmarks actually underestimate AI agents. That's why Washington and Beijing treat these systems as strategic assets. Betting that this trend will fizzle on its own ignores the fact that what drives these models, compute and data, is real and hard to catch up on. For Europe, that means taking reality seriously instead of hoping it goes away.