The Extraterritorial Problem: Can Brussels Govern Chinese AI?

The EU AI Act, like GDPR before it, is written with extraterritorial ambitions. Article 2 extends the regulation’s scope to providers of AI systems placed on the EU market regardless of whether those providers are established in the EU. It extends to deployers established in the EU using AI systems whose outputs are used in the EU. The intention is clear: Brussels will not allow foreign AI companies to operate in the European market under different rules than European AI companies.

This is, in principle, the same logic that GDPR applied to data processing. Any company processing EU personal data faces GDPR obligations. This has been more or less workable for US companies with European business interests and legal infrastructure. Google, Meta, and Amazon all have substantial EU operations with legal entities that national DPAs can investigate and fine. They can be reached.

The question of whether Chinese AI companies can be similarly reached is more complicated, and the answer currently is: not really.

Who’s Actually in the Market

Chinese AI exports to Europe are more substantial than the policy debate typically acknowledges. The most prominent is DeepSeek, whose API has accumulated a significant developer user base across Europe since its breakout in early 2025. DeepSeek is a Chinese company — Hangzhou DeepSeek Artificial Intelligence Co., Ltd. — with no EU legal entity, no EU data centers, and no EU regulatory presence.

When a European startup or enterprise uses the DeepSeek API to build an application, they are the deployer of a system provided by a non-EU entity. The deployer — the European company — faces EU AI Act obligations. The provider — DeepSeek — faces them too, in theory. Enforcement against DeepSeek requires the EU AI Office or a National Competent Authority to take action against a Chinese company with no European presence.

Related reading

The Act contemplates this problem. It requires non-EU providers without EU establishment to designate an authorized representative in the EU — a legal entity that can receive enforcement communications. This is the same mechanism GDPR uses: non-EU data processors must designate EU representatives. In practice, GDPR-required EU representatives have faced limited enforcement. Companies can designate a law firm, a compliance service provider, or a shell entity that can receive notices but has no assets against which fines can be enforced.

DeepSeek has not publicly designated an EU AI Act authorized representative as of August 2026. They are technically non-compliant with the Act’s basic registration requirements. The EU AI Office has not announced enforcement action. There is a quiet acknowledgment in Brussels — not stated publicly, but visible in the gap between the rule and the enforcement — that compelling compliance from Chinese AI companies that don’t want to comply is not a tractable near-term objective.

Alibaba and Bytedance: A Different Story

Not all Chinese AI in the European market comes from companies without EU presence. Alibaba Cloud operates in Europe through data centers in Germany and the UK (technically non-EU post-Brexit, but Alibaba has Belgian and German infrastructure as well). Its Qwen AI model series is available through Alibaba Cloud’s European API. Alibaba has a Brussels policy office and legal entities in several EU member states.

ByteDance’s position is even more complex. TikTok operates in the EU with substantial infrastructure, dedicated legal entities, and an ongoing compliance relationship with the DSA (Digital Services Act) that has produced years of regulatory engagement. ByteDance’s enterprise AI products — marketed under the Volcano Engine brand — have a smaller EU footprint than TikTok but have legal entity coverage that makes EU AI Act enforcement at least theoretically tractable.

These companies have taken the AI Act more seriously than DeepSeek. Alibaba Cloud’s documentation for EU customers addresses GPAI requirements for Qwen. ByteDance’s EU AI Act compliance posture, while not publicly detailed, appears to include the basic GPAI documentation that the regulation requires for general-purpose AI models.

The distinction between Alibaba and DeepSeek is partly about business model and partly about exposure. Alibaba Cloud has major EU enterprise customers whose procurement processes require regulatory compliance documentation. Losing those customers to a compliance failure is a real business cost. DeepSeek’s EU users are predominantly developers paying small API fees — a market that is less sensitive to compliance documentation and more sensitive to price and capability. The enforcement risk calculus differs accordingly.

The Geopolitical Dimension

The EU AI Act’s extraterritorial enforcement problem intersects with EU-China trade relations in ways that complicate the regulatory picture.

The EU has significant economic weight over China — EU-China bilateral trade runs at approximately €700 billion annually, and market access negotiations touch numerous sectors. But applying that weight to AI regulation enforcement is a complex diplomatic move. The EU has simultaneously been pursuing trade dispute resolution in renewable energy (solar panels, electric vehicles), maintaining cooperation on climate, and managing the strategic ambiguity that characterizes EU-China relations. Escalating AI regulation enforcement into a diplomatic confrontation is not obviously consistent with these other interests.

Related reading

The result is a kind of implicit asymmetry: EU companies operating in China face China’s AI and data regulation, enforced with vigorous domestic authority. Chinese companies operating in Europe face EU regulation, enforced with considerably less vigor when the company has no European presence. This asymmetry is not unique to AI — it reflects a broader pattern in how democratic regulatory states interact with Chinese commercial actors — but it means that the EU AI Act’s nominal universality conceals a de facto differential treatment.

There are people in Brussels who are uncomfortable with this dynamic and people who are quietly comfortable with it. The uncomfortable view: regulatory asymmetry effectively subsidizes Chinese AI companies in the European market. The comfortable view: aggressive enforcement against Chinese companies with no European presence would be largely symbolic, would generate diplomatic friction, and might accelerate Chinese market retaliation against EU companies operating in China.

The comfortable view is probably more realistic. Symbolic enforcement actions — fines that can never be collected, orders that can never be executed — weaken the regulation’s overall credibility without achieving the underlying policy goal. The EU’s GDPR enforcement against US companies with no EU presence has been largely ineffective, as years of complaints about US data brokers who simply ignore GDPR have demonstrated.

The GPAI Model Audit Problem

The General Purpose AI provisions of the EU AI Act require providers of models above 10^25 FLOPs of computational training to undergo a systematic risk evaluation and provide technical documentation to the EU AI Office. Several Chinese AI models are plausibly above this threshold. The Chinese government has not published training compute estimates for its domestically developed foundation models. Chinese AI companies providing commercial API access have similarly not published technical documentation meeting GPAI standards.

This means the EU has a transparency gap at the foundation model level for Chinese AI. The US hyperscalers — OpenAI, Anthropic, Google — have published the required GPAI documentation, however imperfectly. Chinese model providers largely have not, and the EU AI Office has limited capacity to compel documentation from entities with no EU presence.

The practical consequence: European developers building on Chinese foundation models are doing so without the GPAI transparency that the regulation guarantees for comparable US-built models. Whether this represents a systemic risk depends on how those models are used. For entertainment, research, or creative applications — minimal to no risk. For applications where the foundation model’s training data, bias properties, and capability limitations matter — which includes any high-risk AI application — the transparency gap is a genuine governance problem.

The EU AI Act’s most ambitious regulatory goal — ensuring that the most powerful AI systems deployed in Europe meet a minimum transparency standard — is not being met for a significant and growing share of those systems. The Act can govern what it can reach. It cannot reach everything it governs.

The Data Governance Shadow

There’s a dimension to Chinese AI in the European market that goes beyond the AI Act’s compliance framework entirely: data governance. Under China’s Data Security Law and the Personal Information Protection Law (PIPL), Chinese companies are obligated to provide access to data — including data about non-Chinese nationals — to Chinese government authorities when requested. The legal architecture is explicit. The scope of what “national security” encompasses under Chinese law is broad.

Related reading

This means a European enterprise using a Chinese AI system is potentially feeding business-sensitive data into a pipeline that has a legally mandated back door for the Chinese state. The AI Act’s GPAI transparency requirements don’t address this. The Act is concerned with model transparency and application risk — not with the data sovereignty question of where information flows after it’s processed.

The European Commission has been aware of this concern independently of the AI Act. The DSA’s rules on data access for very large online platforms touch adjacent ground. The NIS2 cybersecurity directive creates obligations around supply chain security that can apply to AI service providers. But there’s no integrated EU framework that specifically addresses the data sovereignty risk of using Chinese-built AI for sensitive EU commercial purposes.

Germany’s BSI (Federal Office for Information Security) issued an advisory in June 2026 noting that German federal agencies and operators of critical infrastructure should treat Chinese AI service providers with the same supply chain risk assessment they apply to other Chinese technology vendors. The advisory was non-binding and limited to federal actors. It was also not widely reported, because it’s politically inconvenient: telling European businesses that their attractive, cheap, technically impressive Chinese AI services might be a security liability creates friction with EU-China commercial relations that nobody in Brussels is eager to create right now.

The EU AI Act, designed to address the harms AI can cause to individuals and society, wasn’t designed to address the geopolitical risk of AI supply chain dependency. These are different problems requiring different tools. But they exist simultaneously, and the regulation’s silence on the second problem shouldn’t be mistaken for its absence. The Chinese AI presence in European markets is growing, the AI Act can’t meaningfully constrain it for non-EU entities without EU presence, and the data governance implications sit in a regulatory gap between AI law, cybersecurity law, and trade policy that nobody has yet bridged.