China’s proposed AI export rules draw a consequential line between using Chinese artificial intelligence and possessing the capability to operate, adapt or reproduce it independently. That distinction may preserve the global reach of Chinese AI providers while keeping valuable model assets, training knowledge and strategic discretion within China. For European firms, the attraction of competitively priced Chinese AI services could therefore coexist with a longer-term loss of control over deployment, switching and technical learning. [2]

What changed

On 16 October, China’s Ministry of Commerce and Ministry of Science and Technology opened consultation on revisions to the country’s catalogue of prohibited and restricted technology exports. The draft covers 40 technologies across AI, biomedicine and photovoltaics; 23 proposed additions concern large models, AI agents, computing clusters, chip design and information processing. Consultation is scheduled to close on 15 November. [1]

The proposal is not yet a final export-control regime. But it has immediate operational significance: exporters of technologies listed in the draft must register planned transfers while the consultation and revision process continues. Registration covers transfers of ownership, overseas licensing, changes that place ownership or effective control in foreign hands, and other cases specified by the ministry. It also includes transactions already under way but not completed. [1]

The draft would prohibit some AI technologies and restrict others, including core large-model technologies, advanced agents, AI-chip design and large-scale computing-cluster technology. The proposed controls extend beyond AI to biopharma and photovoltaic technologies. [1] [4]

A deliberately asymmetric export model

The most important feature is the ministry’s distinction between cross-border transfer of controlled technology and remote delivery of AI-enabled services. China says overseas localisation of a restricted model is within export-control scope when the underlying technology is transferred abroad. Yet provision through API calls, applications, web interfaces, software-as-a-service offerings and lightweight local edge deployment would generally not count as an export of large-model technology. [2]

That creates an architecture in which foreign customers may retain practical access to Chinese AI capabilities without necessarily receiving model weights, key checkpoints, architecture designs, training-optimisation tools or inference know-how. In commercial terms, it favours a hosted-service relationship over a transfer-and-replication relationship. Reporting on the draft describes this as a distinction between overseas users consuming services and foreign firms acquiring the models, data or expertise required to run them independently. [3]

This should not be read as a blanket Chinese prohibition on AI internationalisation. The official explanation says that most AI technologies would remain freely exportable. Technologies already made public through open source or academic publication are proposed for exclusion, while Chinese companies’ overseas self-built and self-operated computing centres may receive automatic licensing or similar facilitation. [1] [2] The policy is selective: it preserves routes for commercial expansion and parts of the open-source ecosystem while retaining stronger oversight where technology transfer could create an autonomous foreign capability.

Why it matters for Europe

For European adopters, a service-access model can be efficient in the short term. It reduces the need to acquire and maintain costly model infrastructure, and can widen access to capable systems where domestic alternatives are limited. But access is not sovereignty. A company dependent on an externally hosted model is exposed to price and performance changes, service continuity, model updates, data-processing arrangements, contractual limits and the provider’s compliance with its home jurisdiction.

The proposed Chinese rules add a further layer: a European customer may be able to call a model, but face a more difficult route to localising it, obtaining the technical artefacts needed for independent operation, or acquiring the relevant business unit and intellectual property. The immediate dependency would be on a provider’s cloud or application interface; the deeper dependency would be on a regulatory permission structure in which Beijing can distinguish acceptable service provision from sensitive technology transfer. This is an analytical implication of the draft’s design, not a declared Chinese policy objective. [1] [2]

That distinction matters especially for sectors where continuity, auditability and adaptation are essential: public administration, healthcare, finance, critical infrastructure and industrial engineering. European organisations may need access to model logs, security testing, incident response, fine-tuning options and reliable migration routes. A service that remains legally available can still be strategically constraining if the user cannot validate, modify or replace the underlying capability on acceptable terms.

Strategic and policy implications

  • Procurement must distinguish consumption from control. European buyers should assess whether they receive an API service, deployable software, model weights, documentation sufficient for independent operation, or rights to fine-tune and move workloads. These are materially different dependencies.
  • Cloud location is not the whole question. Chinese guidance offers facilitated treatment for a Chinese company’s own overseas cloud deployments, but this does not automatically create equivalent rights for a foreign customer to receive controlled technology or reconstruct the service independently. [2] [3]
  • Open-source claims require scrutiny. Publicly disclosed technologies are proposed to fall outside the controls, but an open-weight release may not convey the training data, engineering processes or compute capacity needed to reproduce a frontier system. [1] [4]
  • European resilience policy should be practical. Contractual exit rights, portability requirements, security review, data-governance assurance and tested substitutes may do more in the near term than abstract calls for autonomy.

What to watch

The key uncertainty is implementation. The draft leaves questions over the benchmarks that determine coverage, treatment of foreign-owned servers, and application to complex model architectures and corporate transactions. [3] Europe should also watch whether the final text preserves the broad exemption for hosted access, how often automatic licensing is used, and whether Chinese providers offer differentiated domestic and overseas model versions. The central policy question is not whether Europe should use Chinese AI. It is whether European users can obtain its benefits without turning a commercial interface into a lasting strategic dependence.