Introduction

Chinese AI models have entered the global first tier in terms of technical capability, with API call prices significantly lower than mainstream overseas products; combined with domestic electricity cost advantages, AI capability going global is turning from an "idea" into real business for a large number of manufacturers. But exporting a product has never been as simple as loading equipment onto a cargo ship. An AI model embedded in Chinese industrial equipment may fall within the Catalogue of Technologies Prohibited or Restricted from Export in China when exported; sensors and payloads on the equipment may fall within the Dual-Use Items Export Control List; data generated and transmitted back by the equipment's operation must pass a data export security assessment; and entering the EU market means simultaneously facing the Artificial Intelligence Act, the new Machinery Regulation, the GDPR and a series of industry rules. This article takes a fictitious but common Chinese industrial equipment manufacturer as its subject, breaks the three questions of "model form, processing location, data destination" into a scenario comparison table, and works through the three lines of export control, cross-border data and EU AI regulation one by one, focusing on answering the three practical questions "how do model weights go abroad, does an OTA update count as continuous export, and do we still need to worry about data kept overseas", and finally provides an actionable self-check list, transaction structure and contract clause recommendations.

01. Introduction and Scenario Setting

I. Going global is becoming reality, and compliance thresholds are multiplying

Several noteworthy things have happened in China's AI industry over the past two years. Models such as DeepSeek, Kimi and Tongyi Qianwen have entered the world's front ranks on multiple benchmarks, training and inference costs have been driven very low, and API prices are far below overseas products such as Claude and ChatGPT. At the same time, domestic electricity costs are relatively low, and some companies have begun packaging "cheap tokens" and "cheap computing power" into products and services sold overseas. Demand on the market side is real: overseas customers have a clear preference for cost-effective industrial intelligence solutions.

But changes on the regulatory side are equally real. In August 2020, the Ministry of Commerce and the Ministry of Science and Technology added "personalized information push service technology based on data analysis" and "AI interaction interface technology" to the catalogue of technologies restricted from export. In September 2024, three departments imposed export controls on specific drones and their payloads, with controlled items including infrared imaging equipment and synthetic aperture radar meeting technical specifications — that is, "components with AI recognition capability" were directly listed as controlled according to product parameters. In December 2024, the Regulations on Export Control of Dual-Use Items and the Dual-Use Items Export Control List came into force simultaneously. In July 2026, according to relevant reports, regulators are discussing with major domestic AI companies the implementation of tiered export controls on frontier AI models, involving model weights going abroad and cross-border flows of core training data; the draft rules have not yet been published.②①

Looking at these two lines together, Chinese smart device manufacturers are at a specific compliance pressure point. When a product goes abroad with a model, the model itself may be subject to export control constraints; the data generated by the equipment must be processed and aggregated, which may trigger a data export assessment; and when the product enters the EU, it must also satisfy EU AI and cybersecurity rules. The superposition of the three is no longer a problem that can be solved by "obtaining a licence when selling goods".

The following text first breaks this type of scenario into several typical forms, then answers three things layer by layer: how the law regulates, how the model goes abroad, and what companies should do.

II. A Typical Scenario: Company A's Equipment Going Global

The subject of analysis in this article is set as a fictitious but common domestic equipment manufacturer, hereinafter referred to as Company A. Company A develops and produces industrial products with embedded self-developed AI models from mainland China; the product form may be industrial robots, AGVs, intelligent inspection equipment, AI medical devices, or intelligent connected vehicle components. The products are exported to the EU, and software and model update services are continuously provided throughout the product life cycle. During use in the EU, the products generate operational data, working-condition data and personnel-related data; these data need to be processed by the model, and in some scenarios are also aggregated, with the aggregation location possibly being local in the EU, within China, or a third country.

This setting is not designed out of thin air. Unitree's humanoid robots have been exhibited in Germany and obtained EU radio equipment cybersecurity certification; United Imaging has obtained EU CE certification for a cumulative 77 products, with more than 35 AI-enabled medical devices; XPeng has built a data centre in Europe and trains models locally to meet GDPR requirements; and Zebra Intelligent's smart cockpit has passed a series of certifications including GDPR, R155, R156 and R171. Each company is answering the same question in its own way: when a product goes abroad with a model, where exactly do the legal obligations fall?⑦⑥⑤④

III. What Problems This Article Can Help You Solve

For manufacturing enterprises in the B2B supply chain, especially manufacturers of IoT devices such as robots, industrial equipment and monitoring facilities, there are often real overseas orders at hand but no dedicated export control and data compliance team; what is most needed is an explanation that can be operated item by item against a product list.

After reading this article, you will get four things. A set of export classification self-check methods that can be done against a product list, a set of compliance judgment standards for model delivery and updates, a batch of clauses that can be written directly into contracts, and implementation recommendations for the three forms of offline, overseas inference and domestic inference.

The author believes that the essence of the going-global compliance problem is that enterprises need to make choices in advance at two levels: product architecture and transaction structure. Waiting until the product is in mass production and orders are in transit before remedying compliance will cause costs and prices to rise exponentially.

02. Think Through Three Things First: Model Form, Processing Location, Data Destination

The same "equipment with a model going global" corresponds to completely different legal obligations under different technical arrangements. This chapter breaks the scenario into three questions, then adds the layer of data nature, to establish a comparison table used throughout the article.

I. First Thing: What Form Is the Model In

The first question is in what form the model goes abroad with the product; there are two.

An offline model means the model weights and algorithms are fixed locally on the device, and inference is also completed locally on the device, without relying on an external network. In this form, the model completes its cross-border movement at the point of export together with the product, and no further model transfer occurs during subsequent use of the equipment. Legal review is concentrated at the export stage.

An online model means the model's capabilities depend partly or entirely on external computing power, and can be further divided into device-local inference and cloud inference. Device-local inference means the model is deployed on the device but requires regular networked updates; cloud inference means inference occurs on a server or cloud platform, and the device is merely a terminal. The legal obligations of an online model are continuous, and the export stage is only the beginning.

II. Second Thing: Where the Model and Data Are Processed

The second question is where inference and data processing occur; there are three possibilities: locally on overseas devices, on overseas servers, and on domestic servers.

When inference occurs locally on an overseas device, the data does not leave the country, and both EU-side data processing obligations and China-side data export regulation may not be triggered, but whether the model itself constitutes a technology export remains a question. When inference occurs on an overseas server, the data is processed overseas; on the China side, one must look at the determination of "providing controlled items to foreign organizations and individuals", and on the EU side, one must look at the application of the GDPR and the obligations of the Artificial Intelligence Act. When inference occurs on a domestic server, the data must be transmitted back to China, and China's data export assessment, standard contract and certification mechanisms come into play one by one; on the EU side, there are also restrictions on transfers of data to third countries.

XPeng's establishment of a data centre in Europe and completion of data storage and model training locally is precisely a real-world example of the "overseas localization" route.⑥

III. Third Thing: Where the Data Ultimately Goes

The third question is the ultimate destination of equipment operational data; there are three possibilities: local aggregation overseas, transmission back to China, and flow to a third country.

The aggregation path determines whether China's data export rules are triggered. If the data remains overseas, whether the equipment vendor is still a data processor under Chinese law depends on the degree of the vendor's control over the data and the remote operation and maintenance model. If the data is transmitted back to China, the assessment, standard contract or certification under the Measures for Security Assessment of Data Export are triggered. If the data flows to a third country, both China-side export rules and EU-side retransfer restrictions must be assessed.

Data nature is the fourth dimension superimposed on the above. Industrial operational data, personal information (such as operator images or physiological data), important data and core data carry completely different levels of obligation. The determination of important data directly affects whether the security assessment route must be taken; this determination rule is currently refined only in some industries, with the automotive industry leading.⑧

IV. Three Unavoidable Additional Questions

In addition to the three questions above, there are three unavoidable additional issues that will change the compliance judgment and are repeatedly addressed below.

The first is OTA updates. Models iterate after the equipment leaves the factory, and whether each pushed update constitutes a new technology export or data export is highly disputed in practice and needs separate discussion (see Chapter 4, Section 3).

The second is open-source weights. The model embedded in the equipment may be an open-source weight model, such as the DeepSeek or Qwen series. Open source does not mean uncontrolled; the redistribution clauses of the Licence and export control rules will superimpose.

The third is retraining. If aggregated data is used for model retraining, the issue of training data export is activated, which is a separate set of obligations from the pure inference scenario.

V. Scenario Comparison Table

Combining the three questions yields a set of scenario combinations, each corresponding to a set of main compliance triggers. To avoid an overly wide table, the table below is arranged by model form in rows, and lists the main legal issues by inference location and aggregation path in columns.

Going-Global Scenario Comparison Table

This comparison table is the skeleton of the whole article. Chapter 3 first answers the legal framework, Chapter 4 addresses the most uncertain question of "how the model goes abroad", and Chapter 5 provides solutions branch by branch according to the key branches of the comparison table.

I. The "Dual-Track" Framework of China's Export Control

China's regulation of cross-border flows of AI technology consists of two sets of rules that run in parallel but do not fully overlap. Understanding the "dual track" is the first step for equipment manufacturers in making export judgments.

The first track is technology export licensing management, based on the Foreign Trade Law, the Regulations on Administration of Technology Import and Export, and the Catalogue of Technologies Prohibited or Restricted from Export in China. Article 18 of the 2025 revised Foreign Trade Law provides that the state may prohibit or restrict the import and export of relevant goods and technologies on the basis of twelve categories of circumstances including safeguarding national security; Article 19 provides for national security exceptions; Article 20 provides for the catalogue formulation and temporary restriction mechanism, under which the Ministry of Commerce, with State Council approval, may temporarily decide outside the catalogue to prohibit or restrict the import and export of specific goods and technologies. Article 28 of the Regulations on Administration of Technology Import and Export further clarifies that the catalogue of technologies prohibited or restricted from export shall be formulated and published by the competent foreign economic and trade department of the State Council together with relevant departments.⑩⑨

The current Catalogue of Technologies Prohibited or Restricted from Export in China was adjusted and issued by Announcement No. 28 of 2025 of the Ministry of Commerce and the Ministry of Science and Technology, and has been in force since 15 July 2025. The restricted-export part of the catalogue retains technology entries directly related to AI, including "artificial intelligence interactive interface technology" and "personalised information push service technology based on data analysis", which were added by Announcement No. 38 of 2020. It should be noted that the 2020 adjustment attracted wide attention because of the ByteDance sale of TikTok case, but the catalogue is not limited to recommendation algorithms: any technology falling within a restricted-export entry requires a licence before it is provided abroad.⑬⑫⑪

The second track is export control administration, based on the Export Control Law and its supporting regime. Article 2 of the Export Control Law defines controlled items as dual-use items, military items, nuclear items and other goods, technologies and services related to safeguarding national security and interests and to fulfilling international obligations such as non-proliferation, and expressly provides that "controlled items include data such as technical materials related to the items", while bringing within the scope of export control both "the transfer of controlled items from within the territory of the People's Republic of China to outside the territory, and the provision of controlled items by citizens, legal persons and unincorporated organisations of the People's Republic of China to foreign organisations and individuals". This article is crucial to the characterisation of model weights going overseas, and Chapter IV will elaborate. Article 12 provides for the export licensing system and, for items not on the lists, sets a catch-all obligation to apply for a licence where the exporter "knows or should know" that the items may endanger national security or be used for weapons of mass destruction or terrorist purposes. Article 34 provides for administrative penalties and liability for unauthorised export, and Article 48 provides for reciprocal measures against countries or regions that abuse export control measures.⑰⑯⑮⑭

The Regulations on Export Control of Dual-Use Items, in force since 1 December 2024, are the core supporting legislation to the Export Control Law. Article 2 follows the definition of dual-use items, Article 10 provides for risk assessment of export destinations and the adoption of corresponding control measures, Article 24 provides for the system of end-user and end-use certification documents, and Article 49 provides for extraterritorial application, that is, where an overseas organisation or individual outside the territory transfers or provides to a specific destination country or region, or to a specific organisation or individual, specific dual-use items originating in China, dual-use items containing items originating in China, or dual-use items manufactured using technology originating in China, the Ministry of Commerce may require the relevant operators to comply with the Regulations by reference. On 15 November 2024, the Ministry of Commerce, the Ministry of Industry and Information Technology, the General Administration of Customs and the National Cryptography Administration issued the Dual-Use Items Export Control List, which took effect on 1 December 2024 together with the Regulations and adopts the internationally prevailing item coding system.⑳⑲⑱

The "product-level" control precedent is equally important. Announcement No. 31 of 2024 of the Ministry of Commerce, the General Administration of Customs and the Equipment Development Department of the Central Military Commission imposes export control on specific unmanned aerial vehicles, covering infrared imaging equipment, synthetic aperture radar, lasers for target designation, inertial measurement equipment, radio communication equipment and counter-UAV systems that meet the technical indicators, and includes a catch-all clause prohibiting the export of UAVs not on the list where the exporter knows or should know that they will be used for the proliferation of weapons of mass destruction, terrorist activities or military purposes. The announcement took effect on 1 September 2024 and was later incorporated into the Dual-Use Items Export Control List. On 5 August 2026, the Ministry of Commerce issued Announcement No. 34 of 2026, strengthening export control on UAV-related dual-use items to the United States.㉒㉑

Comparison table of China's two export control tracks

The author considers that the most common compliance blind spot for equipment manufacturers is checking only the technology catalogue and not the dual-use items list. The two tracks run in parallel, and the export of a single device with AI capabilities may trigger both sets of obligations at once. The UAV announcement signals something more important: China will control AI capabilities by reference to "product plus payload technical parameters", rather than remaining only at the level of the "technology" catalogue.

II. Cross-border data and AI governance

Article 25 of the Data Security Law provides that the State exercises export control in accordance with the law over data that is a controlled item and that relates to safeguarding national security and interests and to fulfilling international obligations. This is the interface between the Data Security Law and the Export Control Law. The cross-border flow of device operating data may therefore fall within two frameworks at the same time: outbound data transfer management on the one hand, and export control of controlled-item data on the other.㉓

The backbone of outbound data transfer management is the three paths established by the Measures for Security Assessment of Outbound Data Transfer. The transfer abroad of important data, and transfers by operators of critical information infrastructure or by those processing personal information reaching the quantitative threshold, require filing for a security assessment of outbound data transfer; the outbound transfer of personal information that does not meet the assessment conditions may proceed via standard contract or certification. In June 2025 the Cyberspace Administration of China issued the Guidelines for Filing Security Assessment of Outbound Data Transfer (Third Edition), which connect with sectoral guidelines such as those for the automotive industry.㉕㉔

The automotive industry is currently the most detailed sample of the rules. The Several Provisions on the Administration of Automotive Data Security (Trial) require important data in principle to be stored domestically, and where it is genuinely necessary to provide it abroad, a security assessment must be passed. On 30 January 2026, the Ministry of Industry and Information Technology and seven other departments issued the Guidelines for Security of Outbound Automotive Data Transfer (2026 Edition), refining the rules for determining important data in scenarios such as R&D and design, production and manufacturing, driving automation, software upgrades and connected operation, and setting out nine categories of exemptions, among which important data relating to security vulnerabilities, security incidents and the source code of OTA upgrade software packages is exempt from filing for security assessment, and making clear that automotive data processors shall file through a domestic legal-person entity.㉘㉗㉖

This industry sample has a dual significance for the scenarios in this article. On the one hand, it shows that the regulator is willing to strike a balance between "facilitation of outbound data transfer" and "national security", providing clear channels such as the nine categories of exemptions; on the other hand, the exemptions currently apply only to the automotive industry, and equipment manufacturers in fields such as robotics, medical devices and surveillance equipment cannot copy them directly, but can only refer to their logic and adopt a conservative approach where the rules have not been refined.

On the AI governance side there is a further layer of filing and content security requirements. China administers generative artificial intelligence services through filing, and algorithm recommendation services through algorithm filing and security assessment; from September 2025, providing generative AI services to the domestic public also requires compliance with the requirements for labelling generated and synthesised content. Whether a model built into a device falls within the filing scope depends on whether the service is provided to the domestic public. Purely offline devices, and scenarios in which no generative service is provided to the domestic public, usually do not trigger filing; once model capabilities are opened to domestic users in the form of an online service, the filing obligation may arise. This layer is independent of export control, and enterprises easily overlook it.㉙

III. The EU side: target market rules

The EU is currently the market with the most systematic AI regulation, and is the target market assumed in this article. Chinese equipment manufacturers need to face four sets of rules at the same time.

The first set is the Artificial Intelligence Act (Regulation (EU) 2024/1689). The Regulation entered into force on 1 August 2024 and regulates by risk level. The prohibitions on prohibited AI practices and the AI literacy obligations have applied since 2 February 2025, and the obligations for general-purpose AI models have applied since 2 August 2025. The obligations for high-risk AI systems listed in Annex III were originally to apply from 2 August 2026, and those for high-risk AI systems listed in Annex I that are embedded in regulated products were originally to apply from 2 August 2027.㉚

The 2026 Digital Omnibus (Regulation (EU) 2026/1744) adjusted the above timetable. The compliance deadline for standalone high-risk systems in Annex III was postponed to 2 December 2027, and that for embedded high-risk systems in Annex I to 2 August 2028. Machinery was removed from Annex I and is instead subject to the new Machinery Regulation as the single regulatory entry point, with the European Commission retaining the power to add high-risk AI safety requirements by delegated legislation. For equipment manufacturers this means that 2026 to 2028 is a real preparation window; the window is not a relaxation but a countdown.㉜㉛

The second set is the Machinery Regulation (Regulation (EU) 2023/1230). The new Machinery Regulation, for the first time, brings AI machinery systems that "use machine learning and have self-evolving behaviour" within the high-risk scope requiring assessment by a notified body, with a direct impact on robotics, intelligent manufacturing and human-machine collaboration industries. The trouble is that its definitions do not align with those of the Artificial Intelligence Act, so the same AI system may lead to different high-risk determinations under the two regulations, and manufacturers can only assess in parallel.㉝

The third set is the GDPR and sectoral rules. The GDPR treats vehicle operating data, driver status data, street-view images and the like as personal data, whose transfer to third countries requires a legal basis. Under the United Nations Economic Commission for Europe framework, R155 (cybersecurity), R156 (software updates) and R171 (driver assistance systems) impose certification requirements on intelligent connected products. The cybersecurity requirements of the EU Radio Equipment Directive (EN 18031) have applied mandatorily since 1 August 2025 to radio equipment exported to the EU. The new EU Product Liability Directive, in force from December 2026, brings AI systems and software within the scope of "product" for the first time, introduces mandatory disclosure of evidence and presumptions of defect, and significantly increases the civil liability risk of AI medical devices and robotics products.㉟㉞

The fourth set is the rules on roles and responsibilities. The Artificial Intelligence Act imposes obligations separately on providers, importers and deployers. Chinese manufacturers entering the EU under their own brands generally bear provider obligations, which are the heaviest; where entry is through a European distributor, the importer is obliged to verify product compliance and the completeness of documentation. There is also an easily overlooked "substantial modification" clause: where a customer makes modifications to the AI system of a product already placed on the market beyond the design scope, the customer may change from deployer to provider and the original certification becomes invalid. The root of this risk often lies in the manufacturer's original product architecture and contract terms.㉝

Comparison table of the four sets of EU-side rules

IV. Reference to US law

US rules do not directly bind Chinese equipment manufacturers exporting to the EU, but because of supply chain realities they cannot be avoided. Here two situations need to be distinguished first.

The first situation is export to the United States. The scenario in this article is set in the EU, so export to the United States is not elaborated, but one reminder is in order: the export of Chinese AI products to the United States faces multiple constraints at once from US export controls, the entity list and investment screening, which is a separate topic.

The second scenario is more common: products exported to the EU or other third countries use US technology or chips, and the question is whether the products are subject to US jurisdiction. The US Export Administration Regulations (EAR) contain two rules that may bring foreign-made products within their scope. One is the foreign direct product rule: products manufactured outside the US using US-origin technology or software may still be subject to US controls because of the source of the technology. The other is the de minimis rule: foreign-made products may be subject to controls when the US-origin content reaches a certain percentage. Put simply, even if an equipment maker ships only to the EU, it must assess US rules whenever the equipment's design software, chips or other components use US-origin technology. The practical approach is to require upstream suppliers to provide item origin and compliance commitments, and to include supply chain compliance clauses in procurement contracts.

In addition, US controls on advanced computing exports to China continue to tighten. In April 2025, Nvidia's H20 chip was brought under export controls targeting China; at the end of May 2026, BIS made clear that overseas subsidiaries of Chinese companies also need a licence to procure advanced AI chips; at the end of 2025, Chinese AI companies such as MiniMax and Zhipu AI were added to the Entity List. This directly determines equipment makers' choice of computing sources, and means that chips and software containing US content must first pass a US supply chain review before entering equipment. ㊳㊲㊱

V. Extraterritorial application: can Chinese rules reach outside China?

Once equipment goes overseas, with models running and data processed abroad, the question B2B clients ask most often is whether Chinese rules can still apply.

The answer is layered. Article 49 of the Regulations on Export Control of Dual-Use Items has already extended controls to onward transfers abroad: MOFCOM may require overseas operators to comply with the Regulations by reference in respect of specific dual-use items originating in China, items containing Chinese-origin items, or items manufactured using Chinese-origin technology. In the rare earths field, the "minimum content tracing rule" was already put into practice in 2025: foreign-made finished products in which Chinese-origin controlled items account for 0.1% or more by value are subject to control. Although the relevant announcement is currently suspended until November 2026, the institutional groundwork has been laid. By analogy to AI, foreign-derived models containing Chinese training data or Chinese model technology content could in theory be subject to retroactive control. ㊴⑲

Article 25 of the Data Security Law provides another link on the data side. If data processed abroad itself constitutes "data that is a controlled item", the processing and aggregation of that data abroad may also fall within the scope of China's export control. Taken together, these two lines mean the assumption that "data kept abroad need not be managed" does not hold; companies must assess case by case according to "the nature of the data and the source of the model technology".

04. Core issue: the compliance boundary of model weights and model delivery

The previous two chapters set out the framework; this chapter answers the question equipment makers care about most: how exactly does the model itself go abroad? The question is complex because current rules contain no provision specifically addressing "model weights"; the conclusion must be derived by interpreting existing regimes.

I. Characterisation: do model weights constitute a "technology export"?

Article 2 of the Export Control Law provides two key handles. First, controlled items include "technical materials and other data related to items"; second, export includes "providing controlled items to foreign organisations and individuals". Model weights are the set of mathematical parameters of a model and the carrier of its capabilities, and it is not difficult to interpret them as falling within the range of "technical materials and other data". More important is the second point: bringing "providing" within the definition of export means an export need not take the form of physical delivery; opening downloads to overseas parties, licensing use, or deploying inference services may all constitute provision. ⑭

The logic under the Regulations on the Administration of Technology Import and Export is similar. In practice, regulators have shifted from "physical exit of the technology carrier" to "whether the technological capability has undergone a substantive cross-border transfer". The yardstick is therefore clear: whether the technological capability has been substantively transferred across the border, not whether source code has been handed over. The authors caution that when assessing any overseas arrangement, three questions should be answered first: who actually obtains it, who continuously uses it, and who controls the technological capability. Practical observations in the legal profession in recent years have identified three high-risk scenarios. First, wholesale relocation of the core team: the domestic R&D team moves abroad, and core algorithms, training frameworks or technical materials are actually obtained, continuously used or controlled by an overseas entity. Second, cross-border licensing and model deployment: licensing domestically trained foundation model weights or algorithm modules to an overseas entity, or deploying inference engines involving export-restricted technology on overseas servers. Third, look-through merger review: an overseas entity acquires a domestic AI company or core technology assets, and the transaction in substance results in the transfer of restricted technology abroad. ㊵

The Manus case in April 2026 sent a signal in another direction. The Office of the Working Mechanism for Security Review of Foreign Investment (National Development and Reform Commission) issued a decision prohibiting foreign investment in the Manus project and requiring the transaction to be unwound. Media commentary interpreted the case as a look-through finding of the logic that "where the technology source is not lawfully severed, a technology export is constituted". It should be noted that the Manus case proceeded through the foreign investment security review route, not the export control route, but both point to the same judgment: whether technological capability has substantively gone abroad depends on the substance of the transaction, not its formal arrangements. ㊶

II. Compliance paths for three delivery forms

In practice, models going overseas with products take three forms, with completely different obligations.

The first is offline embedding: model weights and algorithms are embedded in the equipment and exported with the product. The review point is concentrated at the export stage. The equipment maker must complete three steps. Step one is a self-check of items and technology, comparing item by item against the restricted export entries in the Catalogue of Technologies Prohibited or Restricted from Export and the Dual-Use Items Export Control List, with particular attention to components in the equipment that may be separately controlled, such as infrared imaging and radar payloads with specific parameters. Step two is determining whether a licence is required: where a restricted technology export entry is met, apply to the provincial commerce authority for a technology export licence; where the dual-use items list is met, apply to MOFCOM for a dual-use item export licence. Step three is preparing end-user and end-use certificates, submitting them as required by Article 24 of the Regulations on Export Control of Dual-Use Items and retaining the relevant materials for at least five years as required by Article 27. ⑱

The second is an online API: the model stays in China and exports inference capability to overseas equipment, exporting capability only, not weights. This form sidesteps the "weights going abroad" issue, but the risk of a finding of "providing controlled items to foreign organisations and individuals" then arises. If the model technology itself falls within a restricted export entry, opening a calling interface to overseas parties may constitute a technology export; if an equipment maker, to circumvent controls, deploys a domestically trained model on an overseas server and then provides services externally, it may instead trigger a finding of "substantive exit". In compliance design, a domestically trained, overseas inference architecture is not itself off-limits, but it must be ensured that the overseas deployment entity, the source of computing power and the technology flow path are clear, and that use restrictions and data-not-leaving-the-country clauses are written into the contract.

The third is open-source weights. When equipment embeds open-source weight models such as the DeepSeek and Qwen series, licence redistribution and export control overlap. Open-source licences address authorisation at the copyright level and do not resolve export control issues; export control is a public law obligation and cannot be exempted by licence. An equipment maker embedding an open-source model in a product and re-exporting it must check the licence's redistribution clauses, modification and attribution requirements, and at the same time assess whether the source of the model technology involves restricted export technology. The tiered export control plan reported in the media in July 2026, if implemented, could directly affect the overseas distribution of open-source weights, because the scope of discussion explicitly includes open-source weight models. For such uncertainties, companies should prepare contingency plans for policy shifts and avoid locking themselves into a single open-source dependency. ③

III. OTA updates: a continuing obligation or a one-off obligation?

Model updates after equipment goes overseas are the most disputed issue in practice. Each OTA push involves cross-border transmission of model code and weights; whether each one constitutes a new technology export or data export directly affects companies' operating costs and licensing arrangements.

Current rules do not answer this directly. The closest industry answer is the Guidelines for Data Export Security in the Automotive Sector (2026 Edition), which bring relevant important data concerning security vulnerabilities, security incidents and OTA upgrade software package source code within the situations exempt from filing a data export security assessment, on the ground that this provides companies with a convenient and fast data export channel to promptly fix security hazards. This exemption shows that regulators recognise the efficiency logic that "an update is not the same as a full export process", but its scope is limited to the automotive sector, and it exempts only the data export assessment filing, not technology export licensing. ㉘

The authors suggest that update arrangements should be designed as a package. First, when first applying for an export licence, declare the subsequent update mechanism together, including update frequency, the boundaries of update content and the overseas deployment entity, and seek to have the update arrangements specified in the licence. Second, reserve update rights in the contract, making clear that the equipment maker has the right to push security and functional updates, to avoid disputes with customers over rights. Third, establish a tiered review mechanism for update content, with security patches, model parameter fine-tuning and core algorithm changes each going through a different review process, so that only core changes trigger a full licence assessment. Fourth, retain logs and approval records of each update as compliance evidence.

Flowchart for determining OTA model update obligations

IV. Transaction structure design

Compliance issues in model delivery must ultimately be resolved at the level of transaction structure.

The first level is entity arrangements. Whether the domestic manufacturer exports directly or through an overseas subsidiary or distributor determines where Chinese export control obligations attach and how the EU supply chain role is characterised. Where a domestic entity exports directly, the applicant for the technology export licence and the dual-use item licence is clear; where export is through an overseas subsidiary, it must be examined whether the technological capability has undergone a "substantive exit", to avoid transferring domestic R&D results wholesale overseas and then selling externally.

The second level is contract terms. Equipment sales contracts and technology licensing agreements should be designed as a bundle, covering at least the following clauses.

Checklist of compliance contract clauses for equipment going overseas

The third level is product architecture. One point easily overlooked in practice is that the risk of substantive modification must be addressed by both product architecture and contract terms. At the technical level, version control, parameter locking and update authentication should fix the boundary between permitted operations and modifications that trigger re-certification; at the contract level, specific examples should define which operations constitute substantive modification. For example, a customer adjusting settings within the parameter range specified by the manufacturer is a permitted operation; a customer retraining the model with its own data very likely constitutes a substantive modification, whereby the customer becomes the provider and the original certification lapses. ㉝

The fourth layer is the separation of "safe AI and efficiency AI". Separating AI that performs safety functions (such as collision avoidance and autonomous navigation) from efficiency-tool AI (such as production scheduling and appearance inspection) at the system level can narrow the scope of high-risk determinations while reducing compliance costs. Classify and make a judgment for each AI function individually and keep records; treat ambiguous boundaries as high-risk, and conduct pre-assessment with notified bodies as early as possible.㉝

05. Scenario-based compliance analysis

This chapter returns to the comparison table in Chapter 2, expands branch by branch along the three factual branches, and provides compliance paths that can be directly compared.

I. Branch A: Offline model devices

The scenario is characterized by model weights being fixed within the device, with the device operating locally in the EU and not relying on external computing power. This is the form with the lowest compliance cost and the highest certainty, but it likewise carries two sets of obligations.

The export stage is the first set of obligations. Before export, equipment manufacturers complete an item self-check to confirm whether model technology and equipment components fall within the catalogue of technologies restricted from export or the dual-use items list; where a license is required, apply in advance, and prepare end-user and end-use certificates. The key point of the self-check is "parameter comparison"; for example, drone export controls set clear indicators for infrared imaging equipment such as wavelength range and instantaneous field of view, and imaging, radar, and laser components in industrial equipment must be checked item by item.㉑

EU market access is the second set of obligations. Offline devices entering the EU must satisfy CE marking and the safety requirements of the Machinery Regulation; devices with radio functions must also satisfy the EN 18031 cybersecurity requirements. Where AI serves as a safety function, it may be subject to notified body assessment under the new Machinery Regulation. Unitree Robotics obtained the TÜV Rheinland EN 18031 certificate, showing that Chinese robotics companies are already following this set of rules.④㉞

The remaining problem for the offline form is updates. Pushing model updates via OTA after the device leaves the factory will gradually turn "offline" into "intermittently online". For the compliance arrangements for updates, see Chapter 4, Section 3, and they will not be repeated here.

II. Branch B: Online models, overseas inference

The scenario is characterized by model capabilities relying on external computing power, with inference servers deployed within the EU, possibly in a data center built by the equipment manufacturer itself or in a third-party cloud. Data is processed within the EU, while data aggregation is mainly overseas.

The first question the China side must answer is whether cross-border deployment of model capabilities constitutes "providing controlled items to foreign organizations and individuals". If the model technology falls within the catalogue of technologies restricted from export, deploying inference nodes overseas and providing model weights or algorithm modules to overseas entities may constitute technology export. If an equipment manufacturer deploys a model trained domestically to an overseas server and then provides services externally, the technical capability has substantively left the country, and the focus of review shifts from "export licensing" to "deployment structure and data pathways".㊵

The second question the China side must answer is whether Chinese law can still govern when data is processed overseas. The answer depends on two layers. First, the identity of the data processor: if an equipment manufacturer continuously accesses overseas data through remote operations and maintenance, it may be deemed a data processor under Chinese law, and its data processing activities must comply with the Data Security Law. Second, the nature of the data: if the data constitutes "data of controlled items", the export control provision in Article 25 of the Data Security Law may apply by extension.㉓

The obligations on the EU side are concentrated in three areas. First, the GDPR: the equipment manufacturer or its EU entity may be a data controller and must satisfy requirements such as a lawful basis, data minimization, and restrictions on transfers to third countries. Second, the AI Act and the Machinery Regulation: determine according to product form whether it falls within high-risk, and assume obligations for risk management, data governance, technical documentation, logging, transparency, and human oversight. Third, industry certification: R155, R156, R171, etc. apply according to product type.

XPeng's establishment of a data center in Europe, local storage, and local model training is a complete sample of this route. Its approach shows that overseas localization is not only a compliance need but also a commercial arrangement that meets GDPR and EU regulatory expectations.⑥

III. Branch C: Online models, domestic inference

The scenario is characterized by inference servers located within China, with data generated in the EU transmitted back to China for processing. This is the form with the greatest data compliance pressure and the most obvious reverse pull between Chinese and EU rules.

On the China side, data transmission back constitutes a data flow in the "outbound" direction, and it must be determined under the Measures for Security Assessment of Data Export whether a security assessment, standard contract, or certification is triggered. Equipment manufacturers need to complete data classification and grading, identify industrial operation data, personal information, important data, and core data, and then choose a path according to category and volume thresholds. The exemption situations in industry guidelines must be checked item by item; the OTA exemption in the automotive industry cannot currently be directly applied by analogy to other industries.㉗

On the EU side, data transfers to China must satisfy the transfer rules in Chapter V of the GDPR, requiring an adequacy decision, standard contractual clauses, or other appropriate safeguards, and a data transfer impact assessment must be completed. The current status of the EU's determination of China's level of data protection directly affects the feasibility of the transmission-back path.

Putting the two lines together gives the "localization pendulum". Chinese rules encourage domestic storage of important data and require assessment for outbound transfer, while EU rules require personal data to remain in Europe or be subject to adequate safeguards. Equipment manufacturers make architectural choices in between, and there are three common options. First, full localization: inference and data both remain in the EU, with only a model update channel retained on the China side, suitable for scenarios with high data sensitivity. Second, layered processing: raw data remains in the EU, while desensitized or aggregated data is transmitted back, suitable for training and operations needs. Third, domestic inference plus compliant transfer: data is transmitted back to China, going through full outbound assessment and GDPR transfer safeguards, suitable for scenarios where model iteration is strongly dependent and local computing power is insufficient.

The AIVA case illustrates this problem from the opposite direction. This AI-native automotive company relies heavily on real-time interaction with large models, and analysis points out that its overseas expansion must first solve localized data centers and GDPR compliance; otherwise its core selling point will be blocked by data regulation.㊷

IV. Comparison of the three branches and selection recommendations

Compliance paths for the three factual branches

Comparison of compliance paths for the three branches

The author's view is that architectural choices should be completed at the product definition stage. For most small and medium-sized equipment manufacturers, the offline approach of Branch A or the overseas localization of Branch B is a more stable starting point; the domestic inference of Branch C is suitable for enterprises whose core competitiveness is model iteration, but the costs of outbound data assessment and GDPR transfer mechanisms must be calculated into product pricing in advance.

06. Compliance toolbox

This chapter distills the judgments above into tools that can be used directly, including self-check lists, decision trees, document systems, organizational processes, and key contract terms.

I. Export characterization self-check list

Before launching any overseas project, equipment manufacturers should conduct a round of export characterization according to the following steps.

Decision tree for equipment overseas compliance

The key points of the self-check list are as follows. First, technology identification: list item by item the models, algorithms, software, and key components in the device. Second, parameter comparison: compare against the item codes and technical indicators in the dual-use items list, with key checks on drone, imaging, and radar components. Third, catch-all assessment: determine under Article 12 of the Export Control Law whether there are national security risks in the end user or end use. Fourth, data classification: identify personal information, important data, and core data in outbound data. Fifth, full-process record-keeping: archive the basis for the self-check, conclusions, and approval records for no less than five years.

II. Data classification and selection of outbound paths

Data outbound paths are determined in the following order. Step one: identify whether outbound data behavior exists, including transmission back, overseas access, and provision to third parties. Step two: data classification and grading, distinguishing industrial operation data, personal information, important data, and core data. Step three: check industry exemptions item by item; currently the nine categories of exemptions in the automotive industry are clear, and other industries may refer to their logic but must not apply them directly. Step four: determine security assessment, standard contract, or certification according to path selection standards; file for security assessment where the threshold for outbound important data or personal information volume is met, and use standard contracts or certification for the rest. Step five: simultaneously complete the EU-side GDPR transfer impact assessment and appropriate safeguards.

III. Recommendations on product architecture and document systems

At the product architecture level, separate safety-function AI from efficiency-tool AI, and fix compliance boundaries through version control, parameter locking, and update identity verification. At the document system level, integrate the technical documentation under the Machinery Regulation and the AI safety technical documentation into a unified architecture maintained by the same team, avoiding conflicts between two sets of documents. The core document list includes training data descriptions, risk assessment and testing reports, logging mechanism design, transparency manuals, and records of human takeover paths.

IV. Organization and processes

Compliance responsibility must be assigned within the organization. It is recommended that equipment manufacturers establish a cross-departmental compliance working group covering product, R&D, legal, and marketing, translating regulatory language into product design indicators, test cases, and document templates. For internal compliance system development, refer to the framework of the Ministry of Commerce's Guidelines for Internal Compliance of Dual-Use Items Export Control. Policy tracking should be institutionalized, focusing on three things: the draft tiered export control rules for Chinese AI models, the follow-up implementation of the EU Digital Omnibus and progress on harmonized standards, and adjustments to US advanced computing rules.

V. Key points of commonly used contract clauses

Equipment sales contracts, technology license agreements, and data processing agreements should be signed as a package. The seven categories of clauses listed in Table 1 are the bottom line; in practice, it is also necessary to add obligations to cooperate with data protection impact assessments, audit rights, joint guarantees of export control compliance commitments, and price adjustment and exit mechanisms upon policy changes.

Conclusion

Returning to the core question at the beginning of this article: Are Chinese AI products subject to export control? The accurate statement is that some are already controlled, the main body remains to be observed, and the direction is clear. The algorithm and technology layer already falls within the Catalogue of Technologies Prohibited or Restricted from Export, and the upstream computing power and materials layer already falls within the Dual-Use Items Export Control List, while the large model and model weight layer has no dedicated rules in force; once the tiered control plan reported by the media in July 2026 is implemented, it will directly change the overseas expansion landscape for open-source and closed-source models.

For equipment manufacturers, what really needs to be established is a full-cycle compliance mechanism covering product definition, export licensing, data flows, overseas operations, and contractual arrangements; "whether it can be exported" is only the first question in this mechanism. Product export, model delivery, and data aggregation are intertwined; if any one line fails, the entire sales chain may be blocked.

The author believes that the cost of compliance is essentially the cost of market access. The EU has postponed the compliance window for high-risk obligations of high-risk AI systems to the end of 2027 and mid-2028; the significance of this two-year window is to give enterprises time to complete system building. Enterprises that complete compliance construction first during this window will obtain not only certification but also pricing power and customer trust. Chinese equipment manufacturers have globally leading model capabilities, cost advantages, and a complete supply chain; what they lack is only the step of translating compliance capabilities into the language of product design.

This step is worth starting now.

Article appendix

② Source: Official website of the Ministry of Commerce, "Announcement No. 31 of 2024 of the Ministry of Commerce, the General Administration of Customs and the Equipment Development Department of the Central Military Commission on Optimizing and Adjusting Drone Export Control Measures", link: http://www.mofcom.gov.cn/zfxxgk/fdzdgknr/ztfl/blgg/art/2024/art_03b1626fca9a470486fb693e194b9c52.html .

① Source: China Chang'an Net, "ByteDance Issues Another Statement", which quotes the restricted-export entries of the Catalogue, link: http://www.chinapeace.gov.cn/chinapeace/c100007/2020-08/30/content_12389398.shtml .

⑦ Source: China Industry News, "Cai Ming, Chief Product Officer of Banma Intelligent: Closing the Generational Gap in Overseas Smart Cockpit Products Is the Biggest Opportunity for Intelligent Going Global" (June 2026), link: https://www.cinn.cn/hz/2026/06-15/eryvoA7D.html .

⑥ Source: Phoenix Net, "No Maps, Pure Vision, Local Training: How Does XPeng Bring Smart Driving to Europe?" (15 July 2026), link: https://news.ifeng.com/c/8uoYV32iXlJ .

⑤ Source: China Securities Journal, Voluntary Information Disclosure Announcement of Shanghai United Imaging Healthcare Co., Ltd. on Obtaining Medical Device Registration Certificate (21 July 2026), link: http://app.cnstock.com/zzb/zgzqb/html/2026-07/21/nw.D110000zgzqb_20260721_4-B059.htm .

④ Source: NetEase News, "AI Cooperation: China 'Makes a Move', the World 'Joins Hands'" (July 2026), link: https://finance.sina.com.cn/jjxw/2026-07-16/doc-inihyuqe1326843.shtml ; TÜV eShop report on the EN 18031 certificate for Unitree robots (23 December 2025), link: https://www.tuv-eshop.cn/information/details?id=2003707500764381186 .

⑧ Source: Xinhua News Agency, "Eight Departments Issue Document to Promote Efficient, Convenient and Secure Cross-Border Flow of Automotive Data" (4 February 2026), link: http://www.news.cn/tech/20260204/9ca597741e914bc19d5ed992e571c5b2/c.html .

⑩ Source: PKULaw, "Regulations of the People's Republic of China on the Administration of Technology Import and Export (2020 Revision)" (State Council Decree No. 732), Article 28, link: https://pkulaw.com/chl/1385b55aabf953c9bdfb.html .

⑨ Source: PKULaw, "Foreign Trade Law of the People's Republic of China (2025 Revision)" (Presidential Order No. 67), Articles 18, 19 and 20, link: https://pkulaw.com/chl/fd5fc2cd1f8a3b9dbdfb.html .

⑬ Source: According to public reports, between August and September 2020, ByteDance's sale of TikTok-related technology drew attention to the newly added entries in the Catalogue, and the Ministry of Commerce stated that it would handle applications for technology export licences in accordance with laws and regulations. See the Chinese-language report of Sputnik, link: https://sputniknews.cn/20200924/1032193226.html .

⑫ Source: China Chang'an Net, quoting Article 18 "Artificial Intelligence Interactive Interface Technology" and Article 21 "Personalized Information Push Service Technology Based on Data Analysis" of the restricted-export part of the Catalogue, link: http://www.chinapeace.gov.cn/chinapeace/c100007/2020-08/30/content_12389398.shtml .

⑪ Source: PKULaw, "Announcement No. 28 of 2025 of the Ministry of Commerce and the Ministry of Science and Technology on Adjusting and Issuing the Catalogue of Technologies Prohibited or Restricted from Export in China", link: https://pkulaw.com/chl/088541a3d866e112bdfb.html .

⑰ Source: Ibid., Article 48.

⑯ Source: Ibid., Article 34.

⑮ Source: Ibid., Article 12.

⑭ Source: PKULaw, "Export Control Law of the People's Republic of China" (Presidential Order No. 58), Article 2, link: https://pkulaw.com/chl/81dc811566c06dadbdfb.html .

⑳ Source: PKULaw, "Announcement No. 51 of 2024 of the Ministry of Commerce, the Ministry of Industry and Information Technology, the General Administration of Customs and the National Cryptography Administration on Issuing the Consolidated List of Dual-Use Items Subject to Export Control of the People's Republic of China", link: https://pkulaw.com/chl/d85934d33dd0ad08bdfb.html .

⑲ Source: Text and interpretation of Article 49 of the Regulations on Export Control of Dual-Use Items, see King & Wood Mallesons, "2025 Annual Review and Observations on China's Export Control", link: https://www.ctils.com/articles/24311 .

⑱ Source: PKULaw, "Regulations of the People's Republic of China on Export Control of Dual-Use Items" (State Council Decree No. 792), Articles 10 and 24, link: https://pkulaw.com/chl/44c210372db3ac3abdfb.html .

㉒ Source: PKULaw, "Announcement No. 34 of 2026 of the Ministry of Commerce on Publishing the Announcement on Strengthening Export Control of Drone-Related Dual-Use Items to the United States", link: https://pkulaw.com/chl/8cc9788509e0e455bdfb.html .

㉑ Source: Official website of the Ministry of Commerce, full text of Announcement No. 31 of 2024, link: http://www.mofcom.gov.cn/zfxxgk/fdzdgknr/ztfl/blgg/art/2024/art_03b1626fca9a470486fb693e194b9c52.html .

㉓ Source: PKULaw, "Data Security Law of the People's Republic of China" (Presidential Order No. 84), Article 25, link: https://pkulaw.com/chl/1dd3a2fb498ecb3ebdfb.html .

㉕ Source: PKULaw, "The Cyberspace Administration of China Issues the Guidelines for Declaration of Data Export Security Assessment (Third Edition)", link: https://pkulaw.com/chl/0ed57b6ed4f70a42bdfb.html .

㉔ Source: PKULaw, "Measures for Data Export Security Assessment" (Order No. 11 of the Cyberspace Administration of China), link: https://pkulaw.com/chl/d35dcfdd023f5e32bdfb.html .

㉘ Source: Guangzhou Municipal Bureau of Industry and Information Technology, "Eight Questions + One Chart: Understanding the Guidelines for Automotive Data Export Security (2026 Edition)", link: https://gxj.gz.gov.cn/hd/cjwt/zcjd/content/mpost_10684949.html .

㉗ Source: PKULaw, "Notice of the Ministry of Industry and Information Technology and Seven Other Departments on Issuing the Guidelines for Automotive Data Export Security (2026 Edition)" (MIIT Network Security [2026] No. 27), link: https://pkulaw.com/chl/4482cc11f01221d3bdfb.html ; Xinhua News Agency report, link: http://www.news.cn/tech/20260204/9ca597741e914bc19d5ed992e571c5b2/c.html .

㉖ Source: PKULaw, "Several Provisions on the Administration of Automotive Data Security (Trial)" (Order No. 7 of the Cyberspace Administration of China and Four Other Departments), link: https://pkulaw.com/chl/f1c1ecd7c5caa29abdfb.html .

㉙ Source: Chinese Government Website, "Interim Measures for the Administration of Generative Artificial Intelligence Services" (Order No. 15 of the Cyberspace Administration of China and Six Other Departments, effective as of 15 August 2023), link: https://www.gov.cn/zhengce/zhengceku/202307/content_6891752.htm ; related interpretation of the "Measures for Labeling Artificial Intelligence-Generated Synthetic Content" (effective as of 1 September 2025), link: https://gzwxb.gov.cn/context/contextId/214197 .

㉚ Source: Licentium, "EU Council Adopts Digital Omnibus Amendments to AI Act, Deferring High-Risk Compliance to 2027" (30 June 2026), link: https://www.licentium.io/post/eu-council-adopts-digital-omnibus-amendments-ai-act-deferring-high-risk-compliance-2027 ; Dastra, "New Omnibus Agreement: How the EU AI Act changes", link: https://www.dastra.eu/en/blog/simpler-safer-stricter-where-it-counts-inside-the-eu-ai-omnibus-deal/60025 .

㉜ Source: Forkast, "EU's AI Act 'Teeth Pulled' by Digital Omnibus Deferring High-Risk Obligations 16 Months" (28 July 2026), link: https://forkast.news/eus-ai-act-teeth-pulled-by-digital-omnibus-deferring-high-risk-obligations-16-months/ .

㉛ Source: Licentium, "EU Digital Omnibus Regulation 2026/1744 Enters Into Force, Extending AI Act Compliance Timelines" (31 July 2026), link: https://www.licentium.io/post/eu-digital-omnibus-regulation-2026-1744-amends-ai-act .

㉝ Source: DHH Law Firm (Lou He, Xin Xiaotian), "The EU AI Amendment Act Takes Effect, the Dual-Track Regulatory Path Is Reshaped, and a Practical Compliance Guide for Chinese Robotics Companies Going Global", republished by Maofa Tong, link: https://www.ctils.com/articles/26002 .

㉟ Source: China MedGlobal, "The EU's New Product Liability Directive: Litigation Risk for AI Medical Device Manufacturers Soars, and How Chinese Companies Going Global Should Respond", link: https://chinamedglobal.com/tag/SaMD .

㉞ Source: TÜV eShop's explanation of EN 18031 and the cybersecurity requirements of the EU Radio Equipment Directive, link: https://www.tuv-eshop.cn/information/details?id=2003707500764381186 .

㊳ Source: Yicai, "Chinese Companies in the Period of Global Industrial and Supply Chain Reshaping: Capability Enhancement!", link: https://www.yicai.com/news/102982863.html .

㊲ Source: Al Jazeera, "US says ban on AI chip shipments applies to Chinese firms outside China" (1 June 2026), link: https://www.aljazeera.com/economy/2026/6/1/us-says-ban-on-ai-chip-shipments-applies-to-chinese-firms-outside-china .

㊱ Source: Han Kun Law Offices, "AI Companies Going Global (III): Global Export Control Compliance and Response Strategies", link: https://www.ctils.com/articles/25592 ; Akin Gump, "BIS Rescinds AI Diffusion Rule and Issues New Guidance", link: https://www.akingump.com/en/insights/ai-law-and-regulation-tracker/bis-rescinds-ai-diffusion-rule-and-issues-new-guidance .

㊴ Source: King & Wood Mallesons, "2025 China Export Control Annual Review and Observations", link: https://www.ctils.com/articles/24311 .

㊵ Source: Han Kun Law Offices, "AI Companies Going Global (III): Global Export Control Compliance and Response Strategies", link: https://www.ctils.com/articles/25592 .

㊶ Source: news from the official website of the National Development and Reform Commission, republished on the website of the Sichuan Provincial Department of Commerce (9 May 2026), link: https://swt.sc.gov.cn/sccom/c25030603/2026/5/9/e39a0a2b6d4b49bebbecd17efee520a3.shtml ; Cailian Press report (27 April 2026), link: https://news.qq.com/rain/a/20260427A0656400 .

㊷ Source: AIVA-related report (June 2026), link: https://siteapp.news18a.com/m/news/storys_266968.html .

Questions and answers

Do AI model weights embedded in exported equipment count as controlled technology under Chinese law?

Model weights may fall within "technical materials and other data" under Article 2 of the Export Control Law, and providing them to foreign organisations or individuals can constitute export. The key test is whether the technological capability has been substantively transferred across the border, not whether source code was handed over.

Does an OTA model update after equipment is exported trigger a new export licence or data export assessment?

Current rules do not answer directly. The automotive sector exempts OTA upgrade software package source code from data export assessment filing, but this does not exempt technology export licensing. Companies should declare update mechanisms when first applying for a licence and establish tiered review for update content.

If equipment data is kept on servers outside China, do Chinese export control and data rules still apply?

Yes, in some cases. Article 49 of the Regulations on Export Control of Dual-Use Items allows MOFCOM to require overseas operators to comply for items originating in China or made with Chinese technology. Article 25 of the Data Security Law also extends export control to controlled-item data processed abroad.

What EU AI Act deadlines apply to high-risk AI systems in machinery exported from China?

The 2026 Digital Omnibus postponed compliance for standalone Annex III high-risk systems to 2 December 2027 and embedded Annex I systems to 2 August 2028. Machinery was removed from Annex I and is instead subject to the new Machinery Regulation as the single regulatory entry point.

Can Chinese equipment manufacturers rely on the automotive industry's nine categories of data export exemptions?

No. The nine categories of exemptions in the Guidelines for Security of Outbound Automotive Data Transfer (2026 Edition) apply only to the automotive industry. Equipment manufacturers in robotics, medical devices and surveillance cannot copy them directly but may refer to their logic and adopt a conservative approach.

Laws and cases cited

Laws

  • Foreign Trade Law
  • Export Control Law
  • Data Security Law

Regulations

  • Regulations on Administration of Technology Import and Export
  • Regulations on Export Control of Dual-Use Items
  • Measures for Security Assessment of Outbound Data Transfer
  • Several Provisions on the Administration of Automotive Data Security (Trial)
  • Artificial Intelligence Act (Regulation (EU) 2024/1689)
  • Machinery Regulation (Regulation (EU) 2023/1230)

Policy documents

  • Catalogue of Technologies Prohibited or Restricted from Export in China
  • Dual-Use Items Export Control List
  • Announcement No. 31 of 2024
  • Announcement No. 34 of 2026
  • Guidelines for Filing Security Assessment of Outbound Data Transfer (Third Edition)
  • Guidelines for Security of Outbound Automotive Data Transfer (2026 Edition)