A foreign company selling autonomous driving software, sensors or compute into China will not hold the road-test permit, the mapping qualification or the vehicle type approval. Those sit with a Chinese legal person: the manufacturer named on the announcement catalogue, or the platform operator that runs the fleet. This is not a policy obstacle to be overcome. It is the structure of the regime, and it determines where a supplier’s risk actually lives.
The consequence is that the regulatory questions and the contractual questions are the same questions. Who applies for the test? Who files the over-the-air update? Who is the recall obligor when a perception failure causes a crash? Who may receive the sensor data, in what form, and where may it be stored? Each has a regulatory answer that fixes the OEM, and a contractual answer that allocates the cost between the OEM and the supplier.
A supplier that treats China entry as a licensing exercise — find the permit, hire the consultant, launch — will spend two years discovering that its software sits inside a product it does not control, governed by approvals it did not obtain, with liability that its contract did not cap. The work worth doing before the first vehicle moves is the mapping of obligations, and then the contract that allocates them.
The testing and road-access ladder
China did not legislate a single autonomous driving statute. It built a ladder, and the vehicle climbs it one rung at a time: closed test ground, designated public roads, demonstration or pilot operation, and commercial operation. Each rung has its own applicant, its own evidence and its own permit, and progression is not automatic.
The national framework is a set of trial provisions on road testing of intelligent connected vehicles, issued jointly by the industry, public security, transport and market regulation ministries, supplemented by a later access regime for manufacturers and products and by a pilot programme for vehicles at higher automation levels. The provisions set the common requirements: a vehicle that has passed third-party safety testing, a test plan and an approved route, a trained safety officer, insurance or a bond, a data recording device, and a connection to the local monitoring platform. City rules add detail and differ from one another. Beijing, Shanghai, Guangzhou and Shenzhen operate the most developed regimes, and a permit in one city does not travel to another.
The applicant is a Chinese legal person, and in practice it is the vehicle manufacturer or the operating platform, sometimes with the software supplier named as a joint participant. A foreign supplier without a domestic entity cannot apply in its own name, and even with a wholly foreign-owned enterprise the permit attaches to the test subject — the vehicle, the route and the operator — rather than to a software component. Where the supplier wants its own test fleet, it needs an operating entity, a vehicle it owns or leases, and a local safety team.
What the supplier contributes is evidence rather than an application: a safety case for the software, the test results behind the performance claims, the failure modes and the fallback behaviour, and engineers who sit with the OEM’s test team. That contribution should be named in the test plan, because a supplier whose software is in the vehicle but whose name is not in the application has no standing when the test is suspended and no access to the data that would explain why.
The ladder is also an evidence ladder. Each rung requires accumulated mileage, disengagement records and incident reports from the rung below. A supplier that cannot produce its own logs, because the data sits on the OEM’s platform, is dependent on the OEM for the evidence that its product is safe.
Road access, type approval and the product
A vehicle sold in China must appear on the announcement catalogue maintained by the industry ministry and hold compulsory product certification. For a vehicle with a higher-level automated driving function, the access route is separate and narrower: a pilot programme admits a limited number of manufacturer and product combinations, and admission depends on the vehicle, the system and the operator being assessed together.
The practical point for a foreign supplier is that the software version forms part of the approved configuration. A change to the perception stack, the planning logic or the operational design domain may take the vehicle outside what was approved, and a variation is then required before the change reaches customers. Suppliers used to a continuous deployment model, in which the fleet runs the latest build, find this constraint harder to absorb than any single approval requirement.
Three families of standard bear on the software. Functional safety standards govern the development process and the safety case. Standards for the safety of the intended functionality address the performance limitations that remain when the system is functioning as designed. And a set of mandatory standards on vehicle cybersecurity and on software updates, published with phased compliance dates, sets requirements that a supplier’s code must satisfy and that the OEM will test. The supplier should obtain the applicable standard list from the OEM at the start of the programme, because the evidence pack is built during development and cannot be assembled at the end.
None of this makes the supplier the manufacturer of record. The announcement, the certification, the recall and the regulatory correspondence all belong to the Chinese entity. That is efficient for the supplier and dangerous for it at the same time: the OEM controls the filings, and the supplier’s obligations exist only in the contract.
Data: what leaves the vehicle, and what may not
Vehicles generate location traces, camera images, lidar point clouds, cabin audio, driver monitoring video and biometric signals. The Provisions on the Management of Automobile Data Security, issued by the cyberspace administration together with the transport, industry, public security and market regulation departments, impose a structure on that material: classify the data, process it in the vehicle by default, do not collect outside the vehicle unless necessary, anonymise before any export, and treat personal information and important data according to their own rules.
The personal information regime sits underneath. Driver monitoring captures biometric data, which is sensitive personal information and requires separate consent and a necessity assessment. Cabin audio may capture passengers who never agreed to anything. Location traces can identify a person over time even without a name. A supplier that asks for raw clips for model training is asking the OEM to transfer sensitive personal information to a third party, and the OEM will refuse unless the paperwork is in place.
Cross-border transfer is the constraint that shapes architecture. China provides three routes: a security assessment for important data and for volumes of personal information above the applicable threshold, a standard contract filing for smaller volumes, and certification. The automotive sector has its own rules for identifying important data, and the classification of a data set as important data is a matter for the sector regulator rather than for the supplier’s judgement. Where the data is important data, the assessment route applies and the data generally stays in China.
The design consequence is straightforward. Build the training environment in China, on a domestic cloud, with the OEM as the data controller and the supplier as a processor under a data processing agreement. Receive anonymised, aggregated or synthetic data where possible, and treat anonymisation as a technical claim that must be tested rather than a label. Do not pull raw sensor clips offshore for labelling, however convenient the labelling vendor. And treat the vehicle-to-cloud link as a transfer in its own right: a telematics backend outside China is a cross-border data flow, not a remote management convenience.
Mapping and geographic information
Geographic data is regulated separately from personal information, and more strictly. Surveying and mapping is a licensed activity. The collection and processing of geospatial data at survey-grade accuracy requires a qualification, navigation electronic maps are produced by qualified entities, and foreign-invested entities face restrictions on holding the qualification.
High-precision mapping data used for automated driving is treated as sensitive geographic information. A foreign company generally cannot collect or hold it, and must obtain it from a licensed Chinese provider. The sector regulator has issued guidance for intelligent connected vehicles that permits controlled processing of geographic data under supervision, but the guidance does not remove the qualification requirement, and the boundary between permitted vehicle sensor processing and unlicensed surveying is the question that matters.
That boundary is easy to cross without noticing. A vehicle that records lidar returns, camera features and precise poses, and then fuses them into a map of the road environment, is producing geospatial data at an accuracy that ordinary navigation does not require. A fleet doing this at scale is, on its face, conducting a mapping activity. The compliance answer is architectural rather than documentary: consume the licensed provider’s map layer, keep perception outputs in vehicle-relative coordinates, do not persist geo-referenced point clouds beyond the trip, and do not transmit them to a server outside the vehicle without a legal basis.
The commercial arrangements follow. The mapping provider’s licence, the terms on which its data may be used, the ownership of any derived layer and the restrictions on sharing that layer with the OEM’s other suppliers all need to be settled before the programme starts. A supplier that assumes it can use the OEM’s map licence for its own development is usually wrong, and discovers the point when the map provider audits the data flow.
Cybersecurity and the connected platform
A connected vehicle platform is graded under the multi-level protection scheme, and connected vehicle systems are frequently classified as critical information infrastructure, which brings the level three requirements and, in some cases, the additional obligations that attach to critical infrastructure: a dedicated security management body, background checks for personnel in sensitive roles, local storage of personal information and important data, and an annual security assessment.
The mandatory vehicle cybersecurity standard adds a management system requirement: a security management system covering the lifecycle, incident response, vulnerability monitoring and the supply chain. A supplier is inside that system whether or not it has read the standard. The OEM will require a secure development process, a software bill of materials, a vulnerability disclosure channel, patch timelines and penetration test reports, and it will audit.
These obligations should be priced and contracted rather than absorbed. The audit right, the patch deadline and the incident notification window each cost engineering time, and a supplier that accepts an unlimited obligation to patch within a period it cannot meet has taken on a liability it cannot discharge. The same applies to the security assessment of the update package: if the OEM’s release gate rejects a build, the supplier needs a contractual route to resolve the disagreement without stopping the programme.
Product liability, recalls and the allocation of fault
Chinese product liability law is not written for automated driving. The Product Quality Law makes the producer liable for defects, with a right of recourse against a party whose fault caused the defect, and the Civil Code’s tort provisions allocate liability among producers, sellers and third parties. The Road Traffic Safety Law assumes a driver. There is no national statute that says who is responsible when an automated system causes a collision.
Local rules fill part of the gap. Shenzhen’s regulation on intelligent connected vehicles, in force since 2022 and the first of its kind, allocates responsibility between the driver and the manufacturer or operator depending on whether a driver is required for the vehicle’s operation. Other pilot cities have their own measures. None of them binds a court in another province, and none of them resolves the internal question between the OEM and its suppliers.
That internal question is contractual. The OEM will be the defendant, because it is the manufacturer on the announcement and the seller of the vehicle. It will then look to its tier-one supplier, which will look to the software or sensor supplier. The chain works only if each link fixes four things: the definition of a defect in software terms, the standard of proof, the indemnity and its cap, and the insurance that stands behind it.
The definition is where the argument is won. A clause that indemnifies the OEM against “any loss arising from the software” is unlimited in practice, because every collision involves the software. A clause that limits the indemnity to a loss caused by a failure of the software to perform in accordance with the agreed specification, in the operational design domain for which it was validated, excluding modifications made by the OEM and excluding claims arising from the vehicle’s integration, is defensible. The operational design domain, the validation evidence and the change control record become the boundary of liability.
Recalls follow the same structure. The recall obligation sits with the manufacturer under the defective automobile products regime, and the supplier’s duty is to cooperate and to bear cost where the defect is attributable to it. Recall cost can dwarf the value of the supply contract, and it should be capped, insured or shared on a defined basis rather than accepted as an open-ended indemnity.
One further point deserves attention because it is usually ignored at contracting stage. After a serious incident, the vehicle’s recorded data is the only reliable account of what the system did. If that data sits exclusively on the OEM’s platform, the supplier cannot defend itself. A contractual right to access the incident data, in a defined form and within a defined period, is a liability defence and not a data request.
Over-the-air updates
An over-the-air update is not a maintenance convenience in this regime. It is a change to a product that has been approved, and it is regulated on three axes at once.
First, the type approval. If the update changes a characteristic that was certified, the approval must be varied before the change is released. Second, the recall regime. An update that remedies a defect may itself be a recall, and must be filed with the market regulator under the recall rules rather than pushed quietly. Third, the software update rules issued by the industry ministry, which require filing or approval for updates that affect safety, prohibit pushing an update without the user’s knowledge and consent, and require the update to be validated and reversible.
For a foreign supplier the operative question is who holds the release authority. In most programmes the OEM does, because it is the entity that can file the update and the entity that carries the recall obligation. A supplier that retains a remote update channel into production vehicles has created three problems: a cross-border data flow, a cybersecurity exposure, and the possibility of performing a recall without being the recall obligor.
The contract should therefore fix the release process rather than the technology. A change control board with named members from both sides. A joint sign-off before any build that touches the supplier’s component is released. A notification obligation on the OEM for any update to the surrounding system that could affect the supplier’s function. A rollback plan and a definition of what counts as a failed update. And an escrow of the build environment and toolchain, so that the OEM can maintain the software if the supplier exits, without the supplier handing over its source code on day one.
The last item is where negotiations usually stall. A Chinese OEM will ask for source code; the supplier will refuse. The workable middle is a build-and-toolchain escrow with release triggers, a licence limited to maintenance of the delivered version, and a clear statement that the escrow does not transfer ownership of the background IP.
What a foreign supplier must contract for
The regulatory map converts into a list of contractual allocations. Each item below corresponds to an obligation that a regulator imposes on someone, and the contract decides whether that someone is the OEM or the supplier.
Scope and acceptance: the specification, the operational design domain, the validation evidence and the acceptance test, expressed so that the boundary of the supplier’s responsibility is a technical document rather than an adjective.
Regulatory allocation: which party applies for the test, holds the permit, files the update and acts as recall obligor, and what each party must provide to the other to discharge those duties.
Data: who is the controller, what the supplier may receive, in what form, where it is stored, whether it may be transferred offshore at all, and the data processing terms that satisfy the personal information regime.
Mapping: which licensed provider supplies the map layer, who owns the derived data, and what the supplier may not do with geographic data.
Intellectual property: background and foreground, the licence scope, the escrow terms, and the treatment of model weights and training data sets on termination.
Liability: the defect definition, the standard of proof, the indemnity, the cap, the exclusion of consequential loss, and the insurance that stands behind the indemnity.
Recall: cooperation duties, cost allocation, and the supplier’s right to participate in the investigation.
Cybersecurity: the standards to be met, the audit rights, the patch timelines and the incident notification window.
Change control and updates: the release authority, the sign-off process, the notification obligations and the rollback plan.
Exit: transition assistance, the return of data, the deletion of copies, and the survival of the confidentiality and liability provisions.
One item sits outside the Chinese regime and is easily missed. A foreign supplier’s own technology may be subject to export controls in its home jurisdiction, and a Chinese OEM may be a restricted party. The supply contract should address the export classification of the technology, the end-use and end-user undertakings, and the consequences if a licence is refused. That is the supplier’s problem, not the OEM’s, and it is better raised before the design win than after.
Documents, filings and sequence
| Step | Authority | Typical timing | What foreign suppliers miss |
|---|---|---|---|
| Domestic entity, where the supplier needs its own test fleet | Market registration authority | Weeks to months | The permit attaches to a vehicle and an operator, not to a software product |
| Third-party vehicle safety testing | Designated testing body | Months | The evidence pack is built during development, not assembled at the end |
| Road test application and permit | City transport, public security and industry authorities | Months, city by city | A permit in one city does not travel |
| Demonstration or pilot operation permit | City authorities | After accumulated test mileage | The operator, not the supplier, is the applicant |
| Product access and announcement | Industry ministry | Months | The software version is part of the approved configuration |
| Compulsory product certification | Certification body | Months | Applies to the vehicle, and the supplier’s evidence feeds it |
| Data export assessment or standard contract filing | Cyberspace administration | Weeks to months | Raw sensor data for training is the usual casualty |
| Map data agreement | Licensed mapping provider | Before development | The OEM’s licence does not extend to the supplier |
| Multi-level protection grading and assessment | Public security | Before launch, then annually | The connected platform, not the vehicle, is graded |
| Software update filing | Industry ministry and market regulator | Before each safety-relevant release | An update that fixes a defect may be a recall |
| Recall filing, where applicable | Market regulator | On discovery of the defect | The OEM files; the supplier pays |
Where foreign suppliers get this wrong
Treating software as an unregulated product. The software is part of a vehicle that is approved, certified and recalled as a unit, and the supplier’s obligations are created by the OEM’s approval rather than by a licence of its own.
Assuming the OEM’s compliance covers the supplier. The OEM’s permit does not discharge the supplier’s data, mapping or cybersecurity obligations, and the OEM’s contract will place them back on the supplier.
Promising raw data for model training. The commitment is made in the sales cycle, the data cannot lawfully leave, and the training plan collapses after the design win.
Building a high-precision map from fleet data. The activity requires a qualification the supplier does not have, and the map is the evidence.
Accepting an uncapped indemnity for any loss arising from the software. Every collision involves the software, so the clause has no boundary; the operational design domain and the specification are the only defensible limits.
Retaining a remote update channel into production vehicles. It creates a cross-border data flow, a security exposure and the appearance of performing a recall without authority.
Cost and commercial impact
The cost drivers are not the legal fees. They are the test programme, which runs for months per city and per vehicle model; the evidence pack, which requires a functional safety process most startups do not have; the domestic data environment, which means a China cloud, a labelling supply chain and a data processing agreement; and the compliance headcount to maintain the filings after launch.
The commercial impact of getting the structure right is measured in the design win. A supplier that can tell an OEM, at the technical review, which filings are needed, who holds each one, and how the data will be handled, is easier to integrate than a supplier that arrives with a better model and no compliance story. In a market where the OEM carries the regulatory risk, the supplier’s compliance maturity is part of the product.
The reverse is equally true. A supplier that signs an uncapped indemnity to win the first programme has priced a recall it cannot survive, and the second programme will not arrive.
Practical checklist
- Decide whether you need a domestic entity for testing, and separate that question from the sales structure.
- Obtain the OEM’s applicable standard list and the evidence pack requirements before development starts.
- Map every filing to the party that holds it, and put that allocation in the contract.
- Design the data architecture for China first: in-vehicle processing, domestic storage, anonymised access.
- Confirm the map provider and the permitted use of its data before the first data collection.
- Assume the connected platform will be graded and audited, and budget the engineering time.
- Define the defect, the operational design domain and the acceptance test in the specification.
- Cap the indemnity, insure the recall exposure, and secure access to incident data.
- Fix the release authority and the change control process before the first update ships.
- Check your own export classification against the customer before the design win.
For help with the testing and access regime, data and mapping compliance, or the supplier contract for an autonomous driving programme, please contact Geason Law Offices. Related guides cover the autonomous driving legal framework, automotive standards for 2026 and product liability and recall.
Related Articles
- AI Robotics China Entry Guide for Foreign Companies
- Enterprise SaaS China Entry Guide for Foreign Companies
- China AI Startup Legal Cost Benchmark 2026
- China Hardware & Supply Chain Guide for Foreign AI Companies
Frequently Asked Questions↑ Back to article
Do autonomous driving & nev companies need a Chinese entity to enter China?
It depends on your business model. If you are only selling hardware components or software licenses cross-border, you may not need a local entity. But if you are providing services to users in China, handling Chinese user data, or having a physical presence — you typically need a domestic entity. We assess your specific situation and advise on the right structure.
How long does it take to launch an autonomous driving & nev product in China?
Timeline varies significantly by product type. Entity setup takes 6–8 weeks. Licensing and compliance add 2 weeks to 6+ months depending on your product category and regulatory requirements. Budget 3–9 months for a full compliant launch, depending on what certifications and filings you need.
What are the biggest compliance risks for foreign autonomous driving & nev companies in China?
The biggest risks vary by sub-sector but generally include: data compliance (data localisation and cross-border transfer), product safety and standard compliance (especially for physical products), algorithm and AI regulation, and licensing requirements. The specific risks depend on your product architecture and business model.
What licenses do autonomous driving companies need to test on public roads in China?
China has a tiered autonomous driving testing and licensing system. You need: (1) a test notification from local transport authorities, (2) qualified test drivers with safety officer training, (3) vehicles that meet safety standards and have passed third-party testing, (4) insurance or a bond, and (5) data security and cybersecurity compliance. Different cities have different rules — Beijing, Shanghai, Guangzhou, and Shenzhen have the most mature testing regimes.
Can foreign companies participate in China's NEV market?
Yes, but the market is highly competitive. Foreign NEV companies can set up wholly foreign-owned enterprises (WFOEs) for manufacturing and sales — China removed the 50% ownership cap for automotive joint ventures in 2022. However, you need to navigate complex regulations: production qualifications, CCC certification, data security, cybersecurity, geographic data security, and increasingly detailed autonomous driving regulations.
What data security requirements apply to autonomous vehicles in China?
Vehicles generate massive amounts of data including driving data, location data, and biometric data from driver monitoring systems, all subject to the Provisions on the Management of Automobile Data Security. These rules set specific requirements for data classification, in-vehicle processing, and restrictions on cross-border transfer of automotive data. Connected vehicle systems are often classified as critical information infrastructure, requiring MLPS Level 3 certification and regular security assessments.
What geographic information restrictions affect foreign autonomous driving companies in China?
High-precision mapping data used for autonomous driving is classified as sensitive geographic information in China, and foreign companies generally cannot collect or hold high-precision map data directly. Instead, foreign companies must work with licensed Chinese mapping providers to access the mapping data needed for autonomous vehicle operation. This is one of the key reasons many foreign AD companies form partnerships or joint ventures with Chinese firms that hold the necessary mapping licences.