A board that asks whether the company should enter China is usually asking a narrower question than it thinks. The real question is which of four quite different commitments the company is prepared to make, and whether the return on the heaviest of them justifies the regulatory apparatus that comes with it. China is the deepest concentration of AI adopters in the world and the supply chain on which much of the industry already depends. It is also a jurisdiction in which the entity that serves users must sit inside the country, hold the filings, and answer to a regulator that can suspend a service.

For an AI company the decision rarely resolves into a clean yes or no. It resolves into a choice among entry models, each of which buys a different kind of evidence at a different price. A cross-border sales motion costs weeks and tests demand. A wholly foreign-owned enterprise costs a quarter and a compliance function. A joint venture costs control. The mistake boards make is to authorise the heaviest model first, on the strength of a market narrative, and to discover the legal constraints after the capital is committed.

This note sets out a decision framework a board can actually use: six tests that carry weight, a seventh that measures whether the loss is survivable, a scoring table with a threshold fixed in advance, and three worked cases showing how the same tests produce different answers. It assumes the reader wants a revenue-generating AI business serving Chinese customers, because that is where the legal weight sits. Companies that only want to buy hardware, or license technology to a Chinese operator, face a much lighter set of questions and should not use this framework to answer them.

Four decisions that get confused

“Entering China” bundles four commitments, and most failed plans confuse the first with the second.

Selling from outside the country, by export or cross-border software subscription, is the lightest. No entity, no filing, no local payroll. Operating a service for the Chinese public is heavier by an order of magnitude, because the operator must be a domestic legal person able to hold a website filing, a telecom licence where the service is commercial, and the algorithm and generative AI filings that attach to public-facing AI. Building product or training models in China is heavier again: staff, data, source code and model weights sit inside the jurisdiction, and employment, tax and intellectual property consequences follow the team rather than the contract. Sourcing hardware or licensing a Chinese model is a fourth, largely separate question with its own exposure.

The distinction matters commercially because the second commitment is often triggered by accident. A company sells a software product into China, then adds a Chinese-language interface, a local payment channel and a public sign-up page. At that point it is operating an internet information service, and the filings it needs cannot be held by its overseas parent. The product has become a regulated service without anyone deciding that it should.

Before any other analysis, the board should write down which decision it is making. The framework below addresses the second and third.

Market pull: whether demand exists beyond the anchor

China’s AI demand is real, and it is unevenly distributed. The segments that spend are not always the segments a foreign vendor can serve.

Enterprise buyers in manufacturing, logistics, retail and financial services buy AI for measurable outcomes: yield, defect rates, forecasting accuracy, headcount avoided. They buy pilots readily and production deployments slowly, and they expect the vendor to integrate with systems that were built in China, in Chinese, by integrators they already trust. Consumer demand is enormous but monetises at Chinese price points, through advertising, subscriptions priced for local incomes, or bundled hardware.

Government demand is the largest single pool and the most conditional. Smart city, public security, healthcare, education and manufacturing upgrade programmes carry genuine budgets. They also carry requirements that reshape a foreign product: domestic data residency, security grading, sometimes domestic chips, and procurement preferences for domestic software and domestic intellectual property. A public-sector contract can be the fastest route to revenue and the fastest route to a product you no longer control.

Three questions test pull properly. First, how many qualified buyers exist with confirmed budget, independent of the anchor account? If three customers would make the China year, the company is servicing a relationship, not entering a market. Second, can those buyers legally procure from a foreign vendor, and does their procurement process accept a foreign-owned supplier? Third, will they pay a price that leaves a margin after localisation and support?

A useful discipline is to interview five prospective buyers rather than one, and to ask each what they currently use, what they pay for it, and what would have to be true for them to switch. Answers that converge on a specific budget line and a specific incumbent are worth more than enthusiasm.

Competitive position: the margin you can defend

Chinese foundation model providers and application vendors are numerous, fast and cheap, and several are funded to buy share. A competent model API is close to a commodity. That does not close the market to foreign companies, but it relocates the basis of competition.

Foreign vendors tend to win on reliability at scale, on compliance posture that a Chinese customer needs for its own audit or export business, on vertical depth where the product encodes domain knowledge rather than raw model capability, and on integration with the global systems a Chinese company uses to sell abroad. They rarely win on model quality alone, on price, or on speed of local customisation against a domestic competitor with engineers in the same time zone.

The board test is not “is our model better” but “which buyer, facing which alternative, would pay a premium for what we have, and for how long”. A capability lead measured in months is not a moat in a market where the leading domestic labs ship continuously. A workflow, a dataset, a certification or a global integration is.

Price deserves its own line. Chinese enterprise buyers negotiate hard, expect customisation to be included, and compare against domestic list prices that are typically below United States or European levels. Model unit economics at Chinese prices, not at home-market list price. If gross margin collapses after localisation, the market is not attractive at any volume.

Regulatory cost: four gates, priced as a running cost

Four gates determine whether a plan is executable, and each has a cost that recurs rather than a cost paid once.

Foreign investment access is the first. China governs inbound investment through a negative list naming restricted and prohibited sectors. Most of the AI value chain, including software development, research and AI hardware manufacturing, is open to wholly foreign-owned entities. The binding constraint is telecommunications. Value-added telecom services, which include most commercial internet information services, are generally capped at 50 per cent foreign ownership, and a commercial ICP licence sits in that category. Pilot programmes in selected free trade zones have relaxed some categories and local practice differs, so the current position must be confirmed for the specific service rather than assumed.

Data localisation and export is the second. The Personal Information Protection Law, the Data Security Law and the Cybersecurity Law together govern collection, storage and transfer abroad. Exporting personal information above the volume thresholds set by the Cyberspace Administration of China triggers a security assessment, a standard-contract filing or a certification. Cross-border rules issued in 2024 raised thresholds and simplified routine flows, but a China-facing AI service that sends user data to an overseas model endpoint remains squarely in scope, and the architecture that makes it lawful has to be designed before launch.

Algorithm and generative AI filing is the third. A service using algorithmic recommendation with public opinion attributes or social mobilisation capability must file its algorithm with the CAC. A generative AI service offered to the Chinese public must also complete a security assessment and filing under the Interim Measures for the Administration of Generative AI Services. Provincial offices in practice expect application-layer providers to register even where the underlying model is already filed, which means a foreign company wrapping a Chinese model in its own product is still an applicant.

Content and labelling is the fourth. AI-generated content must be labelled under the AI-Generated Content Labelling Measures in force since 2025, in visible and metadata form. Moderation duties attach to the operator rather than the model vendor. That converts content review from a launch task into a permanent staffing line, and it is the line boards most often omit from the business case.

The regulatory cost is therefore not a legal fee. It is a compliance function: security grading under the Multi-Level Protection Scheme, periodic assessments, filing maintenance, moderation staff, and the engineering time to keep data flows inside the permitted architecture. Price it over three years, and price it as headcount rather than as a project.

Enforcement shapes the cost as much as the rules do. Filing breaches typically begin with a rectification order and escalate to service suspension or removal of the application from Chinese app stores. For a consumer product, removal is not a fine; it is the end of distribution, and reinstatement runs through the same queue that the company failed to join. For a business-to-business vendor the consequence is different and often worse commercially: enterprise customers conduct procurement diligence, and a supplier whose filing is incomplete fails that diligence before any regulator acts.

The gates also interact. A decision to keep data offshore to reduce local exposure can force a cross-border assessment; a decision to keep the core model offshore can limit what the local team may build; a decision to avoid a telecom licence can remove the consumer revenue line that justified the entity. Boards that evaluate the four gates separately tend to approve a plan that fails on the combination, which is why the legal view should be a single written opinion covering the product as it will actually be operated.

IP exposure: what leaves the building

Intellectual property is the risk most often cited by companies that have operated in China, and the one most amenable to structure.

Three channels carry the exposure. Joint development puts your engineers and a partner’s engineers in the same repository, and the ownership of what they build together is decided by a contract drafted before the first commit, not after. Staff mobility moves knowledge between employers in a market where senior AI engineers change jobs readily, and where a departing employee’s next employer may be a competitor funded to hire precisely that person. Supplier access exposes training pipelines, evaluation sets and hardware designs to counterparties whose commercial interests are not identical to yours.

Chinese trade secret law works, and courts do grant relief, but only where the company can show that it took reasonable steps: written confidentiality and IP assignment terms, access controls that are actually enforced, records of what was accessed by whom, and a registered copyright or patent position that gives a cause of action. A company that cannot produce those records has a claim it cannot prove.

The structural answer is to decide, before signing, what stays offshore. Model weights, training data, evaluation harnesses and the core inference stack can often remain outside China while the local entity builds adapters, integrations and vertical applications. Where the product requires the core to sit locally, the protections have to be contractual, technical and evidentiary at the same time.

Export control runs in the other direction and is equally a board matter. Restrictions on advanced chips and on certain model or tooling exports affect what can be deployed in China. China’s own export control regime now covers dual-use items and some technical data. A plan that assumes unrestricted transfer of either hardware or model technology between the two jurisdictions is a plan that will need rewriting.

Partner dependence: the cost of borrowing a licence

Most China entry plans that are not a WFOE depend on a Chinese partner, and partner dependence is a risk category of its own rather than a footnote to structure.

A distributor or reseller gives coverage without control. The partner owns the customer relationship, the pricing conversation and often the service quality, and the foreign company discovers churn and discounting months after they begin. Contracts can require reporting and set minimum performance, but enforcement in practice depends on whether the relationship is worth more to the partner than the dispute.

A licensee gives revenue without operations, and takes the technology into a company whose incentives diverge over time. Licence scope, field of use, sublicensing, improvement ownership, audit rights and termination consequences decide whether the arrangement remains a licence or becomes a transfer. Chinese licensees negotiate hard on all of them, and a licence that does not define improvements and derivative works will be read in the licensee’s favour in practice even where the contract is silent in law.

A joint venture gives access to a licence or to government customers, and takes governance. Minority protection, deadlock, transfer rights, dividend policy and exit mechanics have to be negotiated while both sides still want the deal. Joint ventures in restricted sectors also carry a specific hazard: the Chinese shareholder’s equity is what makes the licence lawful, which means the foreign party’s leverage declines the moment the licence issues.

The test is simple. If the partner disappeared tomorrow, would the company still have the customers, the data and the product? If not, the partner is not a channel; it is the business, and the board is valuing someone else’s asset.

Capital and time: the shape of the commitment

Cost has four lines, and only one of them is a one-off.

Entity formation and licensing takes two to four months for a straightforward wholly foreign-owned enterprise, longer where telecom licences or a joint venture are involved, and the licence timeline runs after the entity exists rather than in parallel with it. Compliance is recurring: security grading, assessments, filing maintenance and moderation staffing. People are the largest line, because senior machine learning engineers in Shenzhen, Beijing and Shanghai command compensation comparable to Singapore and well above Southeast Asia, with turnover that forces repeated hiring. Localisation is the line most often underestimated: Chinese-language product, local cloud, local payment, local support, and the engineering to keep the local build in step with the global one.

Time behaves differently from money. Capital can be increased; a licence queue cannot be shortened by spending. A company that plans to launch three months after incorporation is usually planning to launch before its filings are approved, and the gap is filled by either delay or non-compliance.

Two structural points deserve board attention. Registered capital is a commitment, not a deposit: under the revised Company Law it must be paid within five years, and the amount must be commensurate with the scale of operations. Setting it low enough to be comfortable can leave the entity unable to fund payroll and compute, or unable to satisfy a licence applicant’s balance sheet expectations. Setting it high creates a funding obligation and a liability ceiling that follows the company.

The second point is exit. An entity that has never traded is cheap to close. An entity with staff, a lease, filings and a tax history is not. The board should decide, before incorporation, what evidence would justify continuing and what would justify closing, and should record it.

The scorecard

Score each dimension from 1 to 5, apply the weights, and fix the go threshold before scoring. A threshold set afterwards is a rationalisation.

DimensionWeightScore 1Score 3Score 5
Demand beyond one anchor account20%One buyer, no budget lineThree buyers, one with budgetFive or more qualified buyers with confirmed budget
Competitive position and margin15%Model parity, price competitionOne defensible workflowVertical depth or global integration a domestic rival cannot match
Regulatory path20%Consumer service needing a licence the structure cannot holdB2B service with a clear filing routeSoftware and research only; no telecom licence needed
Data and cross-border feasibility15%Core function requires unrestricted export of personal informationExport possible under a filing or assessmentNo personal information leaves China, or a designed route is approved
IP defensibility15%Core model must sit with a partnerCore offshore, adapters localCore offshore, contractual and technical controls tested
Partner and talent availability10%No credible local leaderLeader identified, not committedLocal leader hired and partner terms agreed
Downside tolerance5%Loss would threaten the companyLoss absorbed but painfulLoss is affordable and ring-fenced

A weighted score above the threshold is necessary, not sufficient. Two dimensions are vetoes regardless of the total: a regulatory path scored at 1, and a data position that cannot be made lawful. A high aggregate cannot cure either, because neither is a matter of appetite.

Three worked cases

The framework is easier to trust when it produces different answers from similar-looking facts.

Case one: enterprise software vendor, enter now. A company sells computer vision quality inspection to electronics manufacturers, has five qualified Chinese buyers with budget, and its product runs as an on-premise deployment inside the customer’s factory. Data stays in the customer’s environment. The core model stays offshore and is updated through a controlled release process. Score: demand 5, competitive position 4, regulatory path 5, data 5, IP 4, talent 3, downside 4. Weighted, this lands comfortably above a threshold of 3.5. The right model is a WFOE in a manufacturing city, a local support and integration team, and no consumer-facing service. Time to revenue is driven by hiring rather than by filings.

Case two: consumer generative AI app, partner first. A company has a consumer writing assistant with strong engagement in its home market and wants to launch in China. The product is a public generative AI service, which means a generative AI filing, an ICP filing, moderation staffing and, if it charges users for information services, a commercial ICP licence that a 100 per cent foreign-owned entity will generally not obtain. Demand is plausible but unproven at Chinese price points. Score: demand 3, competitive position 2, regulatory path 1, data 3, IP 3, talent 3, downside 4. The regulatory veto applies. The rational move is not to abandon China but to change the question: license the product to a licensed Chinese operator, or take a minority position in a joint venture, and treat the arrangement as market research with revenue attached. Entering directly would spend a year and a compliance budget to reach a launch the structure cannot support.

Case three: model company under export control, stay out. A company trains large models on hardware it cannot lawfully export, and its Chinese demand comes from a single state-owned anchor whose procurement rules favour domestic suppliers. Score: demand 2, competitive position 3, regulatory path 2, data 2, IP 2, talent 2, downside 2. The aggregate is below any sensible threshold and two dimensions are close to veto. The correct decision is a documented no, with a review trigger: if the customer base widens beyond the anchor, or if a licensing route emerges that keeps the core offshore, the case can be reopened. A written decision not to enter costs a fraction of an abandoned subsidiary, and it protects the board from revisiting the question every quarter on the strength of a news cycle.

The three cases share a feature worth naming. In each, the answer was determined by two or three dimensions, not by the aggregate. The scorecard’s value is that it forces the board to identify which dimensions those are before the debate becomes about enthusiasm.

Where boards go wrong

Four failure modes recur.

Authorising the entity before testing the licence. The entity is the visible commitment, so it gets approved first. But the licence determines whether the entity can do the work, and the licence question is answerable in a week of advice. Sequence the licence test before the incorporation resolution.

Treating compliance as a launch cost. Security grading, moderation and filing maintenance are running costs, and they scale with users. A business case built on the setup fee will show a profit in year one and a loss in year two.

Underestimating the local build. The Chinese product is not the global product with translated strings. Cloud, payment, identity, mapping, messaging and support all differ, and the local build has to be maintained in parallel. Companies that budget for a one-time localisation pay for it twice.

Confusing a partner with a market. A partner that brings one large customer is a channel. A partner that brings the licence, the customers and the data is the business. Boards should value the two differently and should say which one they are buying.

Practical checklist

Rules in this area change continuously, and this note reflects the position as of late 2026. The entity comparison that follows from the third gate is set out in our guide to WFOEs for AI companies, and the demand-side view is in our note on China AI market opportunities. For structure, licensing and sequencing, see /en/services/incorporation/.

Frequently Asked Questions↑ Back to article

Can a foreign AI company sell into China without setting up an entity?

It can sell software or services cross-border, and many companies start that way. But if the product is offered to the Chinese public as an online service, the operator must generally be a domestic legal person able to hold the website filing, telecom licence and AI filings, which an offshore seller cannot do.

Is China's foreign investment negative list a barrier for AI companies?

Rarely for software and AI hardware, which are largely open to wholly foreign-owned entities. The binding restriction is in value-added telecom services, where foreign ownership is generally capped at 50 per cent, and that is the constraint most AI application companies actually hit.

How much does it cost to enter the China market?

There is no reliable public benchmark, because cost depends on entity type, licences, headcount and localisation depth. Plan for entity formation, recurring security grading and filing maintenance, senior engineering salaries at Shenzhen or Beijing levels, and a localisation budget that is commonly underestimated.

What is the biggest risk of entering China for an AI company?

Intellectual property leakage through joint development, staff mobility and supplier access is the risk most often cited, followed by export control exposure and the possibility that a home-market regulator or customer objects to the China presence. Each is manageable with structure, but only if designed before contracts are signed.

Do we need an algorithm filing before launching an AI product in China?

Not every AI product, but any service using recommendation algorithms with public opinion attributes or social mobilisation capability must file with the Cyberspace Administration of China. A generative AI service offered to the public also needs a security assessment and filing, and provincial offices often require application-layer registration.

Is a joint venture the only way for a foreign AI company to operate in China?

No. A wholly foreign-owned enterprise works for most software, research and hardware activities, and a joint venture is mainly needed where a telecom licence or government access requires a Chinese shareholder. Licensing and channel models avoid an entity altogether.

When should a company decide not to enter China?

When demand depends on a single anchor account, when the product cannot operate without exporting personal information that your home regulator restricts, or when your core IP cannot be protected in a partner-led model. A documented decision not to enter is a legitimate strategic outcome.

Geason Law Offices AI Industry Research Group | Reviewed by He Qianggao (Lawyer License No. 14403201310092337, 10+ years of practice)