China’s core AI industry was reported above RMB 600 billion in 2025, with enterprise adoption growing at more than 30 per cent a year, more AI unicorns than any country except the United States, the highest volume of AI patent filings, and an internet population above 1.1 billion. Those numbers are accurate and largely useless for a go-to-market decision, because none of them tells a foreign company which Chinese buyer will pay it, for what, and at what margin.
The useful question is narrower. In which parts of this market does a foreign AI company have an advantage that survives contact with a domestic competitor, and in which parts is it structurally outmatched? The answers differ sharply by segment, and they are not correlated with the size of the segment. The largest pools of AI spending are the hardest for a foreign vendor to serve. Some of the smaller ones are where foreign companies actually make money.
This note maps the demand side commercially: who buys, what they buy, how procurement works, where a foreign supplier wins and where it loses, and how to test a thesis in a quarter rather than a year. It deliberately avoids the growth language that surrounds this market. The opportunity is real, and it is narrower than the narrative suggests.
What the headline figure contains
The RMB 600 billion figure covers the core AI industry: chips and computing hardware, software and platforms, and AI services. It does not describe a single market with a single buyer. It aggregates at least four distinct businesses with different economics.
The first is infrastructure: chips, servers, computing capacity and data centre equipment. This is where a large share of spending sits, and it is dominated by domestic suppliers for reasons of procurement preference and, increasingly, supply availability.
The second is models and platforms, sold as APIs, private deployments or bundled cloud services. Pricing here has fallen steadily, and a competent model API is close to a commodity. Revenue concentrates among a small number of domestic providers with their own cloud distribution.
The third is applications and vertical solutions: software that embeds AI into a workflow a customer already pays for. This is fragmented, services-heavy, and the segment in which a foreign company with domain depth has the best chance.
The fourth is hardware products: robots, devices, sensors, edge systems. China is both the largest market and the manufacturing base, which makes the commercial logic different from the other three.
Growth of more than 30 per cent in enterprise adoption describes diffusion, not addressable revenue for a foreign vendor. Adoption is spreading through large enterprises and the public sector first, and those are precisely the buyers with the strongest preference for domestic suppliers. A foreign company reading the growth rate as demand for its product is reading the wrong line.
The demand map: five kinds of buyer
Chinese buyers of AI differ more from each other than foreign companies expect, and the differences decide contract terms, sales cycle and whether a foreign supplier is admissible at all.
State-owned enterprises and government units. Large budgets, formal procurement, and a preference for domestic software and domestic intellectual property that is sometimes a rule and always a consideration. Procurement runs through tendering processes with published requirements, qualification thresholds and scoring. A foreign vendor can win, usually as a subcontractor to a domestic prime or in a category where no domestic substitute exists. Payment terms are long and collection requires a local entity.
Large private enterprises. Manufacturing groups, retailers, logistics operators and internet platforms. They buy outcomes, run competitive proofs of concept, and expect the vendor to integrate with systems built in China by integrators they already use. Decisions are faster than in the public sector and more price-sensitive. These are the best buyers for a foreign vendor with a differentiated product.
Foreign multinationals operating in China. The Chinese subsidiary of a European, Japanese or American group often wants the same tooling its headquarters uses, and has a budget in a foreign currency and a procurement process it understands. This is the most accessible segment for a foreign AI company, and the most frequently overlooked. The buyer is not choosing between you and a Chinese competitor; it is choosing between you and nothing.
Chinese companies selling abroad. Exporters in consumer electronics, e-commerce, gaming, automotive and industrial equipment need AI tooling that works across jurisdictions: content moderation in multiple languages, compliance screening, customer support, advertising attribution, fraud detection. A foreign vendor’s familiarity with overseas platforms and rules is an advantage here that it does not have when selling into the domestic market.
Consumers. Enormous, cheap to reach, hard to monetise at Western price points, and heavily regulated where the product is a public generative AI service. Consumer revenue is possible but usually requires a licensed structure and a local operating team.
The practical lesson is that the accessible demand is business-to-business, and within that, the two most accessible buyer types are foreign multinationals in China and Chinese companies selling abroad. Both are segments where being foreign is neutral or positive. Neither appears in the market-size narrative.
How the purchase actually happens
The sales motion differs from the one a foreign vendor knows, and the differences decide whether a deal closes.
Large Chinese buyers run structured evaluations. A request for information narrows the field, a request for proposal sets scoring criteria, and a proof of concept follows. The scoring is often published in advance and weights factors beyond price: local service capability, delivery record with comparable customers, qualification certificates, and sometimes domestic intellectual property. A vendor that scores well technically and poorly on local presence loses to a weaker product with a local team.
Qualification thresholds are the quiet barrier. Tenders and enterprise procurement processes often require a minimum period of operation, audited accounts, a local tax record and references from named customers. A newly incorporated entity meets none of these, which is why the first year of a WFOE is usually spent building a record rather than winning large contracts. A company that plans revenue in month three is planning against the procurement calendar rather than its own.
Channels carry most of the volume. Chinese system integrators and industry solution providers hold the customer relationships, the implementation teams and the qualifications. Selling through them gives reach and costs margin and visibility. Selling direct gives control and requires a local delivery capability that most foreign companies underestimate. The realistic structure is a hybrid: direct for the accounts that justify it, channel for the rest, with the channel agreement drafted to preserve renewal ownership and data rights.
Contract mechanics also differ. Payment is commonly staged against milestones and acceptance, with a retention element, and the acceptance criteria decide when the clock starts. Invoices must be the right type for the customer to deduct the expense. These are not legal technicalities; they determine the cash conversion cycle, which for a foreign parent is often the difference between a viable China operation and a permanent funding requirement.
What policy support does and does not change
AI is a national priority, and municipal governments compete to build local industries. The instruments are real: industrial parks with preferential rent, grant programmes for qualifying projects, support for computing costs, talent schemes, and procurement of AI applications by local government and its platforms.
Two things follow for a foreign company, and neither is a reason to enter on its own. The first is that subsidised demand creates customers. A city that funds AI adoption in manufacturing creates buyers for the tooling that adoption requires, and those buyers are reachable by a foreign vendor with the right product. The second is that the support is local in a way that surprises headquarters: eligibility usually requires local registration, local staff, local tax payments and a project that fits a published category, with an application window that opens and closes. The measures describe intent; the annual notice describes the transaction.
A foreign company should treat incentives as a margin improvement on a decision already taken commercially, not as the decision. A grant that requires a team, a lease and a local project is not free money; it is a subsidy against costs the company has already chosen to incur.
Where the constraints bite
Two constraints limit the opportunity more than competition does, and both are legal rather than commercial.
Data is the first. A product that improves with Chinese user data cannot freely move that data offshore, and the compliance route depends on volume, sensitivity and the architecture. The constraint reaches the product, not just the legal file. A model that trains centrally on global data and serves China from a single endpoint may be the wrong architecture for the market, and rebuilding it for local inference changes the unit economics.
Intellectual property is the second. The categories where foreign companies win are the categories where the asset is a design, a certification or a standard. Those are exactly the assets that a partner most wants access to. The commercial opportunity and the IP exposure are the same asset viewed from two sides, which is why the licence, the joint development terms and the access controls have to be settled before the first workshop rather than after the first prototype.
Where foreign companies actually win
Six categories account for most of the successful China revenue earned by foreign AI companies. None depends on model superiority.
Selling into the supply chain rather than to the consumer. For a large part of the foreign AI industry, China is a place to buy, not a place to sell. Chip designers, robotics companies and device makers come for components, contract manufacturing and iteration speed. Shenzhen concentrates chip designers, sensor makers, robotics component suppliers and contract manufacturers more densely than any other city, and prototyping that takes months elsewhere takes weeks there. The commercial opportunity is procurement and partnership, and it is the single most reliable reason a foreign AI company establishes a presence.
Tooling for model operations and assurance. Evaluation harnesses, red-teaming, observability, data labelling pipelines, model security testing and compliance documentation. Chinese enterprises facing their own filing and security obligations need evidence, and foreign tooling that produces defensible evidence has a market. The buyer is often the compliance or security function rather than the AI team, which changes the sales motion.
Vertical applications where the standard is global. Medical imaging, automotive functional safety, financial risk modelling, industrial inspection for exported goods. In these segments the customer’s own regulator or overseas customer imposes a standard, and a product certified against it has value that a domestic competitor must spend years to replicate. The moat is the certification and the validation data, not the model.
Cross-border infrastructure. Payments, identity, cloud regions, content delivery, ad tech and analytics for Chinese companies operating abroad. The foreign vendor’s advantage is that it already operates in the jurisdictions the customer is entering.
High-value components and design IP licensed to Chinese manufacturers. Where the foreign company’s asset is a design, an algorithm embedded in silicon, or a manufacturing process, licensing to Chinese OEMs can generate revenue without a consumer-facing presence. The exposure is IP leakage, and the structure has to be designed for it.
Enterprise software for multinational subsidiaries. A Chinese subsidiary that must report into a global system of record will pay for a product that integrates with it. The competition is internal standardisation, not a local vendor.
In each of these, the foreign company is paid for something a domestic competitor cannot easily copy: access to a supply chain, a certification, a global footprint, an installed base, or a standard. Where none of those applies, the foreign company is selling a model, and models are cheap.
Where foreign companies lose
The losing categories are equally specific, and companies enter them anyway because the market looks large from outside.
Consumer applications against domestic incumbents. Distribution is controlled by a handful of platforms, user acquisition costs are high, and the domestic competitors iterate weekly with local engineering teams. A foreign app with a better model but a slower release cycle loses on the only axis consumers notice.
Horizontal productivity software priced per seat. Chinese buyers pay for measurable outcomes and cost reduction more readily than for seat-based productivity tools. Price points are below United States and European levels, and enterprise buyers expect customisation and local deployment to be included. A global list price does not survive the first negotiation.
General-purpose model APIs. Price competition is severe and the domestic alternatives are adequate for most tasks. A foreign provider wins only where a specific capability, a data residency arrangement or a contractual commitment justifies a premium, and that advantage has to be re-earned every quarter.
Public-sector prime contracts. Procurement rules, domestic IP preferences and security requirements make the prime role difficult for a foreign entity. Subcontracting to a domestic prime is the realistic route, and it concedes the customer relationship.
Anything requiring a commercial ICP licence on a 100 per cent foreign-owned entity. Value-added telecom services are generally capped at 50 per cent foreign equity, so consumer services that charge for information require a licensed partner or a joint venture. A company that plans to compete in that space without addressing the cap is planning to lose on a legal technicality.
Customisation-intensive integration work. Chinese customers expect the vendor to adapt the product to their systems, in their language, on their timeline. A foreign company without local engineering capacity loses to integrators who are physically present. Services revenue is available, but it is a different business with different margins.
The pattern is consistent. Foreign companies lose where the product is a commodity, where distribution is local, and where the customer expects the vendor to absorb customisation. They win where the product carries something the customer cannot obtain domestically.
The supply chain opportunity is asymmetric
Most analysis of the China AI market treats it as an export market. For a large part of the foreign AI industry, it is an import market, and the asymmetry matters.
A robotics company sourcing actuators, a device maker sourcing sensors, a chip company using Chinese packaging and test, an AI hardware startup using contract manufacturers: each is buying capability that exists at a density and cost no other location matches. The legal work in these relationships is not market entry. It is supplier contracts, quality and warranty terms, IP ownership in tooling and designs, export control classification, and the enforcement problem when a supplier relationship deteriorates.
Two consequences follow. First, a company whose China interest is sourcing does not need a WFOE to start; it needs a supplier qualification process and a contract that survives a dispute. Second, the sourcing relationship is the cheapest way to learn the market. A company that spends a year buying from Shenzhen suppliers learns more about Chinese industrial demand than one that spends the same year on market research.
The supply chain also creates a selling opportunity that is easy to miss. Chinese manufacturers selling into Europe and North America need compliance, traceability and quality tooling that meets foreign standards. A foreign company already selling that tooling at home can extend it to its Chinese suppliers, and the supplier becomes a customer for the compliance layer the foreign company already owns.
Talent and research: what a China centre is for
China graduates a large number of STEM doctorates and AI researchers, and the major platforms have trained a generation of engineers who now work across the ecosystem. Compensation for senior AI talent is competitive with the United States, and mid-level engineering talent is abundant relative to Silicon Valley or Western Europe.
That combination supports a specific use: an engineering centre that builds, tests and localises, rather than a research lab that pursues frontier work. Companies that set up a China R&D centre to do foundational research often find that the best researchers are being recruited by domestic labs with larger mandates and faster publication cycles. Companies that set it up to build product, maintain the local build and support customers get predictable value.
Three constraints shape the decision. Export control limits what hardware and some model technology can be worked on inside China. Data rules limit what Chinese data can be used and where it can go. And IP protection requires the contractual and technical controls described elsewhere; a research centre that cannot protect its output is a leak with a payroll.
City choice by function
The four main hubs are not interchangeable, and the choice should follow the function rather than the prestige of the address.
| City | Strongest for | Weaker for | Who should choose it |
|---|---|---|---|
| Shenzhen | Hardware, robotics, edge devices, supply chain access, government support for tech | Frontier research talent depth; distance from central regulators | Hardware and device companies, robotics, companies whose customers are manufacturers |
| Beijing | AI research talent, universities, proximity to national regulators and policy | Cost, competition for engineers, distance from manufacturing | Research-heavy companies, policy-sensitive businesses, companies needing regulator contact |
| Shanghai | Multinational customers, financial services, enterprise software, international environment | Hardware supply chain; cost | Enterprise software vendors, companies selling to foreign multinationals, financial AI |
| Hangzhou | E-commerce, cloud, consumer internet ecosystem | Hardware; breadth outside the platform ecosystem | Companies selling into retail, logistics and consumer internet |
A company can start in one city and add a second later. What it should not do is register in a city because a subsidy was mentioned, then discover that its customers, engineers and suppliers are elsewhere. Local incentives are real and secondary; the demand and the talent are primary.
Pricing and unit economics
The commercial discipline that most often fails is pricing. Chinese enterprise buyers benchmark against domestic list prices, expect customisation to be included, and negotiate payment terms that strain a foreign balance sheet. A vendor that prices at home-market levels loses at the shortlist stage; a vendor that discounts to win discovers that the account is unprofitable after localisation and support.
Three practices help. Price the localised product, not the global product with a discount. Separate licence from implementation in the proposal, so that customisation is visibly priced rather than absorbed. And model the receivables cycle, because enterprise and public-sector payment terms in China are longer than in most Western markets and collection depends on a local entity with local standing.
Hardware economics are different again. Margin depends on the bill of materials, the yield, and the volume commitment, and the Chinese supply chain rewards companies that can commit to volume. A foreign company that treats sourcing as a procurement transaction rather than a supply relationship will pay more and iterate more slowly than a competitor that invests in the relationship.
Testing the thesis in ninety days
The cheapest way to assess this market is a structured ninety-day test, run before any entity is registered.
Weeks one to four: identify ten prospective buyers across the two accessible segments, foreign multinationals in China and Chinese exporters. Interview five. Ask what they use, what they pay, who approves the purchase and what would block a foreign supplier. The last question surfaces procurement and data objections early.
Weeks five to eight: run a technical proof of concept with one or two of them, in the environment they actually use. A proof of concept that runs in your cloud and not theirs proves nothing about deployment.
Weeks nine to twelve: price the deal as it would actually be contracted, including implementation, local support and the payment terms. Compare the margin with your home-market margin for the same product. If the localised margin is negative, the segment is not viable at any volume, and no amount of market growth changes that.
The output is a decision with evidence: which segment, which buyer, which price, which structure. It costs a quarter and a fraction of an entity, and it prevents the most common failure, which is to incorporate first and discover the commercial answer second.
Where companies misjudge the opportunity
Reading adoption growth as demand for a foreign product. Adoption is concentrated among buyers with the strongest domestic preference. The growth rate describes the market, not the accessible slice of it.
Assuming model quality is the differentiator. Domestic models are adequate for most tasks and improving continuously. The differentiator has to be something the customer cannot buy locally.
Entering as a consumer business. Distribution, pricing and licensing all work against a foreign consumer product. The business-to-business route is slower and more durable.
Underestimating services. Chinese customers expect implementation support. A company that refuses services work loses deals to integrators; a company that accepts it without pricing it loses money.
Treating the supply chain as a procurement detail. For hardware companies the supply chain is the market. The relationship, the IP terms and the quality regime are the commercial substance.
Planning a launch before the filings. A compliant launch takes months, not weeks. Algorithm filing alone can run from four to twelve weeks depending on complexity, and the entity and ICP filing come first.
Practical checklist
- Choose the segment before the city, and the city before the entity.
- Interview five buyers in the two accessible segments: foreign multinationals in China, and Chinese companies selling abroad.
- Test whether the buyer can legally procure from a foreign supplier, and who signs.
- Identify the specific thing you offer that a domestic competitor cannot copy within a year.
- Price the localised product and the implementation separately, and model the receivables cycle.
- If hardware is involved, treat supplier qualification as the entry project and the entity as optional at first.
- Run a proof of concept in the customer’s environment, not your own.
- Map the filing path before promising a launch date: entity, domain, ICP filing, then AI-specific filings.
- Decide what stays offshore, and paper the licence of group IP on arm’s length terms.
- Set a walk-away margin below which the segment is not pursued, and apply it before the second hire.
This note reflects the position as of late 2026; market structure and local practice move faster than the statutes, so verify segment-specific rules before committing capital. The decision framework that should follow this analysis is in our guide on whether an AI company should enter China, and the entity mechanics are in our WFOE guide. For entry planning and market validation support, see /en/services/incorporation/.
Related Articles
- China Hardware & Supply Chain Guide for Foreign AI Companies
- AIGC Apps China Entry Guide for Foreign Companies
- Autonomous Driving & NEV China Entry Guide
- China Manufacturing Cities Compared: Shenzhen vs Dongguan vs Suzhou vs Ningbo for Hardware Startups
Frequently Asked Questions↑ Back to article
Why do foreign AI companies enter China?
Foreign AI companies enter China for five main reasons: access to the world's largest consumer internet market, the deepest AI hardware and robotics supply chain, the fastest commercialisation speed anywhere, top AI engineering talent density, and significant government policy support for AI industry development.
What is the size of China's AI market?
China's AI core industry market exceeded RMB 600 billion (~$83 billion) in 2025, with enterprise AI adoption growing at 30%+ annually. The broader AI-enabled economy is in the trillions of RMB. China has more AI unicorns than any country except the US.
Which AI sectors are strongest in China?
China has particular strengths in computer vision, generative AI application layer, AI chips and semiconductor design, robotics and industrial automation, autonomous driving, smart manufacturing, and AI for healthcare and education.
Do you need a local entity to sell AI products in China?
Not always, but usually yes for any meaningful presence. If you sell directly to Chinese enterprise customers, handle Chinese user data, or need ICP licensing, you need a Chinese entity. For pure cross-border SaaS with no China data footprint, it may be possible without one — but there are risks around data and taxation.
How long does it take to set up an AI company in China?
A WFOE (Wholly Foreign-Owned Enterprise) typically takes 6–8 weeks from start to business licence. Adding ICP filing takes another 2–4 weeks, and algorithm filing (if required) takes 4–12 weeks depending on complexity. Budget 3–6 months for a full compliant launch.
What are the main supply chain advantages for AI hardware in China?
China has the world's most complete electronics and AI hardware supply chain, with Shenzhen hosting the densest concentration of AI chip designers, sensor manufacturers, robotics component suppliers, and contract manufacturers of any city. For products involving hardware, robotics, edge AI, or AI chips, China offers unmatched speed, scale, and cost efficiency that other locations cannot match. Prototyping cycles that take months elsewhere take weeks in Shenzhen.
Which are the main AI hub cities in China for foreign companies?
The three primary AI hubs are Shenzhen, Beijing, and Shanghai, each with different strengths. Shenzhen leads in hardware and manufacturing with the deepest supply chain ecosystem and strong government support for tech companies. Beijing has the deepest AI research talent pool, top universities, and proximity to national regulators. Shanghai is the financial and commercial capital with the largest concentration of multinational companies and a more international business environment. Many foreign AI companies start in one city and expand to others as they grow.