Apple and Alibaba train a custom AI for China

Apple and Alibaba train a custom AI for China

Apple Intelligence faces China's restrictions

Apple reportedly developed an artificial intelligence model specifically intended for the Chinese market with Alibaba's help. The information, reported by The Verge, sheds light on the complex strategy of the Cupertino group in a country where its AI services cannot simply be imported as they are designed in the United States or other international markets.

The issue goes far beyond launching a software feature on the iPhone. It concerns Apple's ability to retain a significant position in China, the way major AI models must be adapted to local rules, and, more broadly, the growing fragmentation of the global technology ecosystem. To operate in Chinese generative AI, Apple cannot rely solely on its own models, servers and US partners: the group must work with local companies and within a distinct regulatory framework.

Apple introduced Apple Intelligence at its WWDC conference in June 2024. The platform brings together several functions based on AI models: writing tools, notification summaries, image generation with Image Playground, creation of custom emojis, enhanced search in device content, and an evolution of Siri. Apple also chose to integrate OpenAI's ChatGPT for certain use cases, with the user's consent, particularly when the assistant needs to draw on expertise or provide an answer beyond its own capabilities.

But the rollout of Apple Intelligence has been gradual and geographically limited. The first features arrived in US English with iOS 18.1 in fall 2024. Apple then expanded language support to other English variants before announcing additional languages. Mainland China remained a special case. Apple itself indicated that certain features would not be available in all countries or languages, and that their availability would depend in particular on applicable regulations.

This caveat is fundamental. Apple Intelligence is not a simple offline software package. Part of its processing is performed directly on the device, thanks to Apple Silicon chips in compatible iPhones, iPads and Macs. But certain more complex needs are sent to Private Cloud Compute, the private cloud infrastructure designed by Apple to run larger models. Other requests can, with consent, be sent to ChatGPT. This hybrid model involves data, infrastructure, providers and models whose international deployment must be compatible with each country's requirements.

In China, this principle comes up against a particularly structuring regulatory environment. Generative AI services accessible to the public must comply with security and compliance obligations. Local providers have been required to register their services or models with the relevant authorities before making them available to the general public. Foreign companies, for their part, must take into account rules concerning data, content, infrastructure and partners that may operate the services.

The choice of Alibaba, if confirmed under the reported terms, is therefore not merely a technical decision. It would also be a response to this regulatory architecture. Alibaba is one of the Chinese groups most active in cloud computing, e-commerce and AI. Its cloud unit develops the Qwen family of models, originally launched under the name Tongyi Qianwen. The group therefore has computing capabilities, Chinese language models, local teams and experience with the operational constraints of the domestic market.

For Apple, China remains a strategic market as well as a field of intense competition. The iPhone faces local manufacturers such as Huawei, Xiaomi, Oppo and Vivo, which have multiplied announcements around smartphones equipped with AI functions. The prolonged absence of Apple Intelligence in the country could become a commercial and symbolic disadvantage at a time when manufacturers are making embedded AI an argument for device upgrades.

The Apple-Alibaba matter thus shows that AI is not deployed according to a uniform global logic. The same manufacturer can present a common platform while having to modify its models, partners, infrastructure and sometimes its user experience depending on the territory. Apple's ambition to tightly control hardware and software integration encounters a political and industrial reality here: in China, market access requires far deeper localization than translating an interface.

What The Verge reports about Alibaba's role

According to The Verge, Apple reportedly trained its own AI model for China with Alibaba's help. The wording is important: this would not necessarily be a simple integration of an existing Alibaba chatbot into the iPhone, nor Apple's abandonment of its own work on foundation models. The scenario mentioned is that of local adaptation, with the Chinese group's participation in training or preparing a system designed to meet the needs of the Chinese market.

The precise technical details of this cooperation are not publicly documented in the reported information. In particular, it has not been established, in the wording repeated by The Verge, which datasets would be used, where training would have been conducted, what share of the model would be developed by Apple or Alibaba, or how requests would be distributed among the device, Apple's cloud and local infrastructure. Yet these points matter enormously, both for regulatory compliance and for the privacy promises associated with Apple Intelligence.

Apple has built its communications around the idea that personal AI should be aware of the user's context while limiting exposure of their data. For requests run on its servers, the company highlights Private Cloud Compute, a system in which data is not supposed to be accessible to Apple and must not be retained after processing. Apple also says that security researchers can inspect the software running on these servers in order to verify the stated guarantees.

Yet turning to a Chinese partner would necessarily raise questions about the practical application of these principles in the country. Apple has not publicly detailed a possible Chinese architecture for Apple Intelligence. It would therefore be premature to state that Private Cloud Compute would be replicated identically, entrusted to Alibaba, hosted in a specific configuration or replaced for certain uses. The only solid conclusion to draw from the reported information is that the Chinese version would require a distinct local solution from the one deployed elsewhere.

Alibaba's exact role warrants the same caution. The group is capable of providing several building blocks: language models, linguistic and cultural adaptation, cloud infrastructure, provision of services compliant with local rules, or assistance with validation procedures. But the article by The Verge does not allow these possibilities to be turned into a list of confirmed features. The cooperation could cover several of these levels, or focus on a limited part of the process.

This nuance distinguishes the Chinese matter from the agreement announced between Apple and OpenAI. In the latter case, Apple presented a visible integration: Siri can offer to send certain requests to ChatGPT, and ChatGPT can be used in writing tools. Apple named its partner, described the consent requested from the user and presented the experience during its conference. For Alibaba, the reported information instead concerns the industrial groundwork needed to make Apple's AI acceptable and usable in a local market.

This difference reveals two partnership models. With OpenAI, Apple supplements its capabilities through an actor whose service is known to the international public. With Alibaba, Apple appears to have to localize the very foundation of its offering. The issue is not merely to provide a more relevant answer to a question: it is to build a value chain compatible with an environment in which model availability, data hosting and service moderation follow specific rules.

Before Alibaba's name emerged in recent reports, several media outlets had reported that Apple was exploring different options in China. Baidu, in particular, had been cited as a potential partner for AI functions intended for Apple devices in the country. That information had already shown that Apple could not automatically transpose its partnership with OpenAI or rely solely on its in-house models for the Chinese market. The reported use of Alibaba fits into this search for a credible local intermediary.

Neither information about previous discussions nor the fact that an actor is cited as a partner is enough, however, to describe a final product. Apple has not announced a definitive public timetable for the arrival of Apple Intelligence in mainland China, nor has it detailed a set of localized features or a business model specific to the country. Any projection about the launch date, the devices concerned or distribution arrangements would remain speculative.

The fact that there is talk of a “custom” model is nevertheless consistent with the nature of language models. Their quality depends on training, alignment, language, reference corpora, connected tools and the safety policies governing their responses. An AI that is genuinely useful in simplified Chinese, integrated with locally used services and compliant with regulatory requirements cannot be obtained through a simple translation layer applied to an assistant designed for US English.

A negotiation between technological sovereignty and user experience

The partnership reported by The Verge highlights a tension that accompanies all major platforms: companies seek to offer a consistent experience worldwide, while states impose territorial rules on data, content, digital services and now AI models. Apple is particularly exposed to this tension because its economic model is based precisely on vertical integration. The company designs devices, operating systems, a significant share of processors, services and, with Apple Intelligence, an AI layer deeply tied to this whole.

In the Apple universe, AI does not take the form only of a website or standalone application. It must be able to rephrase text in Mail, summarize notifications, find information in personal data, help organize content or improve interactions with Siri. This integration increases the system's potential usefulness, but it also increases the sensitivity of the information processed. Calendars, messages, emails, photos, documents and user habits may be involved in a request.

Apple's response consists of distributing computation. The lighter models and certain tasks run on the device, which limits data transfers. Requests requiring more power can be assigned to Apple's private cloud. Finally, third-party models may intervene in certain cases. This strategy makes technical and commercial sense: it allows Apple to retain control over the experience while avoiding having to build a single giant model for every need.

In China, every layer of this architecture can become a subject of negotiation. The embedded model must work in the local language and for local uses. Cloud services must comply with rules applicable in the country. Any external providers must be authorized and capable of operating locally. Generated responses must meet the requirements imposed on providers of AI services accessible to the public. Finally, the brand must explain to its customers what does or does not change in its privacy guarantees.

This situation does not concern Apple alone. Microsoft has a longstanding presence in China, but its products and services also face limits, adaptations and changes in availability there. Google has long had a very restricted presence in Chinese consumer search. OpenAI does not officially make ChatGPT available in mainland China. US groups must therefore balance market access, control over their technologies, legal security and protection of their intellectual property.

Chinese companies also face a form of fragmentation, but from a different angle. They have local models, consumer platforms and considerable domestic markets. On the other hand, their access to certain cutting-edge foreign technologies, particularly advanced computing components, is constrained by US export restrictions. Sino-US rivalry is therefore not limited to competition between AI applications: it affects chips, servers, development tools, investment, cloud infrastructure and the markets where each model can be distributed.

Alibaba occupies a unique position in this landscape. The group is not merely an e-commerce platform. With Alibaba Cloud, it is also a major player in Chinese cloud computing, while Qwen has become an identified family of language models in local competition. The group has released open versions of certain Qwen models, which has helped give it visibility among developers. Its proximity to the Chinese technology ecosystem makes it a logical partner for a foreign company seeking to adapt a sophisticated offering to the local market.

That said, Apple's interest should not be read as recognition of a total inability to build its own models. Apple introduced its Apple Foundation Models and published research describing its approaches for Apple Intelligence. The company has its own AI teams, its own chips and infrastructure designed for its services. Turning to a local partner would primarily address a localization and deployment requirement, not necessarily a lack of fundamental expertise.

The Chinese case could even reinforce Apple's multiple-model logic. The company has already shown that it does not view Apple Intelligence as a product based on a single external provider: Apple's models, ChatGPT and local processing coexist in its presentation. A Chinese version involving Alibaba would extend this approach, but with one essential difference: partner selection would be driven as much by territory as by model performance.

This territorialization of AI can affect the iPhone's uniformity. Two users owning the same device could have access to different functions, different models, different answers and different levels of integration depending on their country. Apple already applies regional differences to its services, app stores and certain functions related to local laws. Generative AI, however, gives this reality a new scale, because it becomes a cross-cutting layer of the operating system.

Chinese competition makes the wait more costly for Apple

The absence of Apple Intelligence in China does not occur in a competitive vacuum. Chinese manufacturers have made AI a visible component of their new smartphone generations. Huawei, Xiaomi, Oppo, Vivo and Honor have all communicated about assistants, editing functions, transcription, summarization or search based on AI. The names, actual capabilities and availability conditions differ from one manufacturer to another, but the market dynamic is clear: AI has become a differentiating criterion in the premium segment and beyond.

Huawei is a particularly significant competitor for Apple in China. After being subjected to US sanctions, the group has managed to return strongly to the high-end smartphone segment with its recent devices. Its HarmonyOS ecosystem and Xiaoyi assistant are part of a strategy in which hardware, software and local services are closely linked. Apple therefore faces a rival that can highlight complete local integration, without depending on an international rollout or an agreement with a US provider.

This competition makes the localization of Apple Intelligence more urgent. Apple has long benefited in China from the iPhone's strength, its brand image and its ecosystem. But when competitors present devices capable of performing AI tasks in Chinese, integrating with local services or promoting their domestic models, the absence of equivalent functions can weigh on product perception. It is no longer enough to market the same iPhone as in the rest of the world; it is also necessary to offer a software experience that appears complete compared with local standards.

The difficulty for Apple lies in the fact that AI functions are not independent of the ecosystem. Voice and text assistants are more useful when they can act with the most used applications: messaging, payment services, e-commerce, mapping, delivery, travel, social networks or content platforms. Yet a significant share of these services is specific to China. Relevant adaptation therefore requires not only a good Chinese model, but also an understanding of usage patterns and, potentially, connections to the local environment.

Apple already has a considerable industrial and commercial presence in China. The group has long worked with Chinese manufacturing partners and sells its devices in the local market through several channels. But the iPhone production chain does not solve the problem of distributing AI services. Industrial relationships, data hosting, language models and cloud services are governed by different frameworks. The Alibaba matter shows that Apple's hardware strategy does not automatically guarantee its lead in intelligent services.

There is also an issue of pace. Apple chose a cautious rollout of Apple Intelligence, with gradual availability, compatible devices limited to the most recent models and to certain iPads and Macs equipped with Apple Silicon chips, as well as expanded language support step by step. This caution allows it to better control quality, server capacity and risks related to models. On the other hand, it creates a window during which competitors can occupy marketing ground and acclimate users to their own assistants.

Apple does not usually communicate iPhone sales figures by country in its financial results. It would therefore be risky to quantify precisely the potential commercial effect of Apple Intelligence's absence in China. The economic mechanism is nevertheless identifiable: if AI becomes a standard expectation in the premium category, a delay in its local availability can reduce the iPhone's perceived advantage and complicate customer retention.

The partnership with Alibaba could also alter the balance of power among Chinese AI champions. A collaboration with Apple would be a powerful industrial signal for Alibaba Cloud and Qwen, as it would associate the group with one of the world's most influential hardware platforms. But no implication regarding exclusivity, financial volume, revenue sharing or the use of Alibaba models in other countries should be inferred from the available information. China's AI market remains highly competitive, notably with Baidu, Tencent, ByteDance and other actors developing their own models and services.

The comparison with Samsung is also instructive. Samsung launched Galaxy AI in 2024, relying in particular on its own capabilities and on Google technologies for certain functions. But the availability and behavior of AI functions also vary by country, language and local agreement. The Apple-Alibaba case confirms that major smartphone brands can no longer regard AI as an identical layer everywhere: the product is becoming an assembly of global components and regional services.

In this context, Apple must protect two promises that are sometimes difficult to reconcile. The first is that of an integrated, simple and consistent experience, which is one of the main attractions of its ecosystem. The second is sufficient local compliance for functions to be genuinely available in China. Cooperation with Alibaba could provide a path between these two objectives, but it would also expose Apple to questions about the consistency of its technical principles across markets.

For France and Europe, the precedent of variable-geometry AI

The Chinese matter is of direct interest to French and European users, companies and regulators, even though Alibaba is not presented as an Apple Intelligence partner in Europe. It highlights a reality that the European market already knows: the availability of an AI function increasingly depends on regulation, compliance obligations and the distribution choices made by platforms.

Apple has in fact illustrated this reality in Europe with certain functions related to Apple Intelligence. In 2024, the company had announced that several features would not be available in the European Union at the planned time, citing uncertainties created by the Digital Markets Act. Apple then announced the arrival of Apple Intelligence in the European Union from April 2025, on iPhone and iPad, with the languages supported at that time. The European case is obviously not comparable to the Chinese framework, but it shows that Apple already adapts its schedules according to jurisdictions.

France and the European Union pursue a different logic from China. The European AI Act aims to establish a risk-based framework, with progressive obligations for certain AI systems and general-purpose models. The Digital Markets Act, for its part, addresses the conduct of major platforms designated as gatekeepers. These texts do not require foreign companies to enter into a partnership with a European champion in order to offer AI. They can, however, influence interoperability, transparency, risk management and conditions of market access.

For French consumers, the main lesson is that a global product name does not guarantee an identical global experience. Apple Intelligence, Gemini, Copilot, ChatGPT or assistants offered by smartphone manufacturers may have different functions, languages, integrations and limitations depending on the country. This diversity is not always visible when purchasing a device. Yet it becomes essential when AI is presented as a major product feature.

French companies deploying AI tools themselves face a similar constraint. They must examine where data is processed, contractual conditions, security guarantees, cloud providers and compliance with European rules. Apple's example in China is on a much larger scale, but it reminds us that choosing a model is never merely a performance decision. It is also a choice of infrastructure, governance and legal responsibility.

For European AI actors, the reported cooperation between Apple and Alibaba reinforces a strategic question: can Europe have local providers strong enough to become natural partners when global platforms must comply with the European market? France has companies such as Mistral AI, while major European groups, telecom operators, cloud providers and public institutions are seeking to develop computing capabilities and sovereign offerings. But Europe does not replicate the Chinese model: its market is governed by common rules, with strong attention to competition and data protection.

It would therefore be incorrect to present the Apple-Alibaba partnership as a model that the European Union would seek to copy. It does, however, demonstrate that a local provider of models and cloud services can acquire strategic importance when a major platform must adapt its services to a territory. In the coming years, companies able to offer multilingual models, compliant infrastructure and clear security guarantees could become decisive intermediaries between global technology giants and local requirements.

Language is also a central issue. French, like other European languages, requires high-performing models, specific evaluations and a good understanding of administrative, professional and cultural contexts. Apple has gradually expanded the languages compatible with Apple Intelligence, but language support does not summarize an assistant's quality. Users also expect reliable understanding, the ability to produce suitable texts and integration with services actually used in their country.

Data protection will probably remain one of the major competitive criteria in the European market. Apple bases part of its argument on local processing and Private Cloud Compute. Other providers highlight different approaches: cloud processing, open models deployed within a company, more compact on-device models or hybrid architectures. The multiplication of local partnerships, like the one mentioned with Alibaba, could make these architectures even more heterogeneous. Regulators and business customers will then demand more transparency about the data transmitted, the models used and each party's responsibilities.

Toward a global map of models and partners

The reported rapprochement between Apple and Alibaba could foreshadow a lasting evolution of the industry: major platforms will not necessarily deploy a single model on a planetary scale. They may maintain a common brand, interface and principles while using distinct models, clouds, partners and security policies depending on the region. AI would then become less of a uniform global product than a federation of local systems connected to the same platform.

For Apple, this evolution represents both an opportunity and a risk. The opportunity is to make Apple Intelligence available in markets where a purely American approach would be insufficient or impossible. The risk is diluting control of the experience, multiplying dependencies and having to explain why certain technical promises do not apply in exactly the same way everywhere. Apple's historical strength has been to reduce complexity for users; regulated and territorialized AI, by contrast, tends to shift that complexity into the infrastructure.

The Chinese case will be watched particularly closely because it brings together all the factors of this transformation: a colossal smartphone market, local AI giants, a demanding regulatory framework, geopolitical tensions with the United States and users already accustomed to highly integrated digital services. If Apple succeeds in launching a local version of Apple Intelligence with Alibaba's help, the result will not only show whether the iPhone can compete better with Huawei or Xiaomi. It will also provide a full-scale test of an American platform's ability to adapt its AI without relinquishing control of its ecosystem.

The trajectory will depend on questions that remain open: obtaining the necessary approvals, the exact definition of Alibaba's role, guarantees concerning data processing, the quality of Chinese models, integration with local services and the reception by users. As long as Apple does not publicly detail its arrangement for China, the reported partnership should not be presented as a completed launch. Rather, it is a strategic signal about the direction taken by the group.

For the French and European market, the lesson is broader. AI competition will not be fought solely between the largest models or the most visible assistants. It will also be fought in the ability to deploy these systems under different rules, forge reliable alliances and make data flows understandable. The association mentioned between Apple and Alibaba is a reminder that technological sovereignty is not only a question of companies' nationality: it depends on practical control over models, cloud computing, chips, data and the conditions of access to each market.

The next phase of AI could thus see regional versions multiply for products consumers thought were universal. In this shifting map, Apple seeks to preserve the unity of its brand while accepting deep adaptation to China. The choice of Alibaba, as reported by The Verge, illustrates a reality set to endure: for global giants, the spread of AI will depend less on a uniform conquest of markets than on the ability to negotiate, territory by territory, the technical and political conditions of its existence.