Almost two years. That is how long it took from the moment Apple promised that Apple Intelligence would come to China "subject to regulatory approval" to the moment it actually got that approval. On July 15, 2026, the Cyberspace Administration of China finally registered Apple Intelligence as an approved generative AI service - and the most interesting part is not the approval itself, but what is running under the hood. Apple Intelligence in China will not be powered by Apple's own models. It will run on Alibaba's Qwen, with Baidu handling additional features. I have been following the slow-motion collision between Western AI products and Chinese regulation for a while now, and this is the clearest picture yet of what "AI sovereignty" looks like in practice.
What exactly happened with Apple Intelligence in China
The facts first, because the story has been dragging on long enough that it is easy to lose track. The Cyberspace Administration of China (CAC), the country's internet regulator, listed Apple Intelligence among seven approved on-device generative AI services for smartphones. The other six are Huawei, Xiaomi, Samsung, OPPO, vivo and Nubia. Apple is the only company on that list that is not Chinese or Korean, and the only one whose AI brand is globally recognizable as a Western product.
The approval covers Apple Intelligence across iOS, iPadOS, macOS and visionOS. An Alibaba spokesperson confirmed that Qwen will power the features on all of those platforms for users in China, describing the scope as "text and image understanding and generation". Baidu confirmed separately that it is also working with Apple on features for Chinese users - most reporting points to Baidu handling parts of the stack like visual intelligence queries, though neither company published an exact split.
What we do not have is a launch date. The regulator's statement does not include one, Apple has not announced one, and given how long the approval took, I would not bet on "next week". But the hard part - the part that took 22 months - is done.
Why Apple needed Qwen at all
If you are wondering why the world's most valuable hardware company cannot just ship its own models in China, the answer is regulation, not technology. China requires generative AI services to be registered and approved by the CAC, and in practice that means the models answering user queries must comply with Chinese content rules, data localization requirements and filtering obligations. Apple's own foundation models, plus the ChatGPT integration it offers in other markets, were never going to clear that bar - OpenAI does not even operate in mainland China.
So Apple did what it has done before with iCloud (operated in China by a local partner) - it found domestic infrastructure. The reported shortlist over the past year included DeepSeek and ByteDance alongside Alibaba and Baidu. Alibaba won the core deal, and it is not hard to see why. Qwen is arguably the strongest open-model family in China and the Qwen ecosystem has become the default backbone for a huge share of Chinese AI products. I wrote recently about how Chinese open-weight models are already grinding through a huge share of tokens even in American APIs - inside China, their position is simply dominant.
There is a delicious irony here that I cannot skip: the same Qwen family that PrismML used to build Bonsai 27B, the 1-bit model that runs on an iPhone, is now officially the brain of the iPhone's assistant in China. Qwen is becoming the Android of language models: it shows up everywhere, in forms its creators never planned.
What Alibaba and Baidu actually get out of this
For Alibaba, this is a prestige win with real strategic weight. The market read it that way too - Alibaba's stock jumped about 4% on the news. Being the model behind Apple Intelligence means Qwen becomes the default AI experience for tens of millions of premium smartphone users, the exact demographic every Chinese AI lab wants. It also validates Alibaba's multi-year bet on Qwen as infrastructure rather than a consumer app.
Baidu's role is smaller but notable: after losing ground to Alibaba and DeepSeek in the model race, staying inside the Apple deal at all keeps Ernie relevant in the highest-profile AI integration in the country.
For Apple, the stakes are bigger than the AI feature list. Greater China generated 20.5 billion dollars in sales in Apple's Q2 2026, up 28% year over year, and Apple recently reclaimed the number two spot in China's smartphone market. That rebound happened without Apple Intelligence. Chinese competitors - Huawei, Xiaomi, vivo - have been shipping increasingly aggressive on-device AI for two years. An iPhone in Shanghai was, until now, the only flagship phone in the store with no meaningful AI assistant. That gap is about to close.
The bifurcation is now official
Here is the bigger picture that I keep coming back to. For two years people have been talking about a "bifurcated AI world" - a Western stack and a Chinese stack, each with its own models, rules and infrastructure. It was mostly an abstraction. This deal makes it concrete in the most visible consumer product on the planet.
Think about what an iPhone now represents. The same physical device, the same Apple Intelligence branding, the same UI - but in Cupertino's home market it talks to Apple's models and optionally ChatGPT, while in China it talks to Qwen and Ernie under CAC oversight. The product is global; the intelligence inside it is jurisdictional. I touched on the other side of this coin when writing about the US letting generative AI into Secret and Top Secret networks - both superpowers are pulling AI infrastructure inside their own borders, just with different tools.
A few consequences worth spelling out:
- Model choice is becoming a regulatory decision, not a technical one. Apple did not pick Qwen because it beat Apple's models on benchmarks. It picked Qwen because that was the price of admission to a market worth over 20 billion dollars a quarter.
- Chinese users get a different assistant, not a worse one. Qwen models are genuinely strong - depending on the benchmark, arguably stronger than Apple's own foundation models. The difference is in what topics the assistant will discuss and where the data lives, not raw capability.
- Every Western AI company now has a template. The Apple deal shows the only realistic path into China: partner with a domestic lab, put their approved model behind your brand, accept CAC registration. Expect others to study this playbook carefully.
- The "one product, many brains" architecture will spread. Apple has effectively built a model-agnostic assistant layer that can swap intelligence providers per jurisdiction. That is an engineering pattern, and other multinationals will copy it.
What I will be watching next
The approval leaves several open questions that will decide whether this is a footnote or a turning point. When does it actually ship - iOS 26.x this fall, or earlier as a point release? How will the feature set compare with the Western version - does Chinese Apple Intelligence get notification summaries, writing tools and visual intelligence at parity, or a trimmed list? And the question I find most interesting as a developer: what happens to the App Intents and on-device model APIs that third-party apps use - do Chinese developers build against Qwen-backed endpoints with the same interfaces?
There is also the unresolved tension with Washington. American lawmakers grumbled loudly when the Alibaba partnership was first reported in 2025. An official CAC approval with Alibaba and Baidu named as providers turns that from rumor into policy reality, and I doubt the reaction will be silence.
Summary: the deal everyone saw coming still matters
Nothing about this outcome is surprising - the Alibaba partnership leaked 17 months ago. But confirmation matters. As of July 15, 2026, Apple Intelligence in China is officially a Qwen product wearing an Apple interface, approved by the Chinese state alongside six other manufacturers. Apple gets to compete on AI in its second most important market. Alibaba gets the most prestigious model deployment in consumer tech. And the rest of us get the clearest demonstration yet that in 2026, the question "which model are you talking to" has a different answer depending on which country you are standing in.
Sources: TechCrunch, The Next Web.