By | MarketCap List, Author | Jia Lele, Editor | Jixin

On July 8, 2026, Apple Technology Development (Shanghai) Co., Ltd. filed for registration of its "Apple Intelligence" generative AI service. On July 15, the Cyberspace Administration of China announced the registration details of seven on-device generative AI services for mobile platforms, with "Apple Intelligence" officially listed.

This marks the removal of the final regulatory hurdle for the China-exclusive Apple Intelligence, which had been in preparation for 1.5 years.

On the day the news broke, Alibaba confirmed that Qwen would be integrated as the underlying AI capability within Apple Intelligence, covering all devices across iOS, iPadOS, macOS, and visionOS in China, providing services including text and image understanding, content generation, and multi-turn dialogue. The capital market reacted swiftly, with Alibaba’s U.S.-traded shares rising over 7% at one point.

With over 200 million existing iPhones in China, Qwen—acting as the foundational AI engine—will be implicitly invoked by default. What does this mean for Qwen, for Alibaba, and for the entire large model industry? As AI becomes an implicit, system-level capability, how will the rules of the game be rewritten across the sector?

One: What Has Qwen Gained?

In February 2025, at the World Government Summit held in Dubai, UAE, Alibaba Chairman Joe Tsai publicly confirmed that Alibaba was collaborating with Apple to provide localized AI capabilities enabling Apple Intelligence to enter the Chinese market.

Tsai revealed at the time that Apple, seeking to enter the Chinese market, had engaged with multiple Chinese enterprises before ultimately selecting Alibaba as its domestic AI partner.

The significance of this collaboration extends far beyond simply securing a major client for Alibaba.

For Qwen, this partnership arrived at a pivotal moment.

Currently, consumer-facing competition in China's large model industry has become fiercely intense: on one hand, Douyin's Dabao is rapidly scaling through sheer volume; on the other, WeChat Agent, backed by its 1.2 billion monthly active users ecosystem, is poised to launch. Under this fragmented landscape, continuing down the path of “users actively downloading AI apps” means engaging in direct, head-to-head combat within already saturated app stores.

Apple’s native integration, however, offers a strategic bypass of this congested channel—granting Qwen access to an entirely different competitive dimension.

Previously, whether deployed within internal Alibaba business applications or accessed via the Tongyi Qianwen App, user engagement required deliberate choice: users actively sought out AI tools.

But with integration into Apple Intelligence, Qwen faces a fundamentally different distribution logic. It is no longer merely a model downloaded by users or called by developers—it now has the potential to become the default AI capability embedded beneath the operating system.

This shift redefines the value metrics of large model competition: from technical benchmarks like parameter scale, inference performance, and open-source ecosystem maturity, to real-world engineering robustness and ecosystem integration indicators.

Based on public disclosures, the collaboration employs a hybrid edge-cloud architecture. On-device lightweight AI computation is powered by Apple’s proprietary silicon chips, ensuring low-latency response and enhanced privacy protection. For cloud-intensive tasks such as long-text processing, complex logical dialogues, and multimodal generation, Alibaba Qwen provides the necessary computational resources. All user-related data is stored exclusively on servers located within mainland China, compliant with regulatory requirements.

According to multiple tech media outlets citing informed sources, Qwen is responsible for “the majority of text generation, summarization, visual-text understanding, and content creation capabilities.” Meanwhile, specific functionalities—including camera-based object recognition, landmark search, image editing, voice wake-up, and Siri’s Chinese language optimization—are handled by Baidu.

For Qwen, the true value of this partnership lies not in short-term increases in API call volume, but in gaining a massive-scale terminal validation.

The large model industry is entering a new phase. Previously, enterprises could validate model strength through benchmark rankings, evaluation tests, and open-source downloads. But once integrated into a smartphone operating system, models must now pass a more rigorous standard: Is the model stable enough? Can it adapt to real-world user demands? Can it sustain performance under complex, dynamic scenarios? And can it meet global-tier consumer electronics companies’ stringent requirements for security, privacy, and reliability?

What Apple’s ecosystem offers is precisely such a high-stakes validation environment.

This deal evokes the early trajectory of Android over a decade ago. Android initially succeeded by offering an open-source OS and free licensing, binding global smartphone manufacturers and capturing the gateway to the mobile internet era, then monetizing through app stores and advertising ecosystems.

While Qwen currently does not control any operating system, both cases share a common goal: becoming the foundational capability behind the next-generation computing platform—so essential that users have no choice but to adopt it.

Two: Commercialization Pathways – Speculation

As noted earlier, the announcement impacted Alibaba’s stock price on the same day. However, neither Apple nor Alibaba disclosed specific commercial terms. Analyst reports suggest several possible revenue streams for Alibaba’s income contribution from this collaboration.

The first layer is fixed annual licensing fees—direct and highly predictable revenue. After the China-exclusive Apple Intelligence launches, Apple will require system-level authorization to use Qwen’s models. Market references to Apple’s prior $1 billion annual licensing fee paid to Google suggest this could represent a substantial fixed income stream.

The second layer involves pay-per-token usage of compute resources. Analysts expect this segment to exceed fixed licensing fees in scale.

Due to regulatory mandates, on-device generative AI inference and user interaction data for Apple Intelligence in China must be processed and stored on servers within the country. Apple lacks its own large-scale AI infrastructure in China, so it procures all cloud computing power for complex AI tasks from Alibaba Cloud’s Lingjun Intelligent Computing Cluster.

This means every complex AI request initiated by a user corresponds directly to token consumption on Alibaba Cloud.

IDC data estimates China’s installed base of iPhones ranges between 220 million and 250 million units. As AI features become standard across devices, user invocation frequency is expected to rise steadily—making this revenue stream more sustainable than fixed licensing fees.

The third layer is e-commerce transaction revenue sharing—a unique income source exclusive to Alibaba.

China-exclusive Apple Intelligence includes dedicated traffic diversion rights tied to Alibaba’s ecosystem. In scenarios such as shopping-related Q&A or image-based product searches, the system generates structured product cards that trigger instant opening of Taobao App, directly entering the transaction pipeline. Apple earns a commission based on GMV generated from these transactions.

This revenue stream does not depend on AI API call volume but on actual conversion rates. Its ceiling is determined by real purchase behavior driven by AI interactions. If Apple Intelligence later introduces a subscription model, Alibaba may also secure future revenue-sharing from subscriptions.

The first three layers essentially represent monetization of three distinct commercial assets: technology licensing, computing infrastructure, and user transaction scenarios.

Additionally, there is potential for a fourth layer: service and operational experience delivery. If Alibaba Cloud participates in the deployment and operations of Apple’s AI system in China through technical outsourcing or joint operations, it effectively converts its deep engineering expertise and cloud operations capabilities into direct service revenue. This form of income is independent of API volume or e-commerce conversion rate, instead reflecting a direct valuation of Alibaba Cloud’s long-term industry service capabilities.

When these projected revenues will materialize in financial statements can be gauged from industry precedents. OEMs like Huawei and Xiaomi typically take between 1.5 to 3 months from regulatory filing completion to official launch of on-device AI features.

Based on this collaboration, analysts have revised upward their forecasts for Alibaba Cloud’s revenue. Guotou Securities International raised its FY27 revenue forecast by 4%, projecting a 47% growth rate for the Cloud Intelligence business. Credit Suisse expects AI-related revenue contributions to reach 50% of external revenue within the next 12 months, driven by strong MaaS (Model-as-a-Service) and AI performance. In the MaaS segment, ARR (Annual Recurring Revenue) is projected to hit $1.5 billion by Q2 2026—advancing toward a $4.5 billion target by year-end.

Provided that Apple Intelligence sees high-frequency usage among Chinese users.

Three: Large Model Competition Enters a New Phase

In the diffusion of technology, widespread adoption often depends less on a standalone product’s strength than on its ability to integrate seamlessly into users’ existing habits and tool environments—becoming an invisible, assumed component.

Electricity is a prime example: early on, whoever controlled large-scale power generation and long-distance transmission held industry dominance. Today, no one questions where electricity comes from—only whether appliances function properly. Similarly, cloud computing evolved from visible infrastructure debates around servers, VMs, and platforms to becoming deeply embedded behind applications, with few users pondering which data center hosts their online documents or video conference streams.

Large models are undergoing a similar transformation—but with higher difficulty.

Different from electricity and cloud computing, large models exhibit greater variability in performance. Users tolerate errors differently: a misstep from Dabao or Qwen is seen as a “tool limitation,” whereas a failure from Siri is blamed on the “phone itself.”

This means system-level AI leaves no room for trial and error. The prerequisite for invisibility is consistent, high-quality output behind the scenes. Thus, entering the system layer does not mark the end of competition—it signals the beginning of a higher-barrier phase.

This seamless integration imposes multifaceted demands on model capabilities.

First is stability: system-level AI must handle billions of interactions across hundreds of millions of users; even rare errors get amplified. Second is response speed: users may wait for an AI app to generate a lengthy response, but they won’t accept prolonged delays from a mobile assistant.

Specifically for Apple’s Apple Intelligence, this includes cross-model orchestration capability.

A collaborative model architecture—where each provider handles distinct functional modules—allows leveraging strengths across players, yet introduces new engineering challenges: the user experience must feel unified, not like a patchwork of disparate models.

When a user poses a request, the backend may involve speech recognition, visual understanding, text generation, search, and task execution—but the final output must present as a continuous, natural interaction. The system must determine which model handles each subtask and how results from multiple models are stitched together.

These experiential nuances directly influence how frequently Chinese users engage with Apple Intelligence.

Thus, future competition among large model providers at the system level will no longer be solely about algorithmic superiority—but also about engineering excellence, ecosystem coordination, and infrastructure robustness.

The Apple-Qwen collaboration may signal a turning point in industry dynamics. More accurately, large model competition is bifurcating into two tracks.

One track operates at the system level, competing for “invisible invocation.” Examples include Qwen’s integration into Apple Intelligence and Dabao’s embedding in Nubia devices—both experiments in achieving scale without owning direct consumer access, instead piggybacking on super ecosystems like smartphones and automobiles.

The other track operates at the application level, competing for “user-driven selection.”

System-level AI inevitably squeezes the space for thin-client AI apps—such as those offering only basic chat interfaces or generic model access. Yet it cannot fully displace applications requiring deep reasoning, complex operations, or domain-specific expertise, nor can it replace services embedded within workflows or private data loops.

The future AI device is likely to follow a hybrid model: a system-level AI (master orchestrator) + countless application-layer AI agents (specialized executors).

Ultimately, whether at the system or application layer, the competition is a battle between “mediocrity” and “irreplaceability.” Only those who either become indispensable utilities—like water, electricity, or gas—to users, or who establish unassailable moats in specialized application scenarios, will survive.

Source: MarketCap List