The major transformation in the smartphone industry: Apple bows to Google, and Doubao is besieged by the ecosystem.
On January 12, 2026, Apple and Google jointly announced a multi - year AI cooperation agreement: Apple will use Google's Gemini model to provide underlying support for Siri. According to Bloomberg, Apple plans to pay about $1 billion annually for this.
In the past two decades, Apple's core competitiveness has been built on the vertical integration of software and hardware - it designs its own chips, writes its own operating systems, and controls its own application ecosystem. However, in the field of AI, which is considered to define the next decade, Apple has entrusted its most core "brain" to the creator of Android.
Moreover, Apple is not the only mobile phone manufacturer making such a choice: Samsung has its self - developed Gauss, and Huawei has its self - developed Pangu. However, their flagship phones are now using external AI capabilities. Looking globally, no mobile phone manufacturer has developed an AI large - model comparable to Google Gemini, ChatGPT, ByteDance's Doubao, or Alibaba's Qianwen.
This is not the failure of a single company but a structural dilemma of the entire industry. However, the deeper problem lies not in "who will provide AI" but in "in what form AI should enter the mobile phone ecosystem". Just over a month ago, ByteDance's Doubao attempted to let AI directly perform cross - application operations - helping users compare prices, place orders, and make payments without opening any apps. As a result, several major Chinese core applications collectively blocked this function within a few days.
One is accepted, and the other is strangled - the difference does not lie in how smart the AI is but in whose interests it touches. This is the real problem behind Apple's $1 - billion order:
The most valuable ability of AI may be precisely the one it is least allowed to exert.
01 It's Difficult for Mobile Phone Manufacturers to Build Top - Tier AI on Their Own
Apple has gone through a rather inglorious process to reach this point.
At the WWDC in June 2024, Apple high - profilely launched Apple Intelligence and the new Siri. However, the vision soon collided with reality. In March 2025, according to Bloomberg, Apple's AI head, Robby Walker, described the delay as "ugly and embarrassing" at an internal all - hands meeting - internal tests showed that the new Siri could only correctly handle 67% to 80% of user requests, with one failure every three to five interactions.
In the process of looking for external solutions, Apple also tested OpenAI and Anthropic. According to multiple media reports, Anthropic performed better in technical tests, but its asking price exceeded $1.5 billion annually. Eventually, Apple chose Google, with which it already had a deep business relationship - the latter pays Apple about $20 billion annually to maintain its position as the default search engine in Safari.
The situation in China is similar. As the mobile phone manufacturer with the strongest technological self - sufficiency ability in China, Huawei has integrated the Pangu large - model into its Xiaoyi assistant in the HarmonyOS. However, after DeepSeek became extremely popular in early 2025, Huawei announced on February 5 that Xiaoyi would access DeepSeek - R1. Honor followed three days later, and OPPO also announced integration during the same period. The enthusiasm of each manufacturer for self - development began to cool down, and accessing third - party models became the mainstream choice.
Why are mobile phone manufacturers collectively absent? The fundamental reason is the difference in strategic priorities.
Google and Microsoft invest nearly $100 billion in AI annually because AI is directly related to their core businesses - Google's search advertising and Microsoft's cloud computing are both at risk of being disrupted by AI. Not investing means waiting for death. OpenAI is expected to lose about $9 billion in 2025 and will not break even until 2029 - 2030, but Microsoft is willing to bear this because it is the core engine for the growth of its Azure. The situation of mobile phone manufacturers is completely different: AI is an additive function for them, not a matter of life - and - death strategy. No board of directors of a mobile phone company will approve spending five to ten years and hundreds of billions of dollars on a technological direction that has little to do with selling mobile phones.
The difference in strategic priorities is directly reflected in the magnitude of resource investment. In 2025, the total AI - related capital expenditure of Amazon, Google, Microsoft, and Meta was about $370 billion; Apple's total capital expenditure during the same period was only $12.7 billion, less than one - thirtieth of the former. Apple has more than $200 billion in cash on hand but chooses to pay Google $1 billion annually to purchase AI capabilities, which in itself explains the problem.
Money is just the surface. The deeper gap lies in data and infrastructure. Google has the world's largest search engine, Gmail, and YouTube. Zuckerberg said in an earnings conference call that there are "tens of billions of publicly shared pictures and hundreds of millions of publicly available videos" on Facebook and Instagram. What do mobile phone manufacturers have? Local photos, contact lists, and app usage habits - highly privacy - sensitive and low in text content, which are not suitable for training general large - models at all.
In terms of computing power, Google has been deploying its self - developed TPU chips since 2015, and they have now reached the seventh generation; the cloud computing businesses of Microsoft and Amazon have allowed them to accumulate the world's largest - scale data centers. The servers of mobile phone manufacturers only need to support the App Store and cloud synchronization and are not designed for training trillion - parameter models - even if they start building now, it will be too late in terms of time.
And we haven't even considered talent. According to foreign media reports, it is estimated that there are less than 1,000 AI scientists globally with the ability to build top - tier large - models. In July 2025, Ruoming Pang, the head of Apple's basic models, was poached by Meta with a compensation package of over $200 million, and Apple didn't even try to match it - this figure far exceeds the compensation of any Apple executive except Cook. When the price of talent competition has exceeded the compensation system of mobile phone companies, this race is not on the same track from the start.
Falling behind comprehensively in five dimensions - talent, capital, data, computing power, and strategy - this is why no mobile phone manufacturer in the world has developed a top - tier AI large - model. Since they can't develop it themselves, relying on third - parties is the only rational choice.
But the question arises: In what form should external AI enter the mobile phone ecosystem?
02 How Does AI Divide the Pie?
To understand the position of AI in the mobile phone ecosystem, we first need to understand the existing business models of mobile phone manufacturers.
Taking Xiaomi's financial report for the third quarter of 2025 as an example, the gross profit margin of its smartphone business was only 11.1%, while that of its Internet services business was as high as 76.9%. The revenue sources of the latter include app store commissions, pre - installed app fees, advertising distribution, and search promotion slots - in essence, mobile phone manufacturers are not selling hardware profits but the entrance for users to access apps. The premise for the operation of the entire system is that users must actively "open" apps to complete tasks, and the entrance to this action is in the hands of mobile phone manufacturers.
Apple's AI cooperation with Google has not shaken this system. According to Bloomberg, Apple pays Google about $1 billion annually to obtain the underlying capabilities of Gemini for Siri. In exchange, Gemini runs on Apple's private cloud servers in a white - label form. The user interface is Apple's, and the data remains in Apple's hands. Users are unaware of Google's existence.
The reason why Google is willing to provide top - tier AI capabilities at a relatively low price is that it is not buying revenue but strategic security - keeping OpenAI and Anthropic out of the Apple ecosystem while maintaining the Safari default search agreement worth about $20 billion annually.
For this arrangement to work, there is a key premise: The most advanced AI models in the United States are closed - source. OpenAI's GPT, Anthropic's Claude, and Google's Gemini do not make their model weights public. Mobile phone manufacturers must pay, negotiate, and sign agreements if they want to use them. Therefore, AI capabilities have become scarce resources that can be locked in through exclusive cooperation.
The situation in China is completely different. DeepSeek's R1 is open - source under the MIT license. The global download volume of Alibaba's Qwen series has exceeded 300 million, and there are more than 100,000 derivative models. Baidu open - sourced its Wenxin large - model in June 2025 - almost all of China's most advanced large - models are available for free. This means that any manufacturer can obtain AI capabilities of the same level. When AI is no longer scarce, it cannot become a moat for differentiation. The focus of competition has shifted from "who can obtain AI" to "who can use AI better".
But the problem follows: If everyone can obtain AI capabilities, who will maintain the existing interest distribution?
The Doubao AI phone in December 2025 provided an answer - not government regulation, not industry agreements, but the immune response of the ecosystem itself. This phone, jointly launched by ByteDance and Nubia, has the ability to automatically perform cross - application tasks. When users say "Help me order the cheapest milk tea", the AI can compare prices, place orders, and make payments across multiple food - delivery platforms. The first batch of devices was sold out within 24 hours, and the second - hand price was once speculated to be ten times the original price. However, the frenzy only lasted for one day. WeChat, Taobao, Meituan, and Alipay collectively blocked the cross - application operations of the AI.
The logic of the blockage is clear: If users no longer open apps, it means that no one will see the splash screen ads, and there will be no opportunity to display the information - flow recommendations. The user paths carefully designed by the platforms will be completely bypassed; when the AI compares prices across platforms, the profit margins of the platforms that rely on information asymmetry will be directly compressed.
For mobile phone manufacturers, the threat also exists - pre - installed fees, app store promotion slots, and negative first - screen ads are essentially fees paid by apps to be seen by users. If the AI bypasses these entrances and directly calls services, the value of mobile phone manufacturers as traffic intermediaries will be undermined. Therefore, it is not difficult to understand why ByteDance finally cooperated with a marginal manufacturer like Nubia.
The fates of the two models reveal the real boundaries of AI in the mobile phone ecosystem. The Apple - Google model is accepted because it does not change the existing value distribution: AI makes Siri smarter, but it does not make choices for users, does not bypass the app store, and does not break information asymmetry. The Doubao model is strangled because it tries to redefine the entrance itself with AI. China's mainstream manufacturers have seen this clearly - Huawei, Xiaomi, OPPO, Vivo, and Honor have all accessed third - party large - models, but they only use them in safe scenarios such as Q&A, translation, and copywriting generation. None of them dares to copy Doubao's cross - application model.
Currently, a preliminary equilibrium state has been formed in the mobile phone ecosystem: AI can become smarter, but it cannot seize the entrance. In the United States, this equilibrium is locked through exclusive agreements for closed - source models; in China, this equilibrium is maintained through the collective actions of the app ecosystem.
Although the forms are different, the logic is the same - the value of AI is strictly limited to "making the existing experience better" rather than "redefining the experience itself".
03 How Long Can the Equilibrium Last?
The current equilibrium - AI provides capabilities but does not seize the entrance - seems stable, but it is based on a premise: AI is not powerful enough. Once this premise changes, the equilibrium will be shaken.
Who will be the variable? The possibility of mobile phone manufacturers is extremely low. They are both beneficiaries of the existing interest pattern and it is difficult for them to make effective breakthroughs in self - developed large - models. The 76.9% gross profit margin of the Internet services revenue depends on the premise that "users must open apps". Actively overthrowing it is equivalent to self - sabotage. The experience of Nubia and Doubao has proven that the adventures of marginal players will only attract the collective backlash of the ecosystem.
More likely variables come from two directions:
One is the AI agent within super apps. WeChat is already a self - sufficient ecosystem - mini - programs cover food delivery, ride - hailing, shopping, and government services, and the payment closed - loop is also within the system. If WeChat deploys an AI agent within this ecosystem, when users say "Help me book a restaurant for tonight", the AI can compare prices, make reservations, and make payments among mini - programs without leaving WeChat. This model does not violate the interests of external apps because it never intends to let users leave; instead, it will further strengthen the moat of the super app - the more users rely on the AI agent, the more they will be inseparable from this ecosystem.
The other is the gradual migration of user habits. The blockage of Doubao is aimed at cross - application operations, but it cannot stop users from completing more and more tasks within AI applications - checking flight tickets, comparing prices, writing weekly reports, and summarizing documents. This does not trigger any blocking mechanism but is gradually eroding the usage time of apps. Once user habits are formed, traditional apps will either actively access AI or watch as new players willing to cooperate take their market share. When the quantitative change accumulates to a critical point, apps will find themselves bypassed - not because AI has seized the entrance but because users no longer need that entrance.
Apple spent ten years and invested more than $10 billion in researching and developing the Apple Car but finally announced its abandonment in 2024. One of the reasons is that if it can only achieve L2 - level assisted driving instead of full - self - driving, the Apple Car will only be "a better car", and it cannot justify why Apple should build cars.
Mobile phone AI faces a similar dilemma - If AI can only perform safe functions such as Q&A, translation, and photo editing, it will only be a smarter assistant and will not change anyone's fate. The real differentiation lies in whether the AI agent can redefine the interaction mode between people and services, but this is precisely what the current ecosystem least allows to happen.
The current equilibrium is essentially an exchange of technological capabilities for ecosystem peace. How long this peace can last depends on when AI becomes powerful enough that users no longer need the action of "opening apps". By then, the definition of the entrance will be rewritten, and all current arrangements will become obsolete.
Disclaimer: This article is for learning and communication purposes only and does not constitute investment advice.
This article is from the WeChat official account "Lying - Flat Index". Author: Sister Lying. Republished by 36Kr with permission.