The "laggard" Apple has begun to roll out its AI features on a limited basis.

The China-market version of Apple Intelligence is finally arriving.

Recently, I spotted a line of fine print in official press releases and at the bottom of the Apple Intelligence product page: some AI features that partially rely on server-side models will have daily usage limits, and additional usage permissions can be obtained via paid subscriptions in the future.

This marks the first time Apple has framed AI as a limited-quantity product, and it also means that after more than two years of delays, Apple Intelligence finally has a confirmed launch timeline. Apple's Executive Chairman Tim Cook recently revealed that the China-market Apple Intelligence will kick off testing at the end of October.

However, what he confirmed is only the "first-generation" Apple Intelligence, and the latest Siri AI will require more waiting.

Managing the Apple Intelligence "brain" via supply chain methodologies

At the same time, major AI giants are collectively hitting the brakes.

Public information shows that Microsoft has canceled hundreds of megawatts of data center leases; Meta cut around 8,000 positions in the first half of the year; at the Anthropic Investor Day event on October 14, attendees stated that its latest valuation stands at 1.5 trillion US dollars, a 25% discount from its previous peak of 2 trillion US dollars.

The root cause lies in the oversupply of AI. What the market and investors are truly worried about is that some tech companies have taken on excessive debt for AI infrastructure construction.

S&P estimates that the six major US hyperscale cloud providers — Amazon, Microsoft, Alphabet, Oracle, Meta and SpaceX — will collectively spend more than 7 trillion US dollars on data centers and other AI-related capital expenditures between 2025 and 2030, and will remain in a negative free cash flow state at least until next year.

Over the past two years, the valuation logic for AI has been "investment equals conviction": whoever pours money into GPUs and data centers will be rewarded with a high valuation by the capital market.

But the market's financing costs are rising, and revenue cannot maintain rapid growth forever. The capital market's focus has shifted from seeing who can afford to burn cash to seeing who generates more profits, and who has more stable cash flow on their books.

Apple, which has long been criticized for "falling behind in AI", has unexpectedly reaped benefits from this situation, picking up a number of cost-effective resources. According to estimates, its annual AI-related expenditure only ranges from 11 billion to 13 billion US dollars, while its competitors each have capital expenditures at the level of hundreds of billions of US dollars.

Some analysts therefore believe that due to its relatively low AI capital expenditure, Apple has been priced as a "safe-haven asset" in the tech stock sell-off. At the same time, strong demand for its first foldable smartphone has further boosted investor confidence.

This has driven its stock price up by roughly 25% in 2026, with its latest market capitalization hitting 5 trillion US dollars, only 148 billion US dollars behind NVIDIA. By contrast, the stock market performance of the other companies in the "Magnificent Seven" has been mixed, with Tesla, Microsoft and Meta all posting declines so far this year.

While competitors are looking for ways to invest in AI infrastructure, Apple is working to control AI supply: limited features, limited regional availability, and even limited daily token quotas.

The most direct action is procurement. Over the past decade, Tim Cook has built a robust supply chain system: for even a single screw, Apple retains at least two suppliers, and maintains deep ties with them, to ensure stable supply, reduce inventory, and secure bargaining power over the supply chain.

The same logic applies to AI. Although Apple has long been criticized for moving too slowly and lagging far behind peers, it still uses its best-in-class supply chain capabilities to manage the "AI brain", setting specifications and managing delivery for model vendors.

At the top of Apple Intelligence's supplier list is Google. Google released the third-generation Apple Foundation Models (AFM) at WWDC26, which includes a total of five models, some running locally on devices and others deployed on the cloud.

Among them, the most powerful cloud-side model Cloud Pro runs on Google's servers, with computing power supported by NVIDIA chips. This is the first time in Apple's history that it has moved computing to a third-party cloud for AI purposes.

Two partners have been finalized for the China region: Alibaba provides model capabilities, including its Tongyi Wanxiang and Qwen series; Baidu provides search and Siri-related capabilities.

Taking Alibaba as an example, it has passed at least three levels of checks: qualification review, adaptation testing, and boundary verification, before finally being included in Apple's supply chain. Among them, Apple has set strict rules for model vendors, such as which apps the model can access on the iPhone, what data it can read, and which chip it can run on.

But there is an ironclad rule in Apple's supply chain management: core components cannot be fully dependent on external procurement. Apple Intelligence has focused all its independent R&D efforts on the device side. Among the five models, the one Apple has put the most effort into is the device-side model, while the cloud-side model is more like leveraging Google's infrastructure to deliver Apple's own services.

For example, the flagship device-side model AFM 3 Core Advanced has 20 billion parameters, uses a sparse architecture, stores the main body of the model in flash memory, and only activates a small portion of parameters per request.

This is essentially using flash memory to replace RAM, a solution that neither Google nor OpenAI can directly replicate, as it requires deep integration with chips, which is Apple's core domain of expertise.

To paraphrase a line from popular short dramas: "With core components firmly in your own hands, you can outsource the AI brain without any anxiety."

Limiting access is a profitable business

Moreover, Apple has poached Amar Subramanya, former Vice President of AI at Microsoft, to head its model business. Subramanya spent 16 years at Google and previously led the Gemini team. Isn't this exactly: if you can't buy the best models, buy the people who build the best models?

As one comment puts it, Apple Intelligence's moat does not lie in the scale of computing power; its killer advantage remains the entry capabilities of its own software and hardware ecosystem, as well as the user data it holds.

Under Apple's precise tiered design, the experience of the free version and paid features are completely different. For example, real-time translation, writing tools, and notification summaries run locally on devices, cost Apple nothing, and have no usage limits; but features that require cloud-side models, such as object erasure and image expansion in photos, come with fixed usage quotas, and users have to pay for additional capacity once they exceed the limit.

This is Apple's long-standing routine: creating scarcity via quantity limits, and scarcity drives willingness to pay. For example, the free iCloud quota is 5GB, and users have to pay for more storage; Apple Music offers three months of free trial, after which it costs 10 RMB per month.

AI is another product placed on the shelves, behind which lies a hundred-billion-dollar business.

According to Apple's financial report, in the first fiscal quarter of FY2026 (ended December 27, 2025), Apple's total revenue reached 143.756 billion US dollars, with iPhone sales hitting 85.27 billion US dollars, accounting for nearly 60% of total revenue but only contributing a gross margin of 40.7%; software service revenue reached 30.01 billion US dollars, with a gross margin as high as 76.5%.

However, the Apple Tax (App Store commission), which accounts for the largest proportion of service revenue, is shrinking. This is because regions around the world are dissatisfied with its strict commission rates, and have started to impose price caps via regulatory policies.

Starting from March 15, 2026, the standard commission rate in mainland China has been reduced from 30% to 25%, and the rate for the Small Business Program and auto-renewing subscriptions after the first year has been cut to 12%; the US has taken stricter measures: in April 2025, a court once ruled in the Epic case that the external link commission should be set to zero.

Tim Cook and Luca Maestri Source: Internet

In my view, selling AI with usage limits adds a new revenue stream to the service business. This is a more certain new charging item than the Apple Tax: selling tokens is the standard model in the AI industry, with higher user acceptance, and Apple has far more flexibility in setting its own prices.

Apple has already started testing the waters: Apple Create Workshops cost 38 RMB per month, and the previously free Magic Fill feature in Numbers has been moved to the paid tier, with charges applied per use.

As the launch date of the China-market Apple Intelligence approaches, Tim Cook and Luca Maestri hope to sell more new hardware such as iPhones on the one hand, and on the other hand, they are more eager to quickly define AI as a product with controllable supply, and push it to users' devices.

Judging from current progress, Apple has "won" again. According to estimates by digital industry bloggers, in the seven days from October 5 to October 11, the cumulative sales of Apple's flagship iPhone 18 Pro series reached approximately 4.0809 million units, setting the fastest growth rate for Apple's flagship models after launch in recent years.

This is naturally the capital that allows Apple to still sit at the main table in the AI era.

Moreover, in this year of AI oversupply, it has proven itself with the AI strategy of "slower is faster": relying on the iPhone to drive more iPhone sales, even with limited AI feature supply, it can still sell at a high price.

This article is from the WeChat official account "Tang Chen's Notes", author: Tang Chen, authorized for release by 36Kr.