Global AI companies have burned through hundreds of billions, and in the end, they all end up competing to "work for" Apple.

The global AI arms race is fought with real money, and it’s a game of life and death.

OpenAI, Google, Anthropic—pick any one of them, and the computational budget, talent salaries, and model training costs poured into it over the past two years have all been calculated in the hundreds of billions of dollars.

In contrast, Apple has seemed like an unconventional “outsider” for years, slow to produce a decent large model response, and was even mocked for falling behind in the AI wave. But this does not mean Apple has been idle in the face of the AI wave’s impact.

Just recently, according to a report from Bloomberg, iOS 27 will introduce a new mechanism called “Extensions,” allowing users to integrate third-party AIs like Google Gemini and Anthropic Claude into Siri through the settings panel, enabling requests to be made directly from Siri just like calling ChatGPT now.

In summary, Siri is set to become the “gateway” for AI applications.

In the beta system, there is an internal prompt that says: “Extensions allow agents from installed apps to work with Siri, the Siri app, and other features on your devices.”

This mechanism fundamentally differs from the previous exclusive collaboration model with ChatGPT. ChatGPT was a special version negotiated by Apple, requiring time and effort for each individual agreement.

Extensions are a set of standard protocols that any qualifying AI application can access. In other words, Siri’s intelligence will be outsourced to more third-party AIs.

Everything is ready, just waiting to collect rent.

In fact, Apple Intelligence was not designed as a single large model system from the beginning, but as a “three-layer decoupled” architecture:

The first layer is the device side. A lightweight model with about 3 billion parameters runs locally on the iPhone, handling text summarization, quick conversations, and privacy-sensitive operations. This model has undergone specialized 2-bit quantization awareness training, optimized specifically for Apple chips, providing fast speeds, no internet connection, and data that never leaves the device.

The second layer is the private cloud. Apple’s self-developed server-side large model uses a PT-MoE (Parallel Track Mixture of Experts) architecture, handling complex reasoning and long texts. Apple claims this cloud computing environment uses end-to-end encryption, with data processed for requests only and not stored; the operating environment can even accept third-party expert reviews.

This system called “Private Cloud Compute” is what gives Apple confidence in privacy issues.

The third layer is the third-party model layer. Third-party AIs connect through standardized interfaces, becoming the “replaceable engines” of the entire system. The key to this architecture is that the three layers are decoupled from each other.

The upper layer’s tool calls, permission management, and user interface are completely indifferent to which model is running at the lower level. What users perceive is “Siri got this done,” while who is doing the work behind the scenes is determined by the system.

To open up to third-party AIs, there is another key prerequisite.

There are several pathways for third-party AIs to enter the Siri ecosystem. One is the App Intents framework, which allows developers to expose their application’s features to the system AI for cross-application calls; the other is the MCP (Model Context Protocol), which Apple is turning into system-level support, integrating it with the App Intents framework.

MCP is not something invented by Apple; it is part of a set of open protocols forming in the AI industry, with Anthropic being one of the early proponents. Apple’s choice to support MCP at the system level means that any AI service complying with this protocol can, in theory, be uniformly scheduled by Siri.

For developers, Apple has also opened up the Foundation Models Framework.

This is an SDK that can natively connect to the core models of Apple Intelligence using Swift, reportedly callable with just three lines of code. It includes guided generation, restricted tool calls, and LoRA adapter fine-tuning.

In short, while developers can build on Apple’s foundation, the rules of the foundation are still set by Apple.

OpenAI’s exclusive honeymoon is over.

Recall that ChatGPT’s entry into Siri was a highlight at the 2024 WWDC when Apple Intelligence was announced.

However, internally at Apple, the collaboration was not entirely harmonious.

Apple’s former AI head, John Giannandrea, was initially reserved about OpenAI, questioning the company’s long-term competitiveness and preferring a partnership with Google.

Ultimately, Apple conducted an “AI competition,” and ChatGPT emerged as the “best choice at the time.” In hindsight, the phrase “at the time” carries significant weight.

Now, with the launch of the Extensions mechanism, ChatGPT’s exclusive status is naturally dissolving.

In an ideal scenario, users will be able to choose which AI answers their questions in the future—Gemini, Claude, Grok, Copilot, Perplexity—as long as they have the corresponding app installed, Siri will be able to call them. ChatGPT is still there, but it is no longer the only option.

It’s worth noting that on the day the news broke, Google’s stock price also dropped in response. This reaction seems contradictory at first glance; isn’t Gemini about to enter Siri? Shouldn’t Google be happy? But a closer look reveals that the open Siri Extensions are two different matters from Apple’s previous push to “rebuild Siri’s underlying layer using the Gemini model.”

The latter involves Apple switching to Gemini’s model capabilities at the technical level, which would be a significant deal; the former is about integrating Gemini’s user-facing application into Siri, placing it on the same level as Claude and ChatGPT.

There is also a deeper concern. According to previously disclosed information, Apple retains access to the weights of the Gemini model in its collaboration with Google to build its foundational model capabilities. In other words, while Apple invites Gemini in as an “equal tenant,” it is simultaneously using Google’s technical resources to build its own foundation.

When Apple levels the playing field for all AIs through Extensions, Gemini loses its priority for the “exclusive underlying layer of Siri,” and its negotiation leverage is significantly diluted. This explains the drop in value.

Moreover, this opening is part of a broader Siri reconstruction plan by Apple, which includes: an independent Siri application, a new Siri interface, integration of Siri with Spotlight search, and new entry points like “Ask Siri” and “Write with Siri.”

These related features are scheduled for official release at this year’s WWDC but may still change or be delayed.

Slow is fast.

After Bloomberg’s report, netizens on X platform commented: “Siri has now become the world’s first AI but doesn’t run any models itself and hasn’t spent a penny on training. Claude, ChatGPT, and Gemini have to come in to reach Siri’s 2.5 billion users. Then Apple collects a 30% fee from subscriptions through the App Store. That’s how AI runs on iOS with Apple hardly lifting a finger.”

There is no exaggeration in this logic.

Apple’s platform strategy has always been like this: controlling the operating environment and distribution channels, allowing others to compete on it while collecting tolls. The App Store once harvested the mobile internet in such a manner, and now this logic is being replayed in the AI industry.

Currently, Apple is already sharing revenue through ChatGPT subscriptions: when users pay for ChatGPT Plus on their iPhones, the App Store payment system is involved, and Apple receives a share. Once the Extensions mechanism is rolled out, every paid AI service like Gemini and Claude entering Siri becomes a new revenue-sharing channel.

The more intense the AI battle, the more subscriptions users take, and Apple benefits as the passive recipient.

In contrast, looking back over the past two years, the competitive script of AI companies has been highly homogeneous: releasing new models, refreshing benchmarks, seizing media headlines, a slightly higher score leads to a higher funding valuation, and then releasing the next new model.

But this “involution” in underlying computing power and scoring obscures the industry’s biggest pain point: the vast majority of AI companies lack a direct entry point to ordinary consumers.

Regardless of how impressive the capabilities of large models are, they can ultimately only exist as independent apps or web pages. Companies must invest heavily in customer acquisition costs to compete for users, and the switching cost for users between different models is extremely low, with no loyalty at all.

Apple’s ability to escape this money-burning cycle comes from its firm grip on “distribution” and “hardware.”

Apple’s total number of active devices worldwide has exceeded 2.5 billion, a distribution scale that no other AI company can replicate on its own. With just one system update, Apple can present a completely new AI ecosystem to all iPhone users the next time they wake Siri.

This distribution efficiency is unmatched by any AI startup. Channels determine exposure, exposure determines habits, and habits determine conversion to paid.

The ultimate battle in AI subscription competition may not be decided at the model level, but rather harvested at the distribution level. The AI service that first secures the iPhone entryway does not gain a one-time traffic bonus but rather a continuous binding of user habits.

From the perspective of AI companies, integrating Siri Extensions is no longer a matter of choice. Once competitors enter, not doing so means actively giving up the largest distribution channel. The true moat for AI remains closest to the user.

This is an unmistakable scheme by Apple, not requiring anyone to join, yet everyone has no reason to refuse.