Apple's new Siri uses Apple Foundation Models across the iPhone and Private Cloud Compute, while Google and NVIDIA play important but more specific roles behind the architecture. Apple collaborated with Google on its new model family and runs its most powerful cloud model on NVIDIA GPUs in Google Cloud. That is considerably more complicated than saying Siri runs on Gemini.

Apple's relationship with Google is one of the most important parts of the new Siri story, but it is also one of the easiest to misunderstand. Apple has confirmed that its third-generation Apple Foundation Models were custom-built in collaboration with Google. It has also confirmed that its most capable server model, AFM 3 Cloud Pro, runs on NVIDIA GPUs inside Google Cloud. What Apple has not announced is a simple replacement of Siri with the Gemini assistant. Apple describes Siri AI as powered by Apple Intelligence and its next-generation Apple Foundation Models. (Apple Machine Learning Research)

The distinction matters because iOS 27 reveals something much larger than a conventional partnership in which one company plugs another company's chatbot into its operating system. Apple has constructed a layered artificial-intelligence architecture that can move between models running directly on Apple devices, models running on Apple's Private Cloud Compute infrastructure, and its most demanding model running on outside hyperscale hardware that Apple has brought inside the security architecture of Private Cloud Compute. The new Siri sits above those layers and connects them to personal context, onscreen information, semantic search and actions inside applications. (Apple Developer)

Did Apple Replace Siri With Google Gemini?

No. That description is too simple for the architecture Apple has disclosed.

Apple announced Siri AI at WWDC 2026 as a new version of Siri powered by Apple Intelligence. Apple's own developer documentation describes Apple Intelligence as being powered by next-generation Apple Foundation Models. Its machine-learning research team separately identifies five models in the third-generation AFM family and says those models were custom-built in collaboration with Google. (Apple)

Those statements can all be true simultaneously.

Google is a significant partner in the development of the new model family. Google Cloud is also part of the infrastructure supporting Apple's most powerful server model. Neither fact means that every request made to Siri is simply forwarded to the Gemini service.

Apple's five-model architecture makes that clear.

The smallest member of the family is AFM 3 Core, a 3-billion-parameter dense model designed for on-device intelligence. Above it sits AFM 3 Core Advanced, a much larger 20-billion-parameter sparse model that Apple says activates only approximately 1 billion to 4 billion parameters for a given request. Apple stores the full model weights in NAND flash rather than requiring the entire model to remain resident in active memory. That approach allows more capable Apple hardware to run a substantially larger model without carrying the conventional memory burden of keeping all 20 billion parameters active at once. (Apple Machine Learning Research)

The server side adds three more models. AFM 3 Cloud is Apple's general server model. ADM 3 Cloud handles image generation and editing. AFM 3 Cloud Pro is the most capable model in the family and is intended for demanding tasks including complex reasoning and agentic tool use. (Apple Machine Learning Research)

That last model is where the infrastructure story changes dramatically.

Why Is Apple Running AI on NVIDIA GPUs in Google Cloud?

Apple designed AFM 3 Core, Core Advanced, AFM 3 Cloud and ADM 3 Cloud around Apple silicon. AFM 3 Cloud Pro is different. Apple says it worked with Google and NVIDIA to extend Private Cloud Compute to NVIDIA GPUs hosted inside Google Cloud while preserving the privacy guarantees associated with Private Cloud Compute. (Apple Machine Learning Research)

That is an important architectural development for Apple.

Private Cloud Compute was designed around a simple privacy proposition: when an Apple Intelligence request becomes too computationally demanding to handle locally, the device can send the necessary information to protected server infrastructure without turning that request into an ordinary cloud-AI transaction.

AFM 3 Cloud Pro expands that idea beyond infrastructure physically built around Apple's own server silicon. Apple is effectively trying to make its security model portable enough that outside hyperscale computing resources can participate without giving the outside infrastructure provider ordinary access to the user's request.

The result changes the old distinction between on-device AI and cloud AI.

Apple now has multiple computational destinations available within one intelligence system. Some work can remain on the device. More demanding work can use server-side Apple Foundation Models through Private Cloud Compute. The most computationally demanding AFM tier can use NVIDIA acceleration inside Google Cloud while remaining subject to Apple's Private Cloud Compute architecture. (Apple Machine Learning Research)

This is why reducing the arrangement to "Apple uses Gemini" misses much of what is technically important.

Google is not merely supplying a chatbot. NVIDIA is not merely supplying GPUs. Apple is attempting to build an orchestration and privacy layer that can span hardware and infrastructure controlled by different companies while maintaining an Apple-controlled user experience.

So What Does Gemini Actually Do?

This is where precision becomes especially important.

Apple explicitly says its third-generation Foundation Models were custom-built in collaboration with Google. Apple's public technical material establishes Google's involvement in model development and confirms that Google Cloud hosts the NVIDIA GPU infrastructure used for AFM 3 Cloud Pro. (Apple Machine Learning Research)

Independent technical analysis following WWDC identifies Gemini-family technology as part of the training and refinement relationship rather than treating Gemini itself as the live runtime assistant answering every Siri request.

That narrower distinction is also consistent with technical reporting after WWDC. Apple executives described the models as jointly developed with Google while emphasizing that Gemini does not simply take control of the Siri request pathway. (TechRadar)

The safest conclusion is therefore not that Google is absent from Siri AI, because it clearly is not. Nor is the evidence best summarized by saying that Siri has simply become Gemini.

Apple has built a family of Apple Foundation Models in collaboration with Google, integrated those models into Apple Intelligence, and created a runtime architecture that includes both Apple infrastructure and, for AFM 3 Cloud Pro, NVIDIA hardware hosted in Google Cloud.

That is the relationship Apple has actually disclosed.

Why Does Siri Need More Than One AI Model?

Different computing problems have different costs.

A model running directly on an iPhone offers enormous advantages for latency and privacy because a request may not need to leave the device at all. But a phone has strict limits on memory, power consumption, thermal performance and computational capacity.

Large server models have the opposite advantage. They can use far greater computing resources and support more demanding reasoning, but they introduce network dependence and the privacy complications associated with sending information away from the device.

Apple is trying to use both environments.

AFM 3 Core provides an efficient local model. Core Advanced expands what high-end Apple silicon can accomplish locally. Private Cloud Compute provides access to larger server models. Cloud Pro adds another computational tier for particularly demanding work. (Apple Machine Learning Research)

This turns Apple Intelligence into an architecture rather than a single model.

That distinction will become increasingly important as consumers compare AI systems. Asking which model powers a phone may eventually become less useful than asking which models are available, where each model runs, how requests are routed, what information each model can access, and what security boundary governs the transition between them.

What Makes the New Siri Different From a Chatbot?

The model architecture is only half of the story.

Apple is also changing the relationship between Siri and applications.

In iOS 27, App Intents allows developers to expose application content and actions to Siri AI through structures that Apple Intelligence can understand. Entity schemas can contribute application content to Spotlight's semantic index. Intent schemas describe actions that can be performed. View Annotations can associate visible interface elements with entities so that Siri can understand what a user is referring to on the screen. (Apple Developer)

This changes the interaction model.

Traditional smartphone software generally requires a person to understand an application's interface, navigate to the correct screen and manipulate the controls required to complete an action.

The architecture Apple is building allows language itself to become another route into those actions.

A user can express an intention. Apple Intelligence can interpret the request and its context. The system can identify relevant entities and available actions. Siri can then use the application's exposed capabilities to help perform the task.

Apple's developer documentation says App Intents makes app content and actions available through natural language and allows Apple Intelligence to use semantic Spotlight search even when someone describes content imprecisely. (Apple Developer)

That is much closer to an operating-system intelligence layer than simply placing a conversational bot on top of iOS.

Is Apple Building an Agentic Operating System?

Apple does not publicly describe iOS 27 itself as an agent operating system, so that phrase should be treated as interpretation rather than Apple terminology.

The underlying capabilities, however, increasingly overlap with what the AI industry describes as agentic computing.

Apple itself describes AFM 3 Cloud Pro as supporting agentic tool use. Its Foundation Models documentation discusses agents and skills as abstractions developers can build. Its current iOS developer material explicitly promotes building agentic app experiences. (Apple Machine Learning Research)

Siri AI adds another layer by combining language understanding with personal context, onscreen awareness and application actions. (Apple)

That does not mean an autonomous AI has unrestricted control over an iPhone.

It means Apple is creating more of the infrastructure required for software to translate natural-language intentions into structured actions across applications.

For developers, that may ultimately prove more consequential than the visible Siri redesign.

Can Developers Use Gemini or Claude Inside Apple's New Framework?

Yes.

Apple has opened the Foundation Models framework beyond its own models.

Its iOS 27 developer documentation says applications can work with Apple Foundation Models as well as cloud models such as Claude and Gemini, or other providers that conform to Apple's Language Model protocol. (Apple Developer)

That is another reason the new architecture should not be understood as a simple Apple-versus-Google story.

At the system level, Apple is building its own intelligence architecture. At the developer level, it is simultaneously creating an abstraction that can accommodate multiple model providers.

That could allow an application to use an Apple on-device model for one task, a larger server model for another, or a third-party provider where appropriate, while developers work through increasingly standardized Apple frameworks.

The long-term significance may be less about which company won the Siri contract and more about Apple positioning iOS as the orchestration layer through which different kinds of intelligence reach applications and users.

Why Does AFM 3 Core Advanced Matter?

The 20-billion-parameter Core Advanced model may be one of the most consequential pieces of Apple's strategy because it demonstrates how the company is attacking one of the fundamental constraints of local AI: memory.

Large models traditionally need their weights readily available in memory during inference. That creates a severe problem on consumer devices.

Apple's Instruction-Following Pruning architecture takes a different approach. The full model is stored in NAND flash, while only a subset of approximately 1 billion to 4 billion parameters is activated for a request. Apple says routing is determined per prompt, avoiding the need to continuously swap model weights for each generated token. (Apple Machine Learning Research)

The importance extends beyond Siri.

If increasingly capable models can operate locally without requiring their complete parameter set to occupy active memory, Apple gains another reason to tightly coordinate its model architecture, storage system, memory design and silicon.

That makes Apple Intelligence inseparable from Apple's hardware strategy.

The AI system is being designed around the device rather than merely delivered to it.

Does Google See Siri Requests?

Apple says its Private Cloud Compute architecture is designed so that user data sent for computation is not stored or made accessible to Apple, and it says the same privacy guarantees extend to AFM 3 Cloud Pro running on NVIDIA GPUs in Google Cloud. (Apple Machine Learning Research)

That claim is more technically significant than simply saying the servers are secure.

Apple's objective is to prevent privileged runtime access and make properties of the computing environment externally verifiable. The company is therefore trying to preserve the Private Cloud Compute security boundary even when the underlying physical infrastructure is supplied by outside companies.

That does not mean every Apple Intelligence request goes to Google Cloud.

On-device models exist specifically to perform work locally. Other server-side AFM models run through Private Cloud Compute. Google Cloud enters the disclosed runtime architecture specifically at the AFM 3 Cloud Pro tier. (Apple Machine Learning Research)

The architecture is therefore better understood as tiered computing than as a universal cloud handoff.

Why This Matters Beyond Siri

For years, the central AI competition in smartphones has often been framed around which company has the best assistant or the largest model.

iOS 27 suggests another competition is developing underneath that one.

The question is becoming who controls the orchestration layer.

Apple controls the operating system, its silicon, its application frameworks, the personal-context layer, the semantic index, App Intents, the interface through which Siri interacts with users and the privacy architecture governing server requests. It can simultaneously collaborate with Google on model development, use NVIDIA's computational hardware, operate infrastructure in Google Cloud and allow developers to connect other models such as Claude without surrendering the operating-system layer itself. (Apple Developer)

That is a substantially different strategy from trying to build every component alone.

It also explains why the simple question "Is Siri powered by Gemini?" produces an unsatisfying answer.

Google matters. Gemini technology matters. NVIDIA matters. Apple Foundation Models matter. Private Cloud Compute matters.

But they do not all perform the same job.

The real story of Siri AI is how Apple is assembling those pieces into a single system while attempting to keep the user, the operating system and the security boundary under Apple control.

That architecture, rather than the name of any one model, may be the most important Apple Intelligence development of 2026.

Fact Summary

Did Apple replace Siri with Gemini? No. Apple describes Siri AI as powered by Apple Intelligence and next-generation Apple Foundation Models. Apple developed its third-generation model family in collaboration with Google, but the disclosed architecture is not simply the Gemini assistant replacing Siri. (Apple)

Does Google participate in the new Apple Intelligence architecture? Yes. Apple explicitly says its five third-generation Foundation Models were custom-built in collaboration with Google. (Apple Machine Learning Research)

Does Apple use Google Cloud? Yes. AFM 3 Cloud Pro, Apple's most capable server-side model, runs on NVIDIA GPUs hosted in Google Cloud through an extension of Private Cloud Compute. (Apple Machine Learning Research)

Does every Siri request go to Google Cloud? No. Apple has on-device models and additional server models. Google Cloud is specifically identified as infrastructure for AFM 3 Cloud Pro. (Apple Machine Learning Research)

What models run directly on Apple devices? Apple identifies AFM 3 Core, a 3-billion-parameter dense model, and AFM 3 Core Advanced, a 20-billion-parameter sparse model that activates approximately 1 billion to 4 billion parameters depending on the request. (Apple Machine Learning Research)

Can developers use other AI models in iOS 27? Yes. Apple's Foundation Models framework supports providers conforming to its Language Model protocol, and Apple explicitly names Claude and Gemini as examples. (Apple Developer)

Can Siri interact with applications through natural language? Apple says App Intents exposes application content and actions to Siri AI, while semantic indexing and schemas allow the system to understand entities and available actions. (Apple Developer)

Is iOS 27 an agent operating system? Apple does not use that description for iOS 27. It does, however, explicitly describe agentic tool use in AFM 3 Cloud Pro and provides developer tooling for agentic application experiences. (Apple Machine Learning Research)

Sources

Apple Machine Learning Research. Introducing the Third Generation of Apple's Foundation Models. June 8, 2026. Apple Machine Learning Research

Apple. WWDC26: Apple unveils next generation of Apple Intelligence, Siri AI, powerful parental controls, and an expansive set of software improvements. June 8, 2026. Apple Newsroom

Apple Developer. What's New in iOS 27. Apple Developer: iOS 27

Apple Developer. Apple Intelligence and Siri AI. Apple Developer: Siri AI and App Intents

Apple Developer. Foundation Models. Apple Foundation Models documentation