Siri AI Under the Hood: Google Gemini Infrastructure & Private Cloud Compute

The Re-engineered Assistant

At WWDC 2026, Apple made a monumental shift in its artificial intelligence strategy by officially partnering with Google to leverage Gemini infrastructure for complex, high-tier reasoning. While Apple's proprietary on-device models handle local tasks (such as app actions, message summaries, and basic photo editing), queries requiring deep semantic knowledge, web retrieval, or multi-step logic are routed to the cloud. This routing is done transparently, with Siri AI determining the optimal destination for each user request.

To maintain Apple's commitment to user privacy, the company designed Private Cloud Compute (PCC). PCC is a custom-built hardware and software system running on Apple Silicon (M-series and custom server chips) inside Apple's data centers. This infrastructure ensures that user data is never stored, logged, or made accessible to Apple or third parties.

The Routing Pipeline

When a user interacts with Siri AI, the request enters a local orchestrator. The orchestrator determines the complexity of the query and routes it accordingly. Here is a breakdown of this routing logic:

On-Device Model

Tasks: App interactions, local search, text formatting, personal schedule lookups.
Latency: Very low (sub-100ms).
Privacy: Local sandboxed execution.

Private Cloud Compute (Apple Model)

Tasks: Summarizing long documents, complex cross-app workflows, advanced photo rendering.
Latency: Medium (200-500ms).
Privacy: Stateless PCC routing.

Public LLM Integration (Google Gemini)

Tasks: Open-domain knowledge queries, deep research, coding help, public data searches.
Latency: Variable (depends on network).
Privacy: Anonymized IP, transient data processing.

The decision tree can be visualized in the following diagram:

text

[Siri AI Request]
        |
        +---> [On-Device Orchestrator]
                    |
                    +---> Can run locally?
                    |          |
                    |          +---> (Yes) ---> [On-Device Model]
                    |
                    +---> (No) ---> Needs Web/Deep Knowledge?
                                       |
                                       +---> (No) ---> [Private Cloud Compute]
                                       |
                                       +---> (Yes) ---> [Anonymized Gemini Gateway]

Security and Verification in PCC

Unlike standard cloud APIs, PCC uses a unique operating system image that is cryptographically signed and published. This allows security researchers to verify the system's claims. Every production PCC node runs a hypervisor that enforces strict virtual machine isolation.

Cryptographic Security and Privacy

For developers, Siri AI offers a set of system APIs through App Intents. By declaring App Intents, developers expose their app's capabilities to Siri AI, allowing it to perform actions on behalf of the user.

Here is a conceptual example of a custom App Intent configured in Swift for a task manager:

swift

import AppIntents

struct CreateTaskIntent: AppIntent {
    static var title: LocalizedStringResource = "Create Task"
    static var description = IntentDescription("Creates a new task in the task manager.")

    @Parameter(title: "Title")
    var title: String

    @Parameter(title: "Due Date", default: nil)
    var dueDate: Date?

    func perform() async throws -> some IntentResult {
        // App-specific logic to save the task
        let task = TaskStore.shared.add(title: title, due: dueDate)
        return .result(value: task, dialog: "I've added '\(title)' to your tasks.")
    }
}

By providing these Intents, the application becomes an active participant in Siri AI's ecosystem, allowing users to coordinate multi-app flows seamlessly. The privacy constraints ensure that even when Siri AI accesses data from multiple apps to fulfill a request, that data is processed transiently and securely.

Secure Data Center servers