Mobile AI as a platform layer: Pixel 10 and iPhone 17 - on-device, private cloud, and trust boundaries

This article will discuss how the Pixel 10 and iPhone 17 route AI requests, what Tensor G5 and A19 provide, how Private AI Compute and Private Cloud Compute are structured, where the AI layer lives in the OS — and how all this changes for developers when AI becomes part of the interface rather than a separate library.

Introduction

Smartphone AI is quickly shifting from "in-app functions" to a platform layer. Three questions come to the forefront: where the request is executed, what data goes into the context, and who can see the result.

The availability of specific AI features almost always depends on language, region, and device model; some capabilities may be unavailable in certain countries or language configurations. We will discuss architecture — the trust boundary, routing, and isolation, but no one can guarantee that a specific feature will turn on "right now" in your region.

On-device and cloud: where the AI request goes

Both platforms use a hybrid scheme: some tasks are performed on the device, while others go to the cloud. The on-device option (inference — calculating the model's response based on input data on the device itself) usually covers quick scenarios and reduces the amount of data leaving the phone. The cloud provides access to larger models and longer context.

In Android, the local generative layer is often associated with the multimodal neural network Gemini Nano, which operates in AICore — a system service in Android for executing and updating foundation models on the device — and relies on hardware accelerators to reduce latency and maintain model relevance.

On the Pixel, this logic manifests as "system" gestures and functions on top of any application. For example, Circle to Search works as a gesture over the screen and allows searching within a highlighted fragment without switching apps. This is important not as a "feature of a specific model," but as an example of how AI is becoming part of the interface.

Some features of the Pixel 10 are built around proactive suggestions. Magic Cue raises actions based on context and operates "on-device and in the cloud" in an isolated environment.

When on-device layers are insufficient, Google connects to Private AI Compute. This is a cloud environment for processing sensitive requests, where privacy must be a technical property of execution, not just an access policy. Developers describe the idea as follows: data and processing results should remain accessible only to the user, and security is ensured through isolation and verifiability of execution. Implementation elements:

• a stack on proprietary tensor operation accelerators;

• hardware-isolated environments called Titanium Intelligence Enclaves;

• remote attestation: verification that code is executed within a trusted environment;

• encryption of the “device-cloud” channel.

On the iPhone 17, Apple Intelligence technology strives to solve tasks on the device, and for functions that require larger foundation models, it uses Private Cloud Compute, or PCC.

Tensor G5 and A19: What Exactly Accelerates AI

The Pixel 10 was released in four variants: Pixel 10, Pixel 10 Pro, Pixel 10 Pro XL, and Pixel 10 Pro Fold. The base level of performance in them is provided by the Tensor G5 chipset, manufactured using a 3nm process. This chipset — a tensor operation accelerator — handles a significant portion of inference: Google claims that it does so on average 34% faster than the central processor. This means developers are trying to execute more scenarios on the smartphone and less frequently rely on the cloud.

The iPhone 17 has a base chip — A19, also manufactured using a 3nm process. It contains updated display engine and ISP blocks, as well as the Apple Neural Engine, or NPU — a neural accelerator for machine learning operations that assists Apple Intelligence. Each GPU core in the smartphone has "neural accelerators" built-in to support the work of generative models.

iPhone 17 Pro and iPhone 17 Pro Max are equipped with the A19 Pro chipset featuring a 16-core Neural Engine, as well as a 6-core GPU with "neural accelerators" in each of them.

In other words, both companies are strengthening the local framework and adding specialized hardware blocks for generative models. At the same time, the entry point to the OS is different: in Android, AICore is allocated to a separate system layer, while in iOS, much of the logic is hidden in Apple Intelligence and in the integration of system applications.

Where the boundary of data privacy lies

In mobile AI, there are almost always three classes of data: input request, device context, and output. Each class can remain on the device or go to the cloud. Therefore, it is more important to focus on the boundaries of trust, an explicit threat model, and the verifiability of mechanisms rather than the thesis of "everything is local."

In Android, they are developing the Private Compute Core (in Google’s materials sometimes abbreviated as PCC, not to be confused with Apple PCC — Private Cloud Compute). This is an isolated environment within the OS for processing confidential data. It is designed to help users control when and how data can leave the trusted zone. Additionally, Android has Private Compute Services, which link the Private Compute Core to the cloud and are published as a separate component.

In Pixel 10, they added a secure coprocessor Titan M2. In AI scenarios, it provides a hardware root of trust, supporting verified boot and key security operations. However, the choice between on-device and cloud is primarily determined by the combination of AICore, Private Compute Core, and Private AI Compute.

Apple builds privacy around Private Cloud Compute. The company claims that personal data residing in PCC is inaccessible to anyone other than the user, including Apple itself — provided that the isolation and attestation mechanisms function correctly. Furthermore, the concept of “verifiable transparency” is used in PCC, allowing verification of PCC builds and trust chains.

That is, Google is moving part of the "device security model" to the cloud through Private AI Compute, attestation, and channel encryption, while Apple is moving the device security model to the cloud through PCC and verifiable builds. These are different implementations of one idea: if the request goes to the cloud, then the boundary of trust should remain technical, not contractual.

What all this changes for developers (and why it matters in fintech)

When AI becomes a layer of the OS, not only the set of APIs changes, but also the "attack surface," especially where the application processes personal information, financial data, and operations with heightened control and audit requirements.

First, "AI on top of the application" ceases to be an exception. Shell-level gestures, such as Circle to Search, allow initiating the processing of a screen fragment without switching applications.

Second, system tools for working with text and context begin to function where the user types, including third-party applications and websites. Visual intelligence technology helps obtain context about objects, places, and text around and on the smartphone screen.

Third, "private cloud" does not eliminate the question of boundaries: which classes of data can leave the device, and at what level is this controlled — by the platform and OS settings or by the application's logic? In practice, this means that when designing data flows, one has to consider not only their own backend but also platform AI contours, as well as their regional/language availability.

Integration into the OS: who manages the models

In Android, Gemini Nano runs in AICore and utilizes the smartphone's hardware accelerators. For mobile developers, this means there is a single entry point into the "system model" instead of multiple libraries and execution environments.

Google is developing the ML Kit GenAI API, which works "on top of AICore," uses Gemini Nano, and provides ready-to-use scenarios like summarization and text rewriting. Because applications use the same model, and requests are isolated at the AICore level, memory savings are achieved.

Apple has taken a different path. In 2025, the company provided developers access to the on-device foundation model Apple Intelligence, which has about 3 billion parameters. Apple emphasizes that this model is suitable for text tasks and is not intended for a "general knowledge chatbot."

Compromises in latency, context, and cloud cost

The on-device approach wins in latency and autonomy but is limited by model size and device resources. The cloud wins in model size and context length but introduces network latency and raises questions about trust boundaries.

Google describes Private AI Compute as a platform that provides all the advantages of cloud models Gemini with privacy at the execution level on the device while claiming on-device processing privacy. In practice, this means that part of the functionality gets a cloud "booster," while the data remains private.

iOS implements a similar compromise: Private Cloud Compute connects for features that require larger foundation models, while access to data is limited and there is an option for external verification. With this approach, the cloud framework requires trust in the claimed verifiability.

UX: when AI takes the initiative

The difference between iOS and Android most often manifests not in individual buttons but in when the system takes the initiative.

In the Pixel 10, there is an emphasis on the device's initiative. It responds to actions, suggests options, complements text, and shows hints before the user requests them. Here, AI is not a tool but part of the overall system logic — from the keyboard to the gallery. For example, Magic Cue connects data from different applications and offers users suggestions "when they are needed." Google emphasizes data isolation and user control.

In the iPhone 17, AI often manifests as a feature integrated into specific locations. This is most noticeable when working with text. For example, Live Translation works as a system function in Mail, Messages, FaceTime, and Phone, while Writing Tools are active in Mail, Notes, and Messages.

Why AI is Most Noticeable in Camera Work

The camera has become a convenient platform for comparing the AI approaches of both companies because it has the strictest requirements for latency, quality of work, energy consumption, and privacy.

The Pixel 10 turns the camera into an extension of AI. With Camera Coach, Face Unblur, and Auto Best Take, the system takes a series of shots and selects the best frame on its own. Camera Coach uses Gemini models and requests a network connection because some tips rely on cloud models. For instance, the system can suggest ideal settings for low-light shots. Best Take compiles group photos from several similar shots to ensure everyone looks great.

In the iPhone 17, the focus is on pipeline processing. The 48 MP Dual Fusion camera processes images carefully and almost imperceptibly. AI does not interfere with the frame but helps convey natural colors and details. In the Pro models, optical zoom has been increased to 8x, and more machine learning is used in pipeline processing to preserve details, reduce noise, and improve color.

Here, the difference is not about "better or worse." In the Pixel, some functions appear as suggestions and assistance before the shot. In the iPhone, part of the logic goes into computational photography and manifests already in the result.

Apple and Google's Position in Mobile AI at the Beginning of 2026

In January 2026, Apple and Google released a joint statement about a multi-year collaboration: the next version of Apple Foundation Models will rely on Gemini models and Google's cloud technology to support future Apple Intelligence features and a more personalized Siri.

In November 2025, Google launched Private AI Compute as a platform for cloud models with private boundaries, and cited the development of Magic Cue on Pixel as one of the first examples.

That is, by the beginning of 2026, both companies arrived at the same concept: the on-device layer remains fundamental, while the cloud adds stronger models. Trust boundaries, certification, and verifiability become critical.