Apple’s approach to artificial intelligence has always been framed around privacy, and the marketing language — on-device processing, Private Cloud Compute — has done its job well enough that most users assume everything happens locally. The reality is a carefully engineered split, and understanding where the line falls explains a great deal about which features work in airplane mode, which need a signal, and why some of the most impressive capabilities arrived later than promised.
What the Silicon Handles
Apple’s Neural Engine has become genuinely capable, and the workloads that live on it share a profile: small models, frequent invocation, latency-sensitive, privacy-critical. Text prediction, on-device dictation, photo subject recognition, background removal, live translation of short phrases, and the notification summaries that appear on your lock screen all run without leaving the device. They work offline, consume modest battery, and never transmit content — which is exactly why Apple leads with them.
The constraint is model size. A phone can comfortably run models in the low billions of parameters. That is sufficient for classification and short-form generation and insufficient for the extended reasoning, long-context comprehension, and image or video generation that define the current frontier.
Where the Cloud Takes Over
For heavier requests, Apple routes to its Private Cloud Compute infrastructure — Apple silicon in Apple data centres, with the stated guarantee that requests are not stored and not accessible even to Apple. For a category of requests beyond that, the system offers to hand off to third-party models with explicit user permission. This is the architecture behind the features people find most impressive and the reason some of them display a network indicator before answering.
It is also the pattern the rest of the industry has settled into, minus Apple’s privacy engineering. Third-party apps on the App Store follow the same split, handling small tasks locally and reaching frontier models over the network — usually through a multi-model AI API that exposes many providers behind one endpoint, so a developer can route cheap requests to economical models and hard ones to premium models without maintaining separate integrations.
What This Means for Buying Decisions
Three practical implications for anyone weighing an upgrade. First, RAM and Neural Engine generation determine how much AI stays local — which affects offline capability and battery, not raw feature availability. Second, features described as “beta” or region-limited are almost always cloud features, because server capacity is provisioned per market. Third, treat the marketing phrase “AI features” as a category, not a spec: ask specifically which ones work without a connection, because that is the part your hardware actually determines.
The Trajectory
Each silicon generation moves another workload from cloud to device. Transcription made that journey, then live translation, and summarisation is halfway there. But the frontier keeps advancing too, so the split will persist even as the line moves. The honest summary for Apple users: your device is becoming a very capable AI computer for the everyday tasks, and a very good client for the hard ones — and the second half of that sentence is not going away soon.
There is one more wrinkle worth knowing about, and it shows up in battery complaints more than in reviews. On-device inference is efficient per request but not free, and features that quietly run in the background — photo analysis while charging, notification summarisation, on-device search indexing — do measurable work overnight. If your battery health looks worse after a major software update, that is usually the cause rather than a defect, and it typically settles once the initial indexing pass completes. Users who want to verify can watch the battery breakdown in Settings for a few days before concluding anything.
