Every time you ask a recent phone or laptop to summarise a page, transcribe a meeting or rewrite a message, a routing decision has already been made behind the scenes: part of the work runs on a dedicated chip inside your device, and part of it travels to a server. Understanding where that boundary sits — and who is making promises about the half that leaves — matters far more than the marketing shorthand “on-device AI” suggests.
- The hardware angle: Neural processing units execute AI workloads far more efficiently than CPUs or GPUs, which is what makes battery-powered inference viable at all.
- The split is real: Apple, Google and Microsoft all keep short tasks over your own content local, then push heavier generative work to the cloud.
- Guarantees differ by vendor: a “private cloud” is a technical claim that can be audited — but it only covers first-party features, never third-party apps.
- You can test it yourself: airplane mode remains the simplest and most reliable on-device detector available to ordinary users.
Why a dedicated AI chip changes the battery math
An NPU is a processor built specifically for the arithmetic that AI models demand. According to Microsoft’s own support documentation, it “uses less energy and is far more efficient at AI tasks than a CPU or a GPU,” while freeing those two components to handle everything else. That efficiency is the precondition for running models on a device that has to survive a full day on a charge.
Microsoft has turned the specification into a commercial badge. Copilot+ PCs are Windows 11 machines whose NPU delivers “more than 40 trillion operations per second” — 40 TOPS — per the company’s official product page.
One caveat deserves stating plainly: none of the official documentation reviewed for this guide publishes battery-life figures broken down by AI task. “More efficient than the CPU” does not mean “free.” The reasonable reading is that an NPU beats the local alternative, but a feature running continuously in the background will still drain the pack.
Where the local/cloud boundary actually sits
The answer depends on the vendor and on the feature, yet all three major platforms follow the same pattern: resolve locally what can be resolved, and hand off whatever needs a larger model.
Apple. In its guidance on Apple Intelligence and privacy for the Mac, Apple explains that the system evaluates each request to decide whether it can be processed on-device, and that “in many cases” the model runs entirely locally. More demanding requests go to Private Cloud Compute (PCC), which runs larger models on servers powered by Apple silicon. The same documentation notes that PCC is not available in every region.
Google. On Android the local model is Gemini Nano, which per the Android developer documentation operates “without needing a network connection or sending data to the cloud.” It runs inside a system service called AICore. Google’s listed use cases are concrete: summarising, correcting short texts, rewriting messages, describing images and transcribing voice.
Microsoft. The Copilot+ PC page enumerates what works without internet: Recall (in preview), Windows Studio effects, real-time translation of live captions and super resolution in Photos all “run locally.” By contrast, Cocreator and generative fill in Paint, along with Image Creator in Photos, “require an internet connection.”
The practical rule that emerges: short tasks over your own content — summarising, correcting, transcribing — are the likeliest to stay on the device. Generating images or reasoning across long documents, less so. And there is a trap in Android’s documentation worth flagging. Google offers developers a hybrid API that, whenever Gemini Nano “isn’t available or isn’t supported” on a given handset, automatically routes the request to Gemini Flash or Gemini Pro in the cloud. An app advertising on-device AI on one phone may not deliver it on yours.
Private cloud, public promises — and what you can verify
A vendor-operated cloud is not automatically less private, but the nature of the protection shifts. Locally, the safeguard is physical: data never leaves. Remotely, it becomes a technical commitment that in some cases can be independently audited.
Apple detailed PCC in its security research blog on 10 June 2024, pledging that “no user data is retained in any form once the response is returned” and that data is “never available to Apple” — not even to staff with administrative access. The company also committed to publishing the software images of every production build so outside researchers can verify the claims.
Google launched an equivalent, Private AI Compute, on 11 November 2025. It runs on the company’s own TPU chips inside isolated enclaves it calls Titanium Intelligence Enclaves, and processed data remains “accessible only to you and no one else, not even Google.” It backs features such as Magic Cue on the Pixel 10 and Pixel Recorder summaries. On the local side, Google states that AICore isolates each request and keeps no log of input data or results, and holds no direct internet access — every connection, including model downloads, passes through an open-source component called Private Compute Services.
As for independent verification, options are thin but not zero. As of October 2026, Apple is the only vendor in this group offering users a record of what leaves the device: an Apple Intelligence Report in the Mac’s privacy and security settings that lists requests sent to PCC, configurable for the last 15 minutes or the last 7 days and exportable as a JSON file. Beyond that, three habits help. Check whether a feature demands connectivity — if it works in airplane mode, it is local. Treat a bare “on-device AI” label with scepticism, given how Android’s hybrid API behaves. And separate “private cloud” from cloud in general: a third-party assistant inherits none of these commitments and answers only to its own privacy policy.
The public debate still frames this as local versus cloud, and that framing has aged badly. Apple and Google have shown that a cloud can be built with serious, verifiable guarantees — the problem is that those guarantees stop at the edge of their own feature sets. The genuine weak point sits with third-party applications, which silently decide whether your request stays on the NPU or ships off to a server you know nothing about. Expect the NPU to keep gaining ground, though not because local models will match the biggest ones. Every request that never leaves the handset is a server cost the manufacturer avoids, and that incentive happens to align with your privacy — a welcome coincidence, but a coincidence all the same.