On-Device AI vs Cloud AI: What Your Phone Really Shares

Every AI feature on your phone runs in one of three kitchens: your own kitchen (on-device — nothing leaves the phone), a restaurant (ordinary cloud — your request is processed on company servers, often logged to your account), or a new third option best described as a sealed private kitchen — cloud servers built so that, by design, nobody can watch the cooking, and outside auditors have checked the locks. Knowing which kitchen each feature uses tells you almost everything about your privacy, and the honest answer changes feature by feature.

The three kitchens, in one table

That third row is the 2025–26 story most explainers miss. Apple’s Private Cloud Compute (PCC) deletes your data after each response, blocks even Apple’s own staff from peeking, and backs the claim with published source code plus bounties up to $1,000,000 for anyone who can break in (Apple Security Research). Google’s version, Private AI Compute, went further on independent scrutiny: security firm NCC Group spent about 100 person-days reviewing it before launch (NCC Group report; Google). And in June 2026 Apple confirmed something remarkable: PCC’s heaviest work now runs partly on Google Cloud hardware, under the same seals (Apple).

So what actually stays on the phone?

Less than the ads suggest. Apple’s on-device models handle the small stuff — call context, voice control, Safari tab sorting — while anything heavy goes to PCC (Apple Newsroom). Google’s Gemini Nano does on-phone summarizing, proofreading and transcription (Android Developers). Samsung keeps translation and call features local but sends Generative Edit to the cloud “as needed” (Samsung Newsroom).

Two catches to remember:

• Features can quietly switch kitchens. An “on-device” feature may fall back to cloud when it’s unsure. The label describes the default, not a guarantee.
• The escape hatch changes the rules. When Siri hands your question to ChatGPT, Apple hides your IP address — but if you’ve linked a ChatGPT account, OpenAI’s normal retention and training rules apply (Apple legal disclosure). Same story we detailed in what chatbots keep about you.

The fine print: hardware and battery

Hardware gate. Google’s newest AI tier requires 12GB of RAM plus a flagship chip — a bar even the Pixel 9 misses (Android Authority). Phone-sized models that punch above their weight exist — we covered a 27B model running on a phone — but shipping products gate the best features to expensive devices, and chip prices are climbing, as our Snapdragon price-hike report showed.

Battery. In the only independent lab measurement we found, running AI locally used 21–37 times more battery per response than the cloud, dropping a phone to about two hours of sustained use (Greenspector). Caveat: that test used generic models on older hardware, and vendors are closing the gap — but it explains why every company ends up hybrid.

The switches you control

• iPhone: Settings → Apple Intelligence & Siri → off. Kills the generative features, frees ~3GB. The ChatGPT handoff is off by default and has its own toggle. There is no “keep AI but never use cloud” switch — the split is per-feature.
• Samsung: Settings → Advanced Intelligence → “Process data only on device.” The clearest switch in the industry: keeps translation and transcription, drops the cloud-dependent extras.
• Other Android: no master switch. You manage it per feature, mainly through Gemini Apps Activity. The weakest control story of the three.

The India angle

For buyers in the ₹15,000–25,000 band, the honest news: real on-device generative AI isn’t there yet. Mid-range chips like the Exynos 1380 carry NPUs built for camera and voice work, not generative models (Samsung Semiconductor), and Samsung’s cheapest Galaxy-AI phone in India launched at ₹22,999 (Samsung Newsroom India). Yet 89% of Indian buyers say AI features influence their choice (Counterpoint x Flipkart) — a gap marketing loves. And the old “on-device saves data costs” pitch has aged: mobile data now runs about ₹6.60 per GB (Mappr/Cable.co.uk). The real Indian case for on-device AI is that it works where the network doesn’t. One more thing on-device does not do: satisfy the law by itself — India’s DPDP Act cares about consent, not where the processing happens.

Take it from here

Three questions to ask about any AI feature before trusting it with something private: Which kitchen does it run in? What happens when it’s unsure (does it fall back to cloud)? And if it’s cloud — is it the logged kind or the sealed kind? If a vendor can’t answer those three plainly, that’s your answer. More guides like this on our how-to and explainers hub.