Guide
At home, or in the cloud?
“Edge AI” runs on the device in your hand or on a box at home; “cloud AI” runs in a data centre. The choice decides how fast an answer comes, how good it is, whether it works offline, and what leaves your home. Most products now use both.
In short. On the device, AI answers without a trip over the internet, keeps working offline and can keep your data at home, but it is limited by a small chip and its battery. In the cloud, the largest models are available, at the cost of sending your request out. Apple and Google now split the work: smaller models on the phone, larger ones in data centres, with rules about what is kept. The question to ask of any product is which requests leave, and what happens to them.
Two places to think
| On the device (edge) | In the cloud | |
|---|---|---|
| Speed | No trip over the internet; limited by the chip | A trip each time; large, fast servers |
| Strength | Smaller models | The largest models |
| Privacy | Data can stay with you | Data travels, under the provider’s rules |
| Without internet | Keeps working | Stops |
| Cost | Paid once, in the device | Paid per request, by someone |
How Apple splits it
Apple calls on-device processing the cornerstone of Apple Intelligence’s privacy, and sends “more sophisticated requests” that need larger models to Private Cloud Compute, servers it built for the purpose. In its June 2024 announcement Apple said the phone encrypts each request to servers it has checked, that the data is used only to fulfil that request and is not kept once the answer is returned, that Apple staff cannot reach it, and that every production build of the server software would be published for security researchers to inspect. Two details are worth knowing: the servers must read a request to answer it, so this is not end-to-end encryption in the usual sense, and these were Apple’s own commitments at launch.
How Google splits it
Google’s on-device model, Gemini Nano, runs inside an Android system service called AICore. Apps can use it to summarise, proofread and rewrite short messages, describe images, transcribe and translate speech, without sending the data out. Google lists the benefits as keeping sensitive data on the device, working offline, removing the network round trip and saving developers the cost of cloud models, while noting that speed depends on the phone’s hardware and that the model has limits. Developers can also choose a hybrid: prefer the device and fall back to the cloud, or the other way round. So the same app can keep one request at home and send the next one out.
What belongs where
Things that must happen fast, or must not leave the home, belong on the device or on a box at home: stopping a robot, keeping it away from people, the camera and the map of your rooms. Things that need the largest models belong in the cloud: understanding a long question, writing a reply, searching the web. The chip that makes on-device AI possible has a name: what an NPU is. And the models themselves: what a large language model is.
Questions to ask any AI product
- Which features run on the device, and which send data out?
- When a request goes to the cloud, is it kept, and for how long?
- Can anyone at the company read it?
- What still works without an internet connection?
- Where are the servers, and under which law? Why that matters in Europe.
How we split it
- Alkyon Concept is designed to keep its camera, raw sound and the map of your home at home, on its dock, and to keep its safety rules on board, where they work without any network.
- Lychnos Ring sends what you say, while you hold the button, to our cloud in the EU, where it becomes text and the voice clip is deleted; the answer comes back to the ring.
More on what leaves your hands and your home in privacy by design.
Sources, read 4 November 2026: Apple Security Research, Private Cloud Compute: a new frontier for AI privacy in the cloud, 10 June 2024; Android Developers, Gemini Nano and hybrid inference.
Questions
Is on-device AI more private?
It can be: data that stays on the device cannot leak from a server. Google says Gemini Nano keeps sensitive data on the device and that the service running it stores no record of inputs or outputs. But many products fall back to the cloud for harder requests, so it is worth checking what goes where.
Does on-device AI work without the internet?
Yes, that is one of its main strengths. Google’s developer pages say on-device models keep features working when the user is offline. The largest models still need a connection.
What is Apple’s Private Cloud Compute?
Apple’s servers for Apple Intelligence requests that need larger models than the device can run. In its June 2024 announcement Apple said request data is used only to fulfil the request, is not kept afterwards and is not available to Apple staff, and that it would publish the software for security researchers to inspect.