When you ask your phone’s AI to clean up a text or summarize an email, the work has to happen somewhere. Sometimes it’s the chip in your pocket. Sometimes your request travels to a server farm and back without you noticing.
You usually can’t tell which, and a lot of apps use both. But the difference affects your privacy, your speed and what the AI can actually do.
On-Device AI
On-device AI runs on the hardware you’re holding: your phone, laptop or tablet. The model is small enough to live on the device itself. Apple’s on-device model for Apple Intelligence is about 3 billion parameters, and Google’s Gemini Nano runs on Pixel phones the same way.
The upside is that nothing has to leave your device, it works without a signal, and there’s no round trip to a server. Google’s developer docs say Nano works “without needing a network connection or sending data to the cloud.” The downside is size. Apple says plainly that its on-device model “is not designed to be a chatbot for general world knowledge.” It’s built for quick jobs like rewriting a sentence or sorting notifications, not deep research.
Cloud AI
Cloud AI runs on big servers in a Data Center somewhere else. That’s where ChatGPT, Claude and the full versions of Gemini live. The models are way bigger than anything that fits on a phone, so they’re smarter, they know more and they can handle long, complicated tasks.
The tradeoff is that your request has to travel there, which means you need a connection, it can take a bit longer, and your data leaves your device. Some companies try to soften that. Apple’s Private Cloud Compute, its cloud backup for harder requests, promises that “no user data is retained in any form after the response is returned.”
The tell
Ask where the thinking happens. If it still works in airplane mode, it’s on-device. If it needs Wi-Fi or cell service, it’s in the cloud.
It’s like doing math in your head versus calling your accountant. Mental math is instant and private, but you’re not doing your taxes that way. The accountant can handle the hard stuff, but you have to hand over your paperwork and wait.
Lots of features now do both (kinda like checking the easy stuff yourself and calling the accountant for the rest). Hardware makers are pushing more onto the device, too. Microsoft’s Copilot+ PCs, for instance, require a special AI chip just to run some of their features locally.

