Apple's own write-up of its cloud AI system admits something most AI marketing leaves out. For requests sent to Private Cloud Compute, complete end-to-end encryption is not an option, because the server has to read your request to answer it.
Inside the iPhone: a small model built to stay home
Apple's local model is a compact, approximately 3-billion-parameter model tuned to run on Apple silicon. Apple says its latest models have improved tool-use and reasoning capabilities, understand image and text inputs, are faster and more efficient, and are designed to support 15 languages. In practice, the phone can handle a lot of everyday language work without sending anything out.
Small businesses should pay attention to the developer side. Apple gives app makers direct access to the on-device model through a new Foundation Models framework, and the AI processing is free of cost and accessible with just a few lines of code. Apple lists text extraction and summarization as examples. A note-taking or invoicing app built this way doesn't need to pay a cloud AI provider for every request, and the request doesn't have to go to one.
Apple sums up its position in one line: "Privacy is the default, not something you have to manage."
Gemini Nano: Google's offline-first pitch
Google describes Gemini Nano as a way to deliver generative AI without needing a network connection or sending data to the cloud. The model runs in Android's AICore system service, which uses the phone's hardware and keeps the model up to date.
Google's developer page is direct about how this works: on-device generative AI executes prompts locally, eliminating server calls. Google lists the benefits as better privacy, offline use and lower costs, and adds that inference speed depends on device hardware. So an older phone may feel slower, but your data stays put either way.
Google also says what the system does after it answers you. AICore is built to isolate each request and doesn't store any record of the input data or the resulting outputs after processing them. Model updates arrive via Private Compute Services. App developers usually reach Gemini Nano through the ML Kit GenAI APIs, which are built on top of AICore.
When a request leaves the iPhone
The on-device model has a larger sibling: a mixture-of-experts server-based model with a novel architecture tailored for Private Cloud Compute. Apple uses Private Cloud Compute (PCC) for advanced features that need to reason over complex data with larger foundation models.
Apple's main promise is that personal data sent to PCC isn't accessible to anyone other than the user — not even to Apple. Since the server has to read your request, Apple relies on hardware and software limits instead of encryption alone. Each PCC computer must have technical enforcement for the privacy of user data during processing, and must be incapable of retaining user data after its duty cycle is complete.
Apple also says it removed the tools data centers normally use to look inside running systems, such as remote shells and system introspection and observability tools. Operations staff get only a small, restricted set of operational metrics. Apple's design rules are stateless computation, enforceable guarantees, no privileged access, non-targetability, and verifiable transparency. Apple also invites outside checks: the Apple Security Bounty will reward research findings in the entire Private Cloud Compute software stack, with especially large payouts for flaws that undermine its privacy claims.
A smartphone lying on a wooden desk in the foreground, with rows of dimly lit server racks in a data center blurred in the background.
| Question | Gemini Nano | Apple on-device | Apple PCC |
|---|---|---|---|
| Where it runs | Phone (AICore) | Phone (Apple silicon) | Apple's cloud servers |
| Needs a network? | No | No | Yes |
| Can the server read it? | No server involved | No server involved | Yes, during processing |
| Kept afterward? | No record stored | Stays on the device | Can't be retained |
What this means for your data
The useful way to see this is not local good, cloud bad. There are two different kinds of promise.
Gemini Nano and Apple's on-device model make the simpler one: your request never leaves the phone. You don't have to trust any server, because none is involved.
Private Cloud Compute makes a harder promise. Your request does leave, the server can read it while working on it, and Apple says the system is built to forget it and is open to outside inspection. That is a stronger guarantee than ordinary cloud AI offers, but it depends on the design working as described rather than on the data never leaving your hands.
My advice for small businesses: For client records, health details or anything covered by a confidentiality agreement, choose apps that say their AI features run on the device. Treat cloud processing as a choice you make deliberately, even when it's well protected. Ask app vendors directly which of their features run locally. The good ones will know the answer.
Sources
- machinelearning.apple.com – "Updates to Apple’s On-Device and Server Foundation Language Models"
- developer.android.com – "developer.android.com"
- security.apple.com – "Private Cloud Compute: A new frontier for AI privacy in the cloud - Apple Security Research"