A month after WWDC, the details of Apple Intelligence have settled enough to think through what they actually mean. The marketing was compelling. The engineering underneath is where things get interesting.
The three-tier model
Apple Intelligence does not run everything on-device. It uses a three-tier architecture. Small models run entirely on-device with no network access. More complex tasks go to Private Compute Cloud, Apple's server infrastructure running on Apple Silicon. The most complex requests, the ones that need GPT-4 class reasoning, route to OpenAI through a privacy-preserving relay.
Apple's key claim on the second tier is that Private Compute Cloud uses cryptographic attestation to verify the user's device is talking to genuine Apple hardware running genuine Apple software. The server cannot store the request or any personal data. Apple themselves cannot access what you asked. That is an unusual privacy model for cloud AI and the technical implementation is genuinely novel.
The Neural Engine requirement
Apple Intelligence requires at least an A17 or M-series chip. The Neural Engine in these chips handles the matrix operations that LLMs need efficiently enough to run a useful model at real-time speeds without draining the battery in 20 minutes. This is the hardware threshold Apple drew.
iPhone 15 Pro and Pro Max qualify because of A17 Pro. The non-Pro iPhone 15 uses A16 and does not make the cut. Every iPhone 16 qualifies, including the base model with A18. Apple clearly designed the iPhone 16 lineup with this cut-off in mind.
What on-device means for privacy
The on-device tier is where the privacy argument is strongest. Your personal context, your messages, your email content, the documents you asked the system to summarise: none of that leaves the device for on-device tasks. The AI sees your data but it never transmits it anywhere. For enterprise use cases involving sensitive data, this changes the compliance conversation significantly.
The OpenAI integration on the third tier is opt-in. You explicitly choose to send a query to ChatGPT. Siri will ask before routing anything off-device to a third party. That is a meaningful boundary.
What it means for developers
The App Intents framework is how apps get access to this ecosystem. If you surface your app's actions through App Intents, the upgraded Siri can execute them with context. A productivity app that registers its "create new task" action can have Siri create tasks based on email content, calendar events, or messages, without your app code seeing any of that data. The OS brokers it.
This is Apple doing what Apple does: abstracting the complexity behind a framework and controlling the privacy boundary at the OS level. For developers who build within the framework, it is powerful. For those who want direct model access, it is closed.