Someone dictates a private journal entry into their phone, and Apple Intelligence summarizes it using a model running entirely on the device's own chip. No request leaves the phone, a distinction Apple built its entire 2024 marketing pitch around, down to publishing technical papers on Private Cloud Compute specifically to describe what does and does not get sent out when a task is too large for the device alone.
The bargain was straightforward even if it was rarely stated plainly: send data to the cloud, get more capability than a local device could provide, and trust the operator of that cloud with whatever was in the request. On-device chips, Apple's Neural Engine, Google's Tensor silicon running Gemini Nano on Pixel phones, the Qualcomm processors inside Microsoft's Copilot+ PC category, break that trade in half, offering meaningfully capable AI without the data ever leaving the device, at some cost in raw capability compared with the largest cloud-hosted systems.
The people most affected are not primarily technical. They are ordinary users of features like on-device transcription, local photo search, or offline assistants, for whom the cloud was previously invisible infrastructure they neither consented to nor could inspect.
What breaks, at least partially, is the business logic that made centralized processing attractive in the first place: aggregated data as a byproduct of service delivery, useful for improving models or targeting features. A more capable on-device model is a worse data-collection funnel by design, which means companies whose business model depends on that funnel have real incentive to keep processing in the cloud even where local alternatives now exist.
What is opening up is a genuine architectural choice for product builders that did not previously exist in practice: deciding explicitly what has to leave the device versus what can be handled locally. That decision is becoming a marketable feature in categories, health, journaling, communication, where users have specific reason to want processing to stay put.
Apple did not build Private Cloud Compute because on-device chips got good enough on their own. It built it because "trust us with your data" had stopped being a sufficient answer, and a phone that can prove what it did not send is now worth more to some customers than a cloud that only promises what it will do with what it did.