The marketing name is "on-device intelligence." The filing name is more honest: "Kernel-level load balancing across neural engines." That is the title of Apple's granted patent US12277492B2, issued April 15, 2025, and it describes the part of on-device AI that no keynote slide bothers to show: how the operating system divides machine-learning work across a chip's dedicated neural-processing cores.

Here is what on-device AI actually is, before the adjectives. Instead of shipping your photo or your voice to a data center and waiting for an answer, the device runs the model itself, on a block of silicon — Apple calls its version the Neural Engine — built specifically for the matrix math that neural networks need. The privacy and latency benefits are real: the data never leaves the phone, and there is no round trip to a server. But a dedicated AI accelerator is only as good as the work you can keep feeding it.

“An electronic device may receive, at a first system routine from a client application, a provisioning request indicating that the application includes code for evaluating a machine learning model, wherein the first system routine executes in user space of memory on the device.”— U.S. Patent No. 12,277,492 source

That feeding problem is exactly what the grant addresses. The patent's framing — load balancing at the kernel level, the lowest layer of the operating system — points to the OS itself deciding which neural-engine resources handle which slice of a model, so the cores stay busy rather than stalling. The named inventors (including Sundararaman Hariharasubramanian and Andrew Yanowitz) and the CPC classifications in G06N (neural-network computing) and G06F 9/50 (resource allocation) line up with that reading: this is scheduling IP, not model IP.

Why does a scheduling patent matter more than another model announcement? Because the bottleneck in consumer AI has quietly moved. The models that run on a phone are largely commoditized techniques; the differentiator is whether the hardware can run them fast and cool enough that a user never notices. A grant on splitting work across accelerator cores is a claim on that differentiator — the efficiency layer, not the intelligence layer.

Apple's broader on-device posture shows up elsewhere in its portfolio too. The company holds US11893486B2 (issued February 6, 2024) on digital watermarking of machine-learning models — a method for marking a model so its provenance can be checked — and US12651020B2, a June 2026 grant titled "Digital assistant intelligence engine." Together they sketch a company building the plumbing around on-device models: how to run them, how to mark them, how to wire them into an assistant.

The discipline here, in this column's house style: a granted claim covers its specific language, not the entire idea of "AI on a phone." None of these patents proves Apple ships exactly the claimed method in a given product. But read together they answer the question the marketing dodges — what is actually novel about on-device intelligence? Less the model, more the scheduler that keeps the silicon fed.