Open with the records, then the verdict. Apple holds three granted patents that, read in isolation, look like unrelated engineering. Read together, they are a strategy. The first, US12651020B2, "Digital assistant intelligence engine" (issued June 9, 2026), is the application layer — the part a user talks to. The second, US12277492B2, "Kernel-level load balancing across neural engines" (issued April 15, 2025), is the efficiency layer — how the work gets spread across the chip. The third, US11893486B2, "Digital watermarking of machine-learning models" (issued February 6, 2024), is the provenance layer — how a model is marked so its origin can be verified.
Here is what the docket actually says, layer by layer. The assistant engine is the thing the keynote demos: ask a question, get an answer drawn from on-device understanding. But an assistant is only as good as the compute it can call on instantly and privately, which is where the scheduler comes in — the load-balancing grant is about keeping the neural-engine cores busy so the assistant feels immediate rather than laggy. The keynote sells the first; the second is what makes the first usable.
The watermarking patent is the one that reveals the most, because it answers a question users never ask. If models run on-device, downloaded and updated over time, how does Apple know a given model is genuinely its own, unmodified and untampered? A watermark embedded in the model is a method for checking exactly that. It is a control-and-trust mechanism, and its presence tells you Apple is thinking about on-device models as managed assets with a chain of custody, not just files.
Stack the three and the bet comes into focus: own the run (the assistant), own the route (the scheduler), own the receipt (the watermark). Each is a different discipline — natural-language interaction, low-level resource scheduling, model security — and a company that files across all three is not dabbling. It is building the full stack of on-device intelligence and claiming IP at every layer it controls.
What the filings do not say is as important as what they do. None of these grants discloses how well the assistant performs, how much the scheduler actually speeds things up, or whether the watermark survives every model update. Claims describe methods, not benchmarks. Anyone telling you Apple's on-device AI is definitively ahead or behind is reading a verdict the patents do not contain.
But the structural conclusion is sound, and it is the one this desk cares about: Apple's on-device AI is not a feature it announced — it is infrastructure it has been quietly patenting in layers. The demo showed you the assistant. The docket shows you the scheduler and the watermark underneath. Follow the filing, not the demo, and the three records become one coherent bet on owning intelligence at the edge of the device.