Here is the tell, and it is in the repetition. Google holds US11403540B2, "On-device machine learning platform" (issued August 2, 2022), and then US12282869B2, with the identical title "On-device machine learning platform," issued April 22, 2025 — a continuation in the same family, with the same lead inventors (Pannag Sanketi and colleagues). When a company keeps re-filing on the same titled invention across years, it is not chasing a feature. It is fencing a platform.
The word "platform" is doing the work. A feature is a thing a user sees — face unlock, smart reply. A platform is the layer underneath that many features call into: a system service that handles loading models, running them on the device's accelerators, and managing the lifecycle, so that any app can request on-device inference without building the machinery itself. The grants describe that service layer, classified in G06N 20/00 (general machine learning) and G06N 5/04 (inference) — the infrastructure classes, not an application class.
Why does owning the platform layer matter more than owning a model? Because the layer is where lock-in lives. On Android, the apps that depend on Google's ML platform to run their on-device features are, in a quiet way, dependent on Google — the same structural position Google occupies with Play Services, location, and the rest of the system services that make Android Android. A model can be swapped. A platform that thousands of apps build against is sticky.
This is the platform-economics read of a technical patent. The interesting question is never just "is the invention clever?" It is "what position does owning it create?" Repeated grants on a system-level ML platform create the position of being the default place on-device intelligence gets run — which is leverage over both app developers (who depend on the layer) and the broader question of whose silicon and whose runtime become standard.
Contrast this with how the same capability is framed for users. To a consumer, on-device ML is sold as privacy and speed — your data stays on your phone, answers come instantly. Both are true. But the durable corporate value is not the privacy story; it is the infrastructure ownership the platform patents stake out. The user gets a fast, private feature; Google gets a layer everyone has to go through.
The honest limits: these are granted claims on specific platform mechanisms, and a continuation re-asserting a title does not by itself prove market dominance. Competing on-device ML runtimes exist. But the filing pattern is unambiguous about intent — Google is treating on-device ML as owned infrastructure, the kind you patent across a decade, not a feature you ship and forget. Read the portfolio, not the keynote, and the strategy reads as platform from the first grant.