Read the strategy again, slowly: not every model has to run in a data center. Samsung's application US20200218962A1 (“Method and apparatus for neural network quantization,” published 2020-07-09) is about the other path — squeezing a model down to low precision so it runs on the device in your hand. Assigned to Samsung Electronics Co., Ltd. and classified CPC G06N 3/0454, it is on-device AI's cost engineering, dated 2020.
The line that matters for a segment reader is where the cost lands. Cloud inference is a recurring operating expense on the service provider's books. On-device inference, by contrast, shifts compute to hardware the customer already bought — the cost is amortized into the device, not the cloud bill. Quantization is what makes that shift technically possible, because phone silicon cannot run full-precision frontier models.
Samsung does not report an “on-device AI” revenue segment, and its disclosures discuss AI features in product terms rather than isolated financials. So this is a strategy signal, not a segment number. The patent tells you the company was investing in the compression techniques that make device-side AI viable while most of the market was still assuming everything ran in the cloud.
For the business reader the contrast is the point. A hyperscaler's AI story is a capex-and-inference-cost story. A device maker's AI story can be an IP-and-silicon story where the marginal inference cost is borne by the user. Same technique — quantization — two completely different places it hits the income statement.
The discipline: this is a published application, not a grant, so scope is unsettled, and we attribute no revenue to it. What it documents is that the on-device cost path was being patented by a major device maker in 2020 — a useful counterweight to the assumption that AI economics are only ever cloud economics.