One assignee dominates this dataset, holding 85 of the 137 records in scope — roughly 62% of the entire field on its own. The next four assignees combined bring the top-five total to 109 records, or 79.6% of all records. Below the top five, holdings drop to single digits, with tenth place at just 4 records, so the field is best described as one leader followed by a flat mid-tier rather than a gradual long tail.

Filing peaked at 30 records in 2020 and has declined since, falling from 25 in 2021 to 5 in 2024 — an 80% drop over that three-year window. That said, patent publication lags filing by roughly 18 months, so the lower counts in 2025 and 2026 (14 so far) are not yet reliable evidence of continued decline; they will likely revise upward. The safest reading is that the field cooled from its 2020 peak through 2024, with the most recent years still incomplete.

Digital information transmission (IPC class H04L) appears in 62.0% of the 137 records, making it the dominant class by a wide margin. Wireless communication networks (H04W) follow at 31.4%, and AI-based computing (G06N) appears in 24.8% of records, reflecting the growing overlap between edge transmission architectures and on-device model inference. Because a single record can carry several IPC classes, these percentages are shares of the full 137-record total and add up to more than 100%.

The smallest IPC shares in this dataset — G06Q (business/commerce), G06V (image/video recognition) and G16H (healthcare informatics), each at 3.6% of records — mark branches where filing density is comparatively low next to the dominant transmission and wireless classes. That does not guarantee an easy path to allowance, but it does suggest fewer competing claims to design around in edge-adjacent commerce, vision, or healthcare applications. Sub-areas like clock synchronization across intermittent links and data prioritization on resource-constrained nodes also show up as recurring but not saturated themes across the corpus.

US12591804B2, assigned to Intel Corporation and published 2026-03-31, covers an edge computing node apparatus that performs federated machine-learning training rounds — including selecting a candidate set of clients, distributing a global model to them, and training on coded data from those selected clients. It sits within the same portfolio driving this field's top-heavy concentration, so claims that overlap its specific client-selection and coded-training-data mechanics are likely to face close scrutiny. Filings that use different aggregation logic, different client-selection criteria, or move the inference step off the federated-training model would sit further from its claim boundary, but a formal freedom-to-operate review of its full claim set is the only way to confirm clearance.