The popular story of AI ownership is a story of a few names: a handful of labs and silicon vendors that, in the telling, hold the field. A patent publication drop is a useful place to check that story against a document, because it shows not what a company announced but what it spent research money on roughly eighteen months ago, surfaced only when the application publishes. The applications published the week of June 25, 2026 do not read like a field owned by a few. They read like a field that has spread out.

Start with the counts, because they set the scale. A search of that day's published applications returns 1,577 hits on "machine learning" and 1,274 on "neural." Those two slices overlap heavily and together describe a large fraction of a single week's filings. The classification data confirms where the weight sits: the dominant patent class in both slices is G06N 20/00, the catch-all code for machine-learning methods, with 88 applications in the machine-learning slice. The next-largest AI-specific class is G06N 3/08, the code for neural-network training, with 29 applications in that same slice and 37 in the neural slice, followed by G06N 3/045 and G06N 3/084. G06N is the corner of the classification system reserved for machine learning and neural architectures, and it is the busiest AI corner of the drop. So far this matches the narrative: AI filings cluster, and they cluster in G06N.

The narrative breaks at the next question, which is who filed them. Across the machine-learning slice of more than 1,500 applications, no assignee holds a commanding share. The largest single filer is Toyota, with 18 applications. InterDigital, the wireless-standards licensor, is next at 17. Samsung follows at 16, then LG Electronics and NEC at 9 each, Huawei and Microsoft at 8 and 7, Advanced Micro Devices at 7, Sony at 6, and Capital One Services at 5. A large share of the slice carries no normalized assignee at all in the index. Put plainly: the biggest AI filer in the drop accounts for roughly one application in eighty, and the top of the table is a list of companies from different industries rather than a leaderboard of model labs.

The leaders are an industry cross-section, not a lab roster

Read the names by sector and the horizontal pattern is hard to miss. The telecom and wireless contingent is the largest: InterDigital's filings sit in the machine-learning slice because modern radio standards now fold AI/ML into the network. Its application US20260181701A1 describes initial-access assistance information exchanged between a wireless transmit/receive unit and a network node, the kind of physical-layer optimization where machine-learning methods are increasingly disclosed. The hero record of this drop sits in the same telecom lane: US20260181740A1, assigned to Nokia Technologies, covers dataset-sharing instructions for training a model split across a device and a base station.

The present disclosure relates to a method and a system for improving the establishment of a two-sided AI/ML based model implemented in an apparatus AP1-APX, such as an user equipment, UE, and a device DE1-DEY, such as a server or a gNodeB, gNB, in a wireless network system in which, by sending configuration information providing instructions regarding at least what to share for data set sharing between the apparatus AP1-APX and the device DE1-DEY, an efficient separated or sequential training of the two-sided model can be generated.— Dataset Sharing Transmission Instructions for Separated Two-Sided AI/ML Based Model Training, US20260181740A1

The silicon contingent is the second cluster, and it is the one the popular narrative expects. Advanced Micro Devices appears with 7 machine-learning and 11 neural applications; its memory-module application US20260181777A1 describes pin cutouts on a DIMM whose stated purpose is supporting "machine-learning applications, neural network" workloads. Intel is the single largest filer in the neural slice at 14 applications, almost all of them packaging and substrate work, glass-core substrates, through-glass vias, chiplet interposers, whose specifications cite neural-network processing units as the workload being served, as in US20260182456A1. Samsung leads the neural slice when its two index spellings are combined and brings device-level filings such as US20260181944A1, a gate-all-around transistor structure whose disclosure points to neural-network devices among its applications. The chip names are present, but they are filing on the substrate beneath AI, not on a monopoly over it.

The business read: AI IP has gone horizontal

The third cluster is the one that completes the picture, because it is the one the narrative leaves out: enterprise and finance. Capital One Services, a bank, sits in the top fifteen machine-learning filers, and its own classification data shows applications landing in G06N 20/00. Its application US20260179412A1 describes a liveness-verification method that analyzes image depth data to determine whether a captured subject is real, an applied-AI filing from a financial institution rather than a technology vendor. When a bank, an automaker, a wireless licensor, and a chip designer all appear in the same fifteen-name list for the same week, the through-line is not which lab owns AI. It is that AI methods have become a standard ingredient that companies across unrelated industries now file on as a matter of course.

That is the commercial signal worth naming, and it is grounded entirely in the distribution. When a technology is owned by a few, the filing data concentrates: a handful of assignees carry an outsized share of the relevant class, and the rest is a long tail. This drop does not look like that. The relevant class, G06N, is busy, but the assignee column underneath it is flat, with the top filer at 18 against a pool above 1,500 and the leaders drawn from four distinct industries. The applications, taken together, suggest that AI intellectual property in this week's publications is broad-based rather than concentrated, distributed across telecom, automotive, chips, and enterprise rather than pooled in a single lab. Read as a market indicator, that is the difference between a technology a few companies sell and a technology every company builds with.

The standard caveat is load-bearing here, and it cuts toward the same conclusion. These are published applications, not granted patents; the counts reflect where research money was directed roughly a year and a half ago, not what any of these assignees can yet enforce, and the classification facets are sensitive to how the index normalizes assignee names and codes. For reading direction rather than litigation exposure, that is exactly the value. The drop will not tell you who is winning AI, and nothing here should be read as a claim that anyone is. What the distribution does show, in the counts rather than in any single filing, is that the act of patenting machine-learning and neural-network methods is no longer the province of a few names. It has become horizontal, a layer that a telecom licensor, an automaker, a memory designer, and a bank all reached for in the same week.