Generative AI Patents: Who Leads, Where the Gaps Are 2026
- Filings grew 467% from 2021 to 2024, peaking at 6,361 records in 2024 before the expected 18-month publication lag pulls the two most recent years down artificially.
- The top 10 assignees hold only 22.8% of all 29,906 records, leaving a long tail of single- and few-filing entrants rather than a market locked up by a handful of firms.
- G06F and G06N together anchor over half the filing activity, while speech/audio (G10L) and healthcare informatics (G16H) each sit near 4%, marking them as comparatively open branches.
Filing growth compares 2021 (1,121 records) with 2024 (6,361) — a three-year span. 2024 is the most recent year we treat as complete: publication lags filing by roughly 18 months, so 2025 onwards are still filling in and any growth rate that ends there would understate the field. Top-5 share is the combined record count of the five largest assignees divided by all 29,906 records in scope (CR5), not by the ranked leaders only.
What the generative AI patent record actually shows
The dataset covers 29,906 published records filed against generative model and training-data claim language between 2015 and the 2026-08-31 cut-off. Filing activity was modest through the late 2010s and accelerated sharply from 2021, when annual output stood at 1,121 records, to a peak of 6,361 in 2024 — a 467% increase over that three-year span. Because publication trails filing by roughly 18 months, the 2025 and 2026 figures in the trend chart understate real filing volume and should not be read as a slowdown.
Ownership is not tightly held: the leading assignee accounts for 1,495 records, and even the top 10 combined reach only 22.8% of all records in scope. That leaves most of the corpus distributed across a long tail of assignees filing in narrower technical niches, which is where freedom-to-operate analysis usually needs to look first.
Let an AI agent run this analysis on your own technology
Pick a task. Every answer cites the patents behind it.
Filing trends and technology composition
Two views of the same 29,906-record dataset: how filing volume moved year over year, and which IPC subclasses the claim language actually falls under.
A three-year surge, then a lag-distorted tail
Annual filings rose from 288 in 2017 to a peak of 6,361 in 2024, with the +467% jump between 2021 and 2024 marking the period generative model claims moved from a niche filing category into mainstream patent strategy. The 2025-2026 figures (488 in the partial latest year) are undercounts caused by publication lag, not a genuine drop in activity.
Core compute and image classes dominate, but not exclusively
G06F (electric digital data processing) and G06N (AI-specific computing) lead at 26.9% and 23.9% of the 29,906 records respectively, confirming that most claims are still anchored in general-purpose data processing and model architecture rather than a single application. Image processing (G06T, 12.8%) and recognition (G06V, 9.4%) follow, while business/admin (G06Q), telecom transmission (H04L), speech/audio (G10L) and healthcare informatics (G16H) each sit in the 4-6% range — smaller footprints that mark where application-specific claim density is still thin.
Shares are the percentage of the 29,906 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Go deeper on Generative AI Patent Landscape with Eureka
This page is one run against one query. Ask Eureka your own question about generative ai patent landscape and every answer comes back with the patent numbers behind it.
Try EurekaA representative claim and the most-cited prior art
US12646296B2 — Training data generation apparatus (NEC Corporation, 2026-06-02)
A training data generation apparatus uses spatial data of an actual object and thereby trains a generative model to perform conversion from the spatial data to a feature vector and conversion from the feature vector to spatial data. The apparatus generates a sample of the feature vector as a realization value of a probability distribution defined by a set of parameters, then generates training data used to train an object recognition model based on spatial data the generative model outputs when that sampled feature vector is fed back in.Filed by NEC, this is one of the most recent records in the dataset and illustrates a specific pattern: using a generative model as a synthetic-data engine to train a separate downstream recognition model.
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US20180075581A1 | Super resolution using a generative adversarial network | 863 |
| 2 | US20220066456A1 | Obstacle recognition method for autonomous robots | 760 |
| 3 | US20220126864A1 | Autonomous vehicle system | 675 |
| 4 | US10388272B1 | Training speech recognition systems using word sequences | 484 |
| 5 | US20080292194A1 | Method and System for Automatic Detection and Segmentation of Tumors and Associated Edema (Swelling) in Magne… | 455 |
| 6 | US10573312B1 | Transcription generation from multiple speech recognition systems | 445 |
| 7 | US7752152B2 | Using predictive user models for language modeling on a personal device with user behavior models based on st… | 437 |
| 8 | US20200175961A1 | Training of speech recognition systems | 410 |
| 9 | US20190108396A1 | Systems and methods for object identification | 379 |
| 10 | US20240386015A1 | Composite symbolic and non-symbolic artificial intelligence system for advanced reasoning and semantic search | 374 |
Citation counts accumulate over time, so older records such as the 2018 super-resolution GAN filing lead the table by influence rather than by current filing relevance.
Each row carries its publication number; clicking a row searches Eureka by that number.
Put your own technology through the same analysis
Eureka on the web
When you want the answer in the next five minutes.
The agent works the prompt against patents and technical literature, citing every source.
Run your analysis now →MCP server & REST API
When it has to run inside your own pipeline.
Patent search, landscape analysis and assignee resolution as MCP tools. Drop them into any agent framework, or call REST directly.
Browse MCP servers →What the data means for filing strategy
Three findings that change how a freedom-to-operate or whitespace search on generative AI should be scoped.
Ownership is diffuse, not cornered
With the leading assignee at 1,495 records and the top 10 combined reaching only 22.8% of the field, no single firm's portfolio forecloses the space. Clearance work should expect to clear many small and mid-size holders, not just the largest names.
The surge is real, the tail-off is not
Filing volume nearly sextupled between 2021 and 2024. The apparent decline in 2025-2026 tracks the 18-month publication lag rather than a cooling of interest, so recency-weighted searches should treat the last two years as incomplete.
Core compute claims dominate application claims
General digital data processing (G06F) and AI-specific computing (G06N) together cover more filing activity than any application-layer class. Application areas such as speech/audio and healthcare informatics remain comparatively thin, at roughly 4% of records each.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to generative ai patent landscape, with the prior art for and against each one.
Assignee landscape and recent momentum
The ranked leaders span compute, cloud and device platform companies, but their most recent-year filing counts have dropped sharply — consistent with the publication lag rather than retreat from the field.
A single leader, no runaway gap
The top-ranked assignee holds 1,495 records, well ahead of fifth place at 518 and tenth place at 299 — a gradual taper rather than a cliff, which is typical of a field still being staked out rather than consolidated.
Latest-year counts look weak across the board
Every major assignee shows a steep year-over-year drop in the most recent year, from roughly -83% to -92%. Given the 18-month publication lag, this is expected: it reflects filings still working through the pipeline, not falling investment.
Joint filing is limited and concentrated
Only 10 co-assignee pairs appear in the dataset, the strongest linking related corporate entities of the same group rather than independent competitors partnering. Cross-company collaboration on generative AI filings remains the exception.
| Assignee | Recent year | YoY |
|---|---|---|
| NVIDIA Corp | 26 | -88% |
| Citibank | 20 | -86% |
| Microsoft Technology Licensing, LLC | 16 | -90% |
| Google LLC | 12 | -92% |
| Samsung Electronics Co., Ltd. | 9 | -88% |
| Adobe Inc. | 4 | -83% |
| Toronto-Dominion Bank | 3 | -89% |
| International Business Machines Corporation (IBM) | 0 | -100% |
Where to take this next
The landscape points to a diffuse ownership structure with a few clearly thin branches — the next step is turning that into a scoped search or a claim-drafting position.
Run a freedom-to-operate check on a specific application
General compute and vision classes are dense, but speech, healthcare informatics and business-workflow applications of generative models remain comparatively open. A targeted search on one of these branches will surface fewer blocking claims than a search on core model-architecture language.
Explore with Patsnap EurekaTrack the leading assignees' pipeline, not just their granted counts
Because publication lag depresses the last one to two years of data for every filer, current portfolio comparisons should weight 2023-2024 filings more heavily than the apparent 2025-2026 drop-off.
Explore with Patsnap EurekaCommon questions about the generative AI patent landscape
This dataset covers 29,906 published records matching generative model and training-data claim language, filed between 2015 and the 2026-08-31 data cut-off. Annual filings grew from 288 in 2017 to a peak of 6,361 in 2024, a 467% increase between 2021 and 2024 alone. The 2025 and 2026 counts appear lower only because publication typically lags filing by about 18 months, so those years are still filling in.
The dataset's ranked leader holds 1,495 records, with the fifth-placed assignee at 518 and the tenth at 299 — a gradual decline rather than a sharp cliff. Even combined, the top 10 assignees account for only 22.8% of all 29,906 records in scope, so the field is not controlled by a small handful of companies. Most filing activity sits in a long tail of assignees with smaller, more specialised portfolios.
General digital data processing (IPC class G06F) and AI-specific computing (G06N) lead, covering 26.9% and 23.9% of the 29,906 records respectively. Image data processing and recognition (G06T and G06V) follow at 12.8% and 9.4%. Because a single record can carry multiple IPC classes, these shares add up to more than 100%, and they reflect claim breadth rather than a ranking of importance.
The comparatively thin classes are speech and audio synthesis (G10L, 4.2% of records) and healthcare informatics (G16H, 4.1%), alongside business and administrative applications (G06Q, 5.6%). These lower shares indicate less claim density relative to core compute and vision classes, which typically means more room to file a defensible claim without stepping directly on existing art. It does not mean the underlying technology is immature, only that fewer claims have staked it out so far.
Publication of a patent application normally lags its filing date by roughly 18 months, so any dataset's most recent one to two years will always look artificially low. In this dataset, filings fell from a 2024 peak of 6,361 to 488 in the most recent partial year, and every major assignee shows a steep year-over-year drop at the latest year. That pattern reflects filings still moving through the pipeline, not a genuine retreat from generative AI as a filing area.
Research Generative AI Patent Landscape in depth with Eureka
Go past this page: query the whole generative ai patent landscape corpus yourself, in your own scope.
Every answer comes back with patent numbers you can open.
Disclaimer. This page is generated from Patsnap Eureka data drawn from a limited snapshot of global patent and scientific-literature records, and is provided for general information and reference only.
Patent data carries inherent limitations: recent filings (typically the most recent 18–24 months) are under-counted due to standard publication lag; counts may be reported at either a patent-family or a patent-record basis and are not always directly comparable; classification, applicant-name, and citation data may contain errors, duplicates, or omissions; and the underlying search query defines and constrains the scope shown. As a result, the analysis may be incomplete or inaccurate and may not reflect the full technology landscape.
Nothing on this page constitutes an exhaustive prior-art, novelty, freedom-to-operate, or validity search, nor does it constitute legal, financial, investment, or professional advice, and it should not be relied upon as such. Any patent, commercial, or strategic decision should be verified independently and reviewed with qualified patent, legal, and domain professionals. Patsnap makes no warranties, express or implied, as to the accuracy, completeness, or fitness for any particular purpose of the information presented.
Machine translation. Assignee and organisation names originally recorded in Chinese, Japanese or Korean have been rendered into English by an AI translation step so that the tables stay readable. These renderings are best-effort and may not match a company’s registered English name; the original name is what the underlying patent record carries, and it is what any Eureka query launched from this page uses.