AI Accelerators Patents: Who Leads, Where the Gaps Are 2026
- 56.1% concentration. The top five assignees alone account for 6,123 of the 10,910 records in scope — over half the entire field sits with a handful of filers.
- +67% filing growth, 2021 to 2024. Filings rose from 1,318 in 2021 to a peak of 2,200 in 2024, the last year publication lag lets us treat as complete.
- Momentum is cooling at the very top. Several of the leading assignees show sharp year-on-year declines in their latest reported year — a sign the leaders may be shifting filing strategy, not that the field is shrinking.
Filing growth compares 2021 (1,318 records) with 2024 (2,200) — 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 10,910 records in scope (CR5), not by the ranked leaders only.
What this landscape covers
This landscape covers 10,910 published patent records filed between 2015 and the 2026 data cut-off, drawn from filings that combine AI accelerator hardware language — neural processing unit, machine learning accelerator — with the surrounding software and semiconductor context: training datasets, model inference, feature vectors and process integration. It spans records filed with the United States, WIPO, the European Patent Office, India and Germany, among other offices.
The dataset mixes chip-architecture claims with the model-level and training-pipeline claims that increasingly ride alongside them, so it captures both the silicon and the workloads it is built to run.
Filing trends and technology composition
Two views of the same 10,910 records: how filing activity has moved year over year, and which technical branches carry the claim density.
Filing trend, 2017–2026
Filings climbed from 98 in 2017 to a peak of 2,200 in 2024, with 2021's 1,318 rising to that 2024 peak — a +67% increase over three years. 2025 and 2026 figures (339 so far in 2026) will keep revising upward as publication catches up with filing, typically an 18-month lag, so treat the most recent two years as undercounts rather than a slowdown.
Technology composition by IPC subclass
G06N (AI-model computing) leads at 37.8% of the 10,910 records, followed by G06F (digital data processing) at 31.0% and G06T (image processing) at 14.9%. Because records can carry multiple IPC classes, these shares sum to well over 100% — they show where claim density concentrates, not a partition of the field.
Shares are the percentage of the 10,910 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Go deeper on AI Accelerators Patent Landscape with Eureka
This page is one run against one query. Ask Eureka your own question about ai accelerators patent landscape and every answer comes back with the patent numbers behind it.
Try EurekaRepresentative and most-cited filings
Image processing method and neural processing unit for region-specific blurring
An image processing method performed by a neural processing unit is disclosed. The method includes receiving an input image including at least one object and processing the input image using a first model, via the neural processing unit, to detect a particular object among the at least one object — the first model being an artificial neural network-based object detector. A second model, also run on the neural processing unit, is trained to blur the region corresponding to that detected object.Filed by DEEPX CO., LTD., published 2026-07-09 as US20260196034A1.
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US20070192863A1 | Systems and methods for processing data flows | 961 |
| 2 | US20120240185A1 | Systems and methods for processing data flows | 864 |
| 3 | US20220126864A1 | Autonomous vehicle system | 677 |
| 4 | US20080229415A1 | Systems and methods for processing data flows | 637 |
| 5 | US20110238855A1 | Processing data flows with a data flow processor | 618 |
| 6 | US20080262990A1 | Systems and methods for processing data flows | 441 |
| 7 | US20080262991A1 | Systems and methods for processing data flows | 433 |
| 8 | US8135657B2 | Systems and methods for processing data flows | 426 |
| 9 | US8402540B2 | Systems and methods for processing data flows | 417 |
| 10 | US20110214157A1 | Securing a network with data flow processing | 414 |
Citation counts favour older records simply because they have had more time to be cited — read them as a signal of influence within this corpus, not as a ranking of current technical importance.
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 should inform where a new filing is likely to land relative to existing claim space.
The field is top-heavy
Just five assignees account for 6,123 of the 10,910 records in scope, and the top ten extend that to 63.4%. A long tail of single- and few-filing entrants fills the remainder, which means most freedom-to-operate analysis should start with the leaders' portfolios rather than the tail.
Growth is real, but the latest years understate it
The three-year rise from 1,318 filings in 2021 to 2,200 in 2024 marks the last period publication lag lets us call complete. 2025 and 2026 numbers will rise as filings still in the pipeline get published, so current-year totals should not be read as a plateau.
Model-level computing claims dominate the classification mix
G06N and G06F together cover the bulk of filings, ahead of the image-processing classes G06T and G06V. This points to claim density sitting as much in the AI-model and data-processing layer as in accelerator silicon itself — a useful check before assuming a chip-architecture claim is novel.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to ai accelerators patent landscape, with the prior art for and against each one.
Who holds the claim space, and where it is thinning
The leading assignees hold the bulk of granted and pending claims, but their most recent filing activity is cooling sharply, and several adjacent sub-areas remain lightly claimed.
One filer well ahead of the field
The leading assignee's 2,509 records dwarf the fifth-place total of 286 and the tenth-place total of 142 — a steep drop-off rather than a gradual one, typical of a field anchored by one dominant platform holder.
Leaders are pulling back in the most recent reported year
Every leading assignee tracked shows a sharp year-on-year decline in its latest reported year, ranging from roughly -74% to -98%. Given the 18-month publication lag, this likely reflects incomplete recent data rather than an actual retreat from the technology — but it is worth re-checking once later years fill in.
Automotive OEM pairing stands out
The strongest co-assignee pairing in the dataset links two automotive manufacturers on 148 shared records, far ahead of the next pairings. That points to joint accelerator-related filings tied to autonomous-driving or in-vehicle inference work rather than general-purpose chip design.
| Assignee | Recent year | YoY |
|---|---|---|
| Qualcomm Inc | 63 | -74% |
| Samsung Electronics Co Ltd | 47 | -85% |
| Nvidia Corp | 11 | -85% |
| Huawei Technologies Co Ltd | 5 | -81% |
| Shopify Inc | 4 | -83% |
| Intel Corp | 2 | -98% |
| Semiconductor Energy Laboratory Co Ltd | 2 | -99% |
| Microsoft Technology Licensing LLC | 1 | -98% |
Where to take this analysis
The figures above establish the shape of the field. Turning that into a filing or freedom-to-operate decision means going deeper on specific claims and specific competitors.
Check freedom to operate against the leaders first
With 56.1% of records held by five assignees, any new filing in accelerator architecture or model-inference claims should be checked against their portfolios before the long tail.
Explore assignee portfolios in Eureka →Watch the under-claimed branches
Sub-areas like region-specific inference processing and wireless-integrated accelerator scheduling show thinner claim density than the core classes — early filings there face less prior art.
Run a white space search in Eureka →Re-run the trend once 2025-2026 data settles
Because publication lags filing by roughly 18 months, the apparent pullback in 2025-2026 filings needs revisiting once those years are fully published.
Track filing trends live in Eureka →Common questions about AI accelerator patents
The assignee ranking in this dataset is led by a single company holding 2,509 of the 10,910 records in scope, well ahead of the fifth-ranked assignee at 286 and the tenth-ranked at 142. The top five assignees combined hold 56.1% of all records, and the top ten hold 63.4%, so the field is concentrated rather than evenly spread. Beyond the leaders, a long tail of companies each hold a small number of filings, which is typical of a fast-growing hardware-and-software field still attracting new entrants.
Filings rose from 1,318 in 2021 to a peak of 2,200 in 2024, a growth of 67% over that three-year span, and 2024 is the most recent year that can be treated as complete. Figures for 2025 and 2026 appear lower, but that reflects the roughly 18-month lag between filing and publication rather than an actual slowdown. Anyone reading a recent-year dip in this space should wait for later publication rather than concluding the technology has plateaued.
Classification data shows G06N (computing arrangements based on AI models) as the largest category at 37.8% of the 10,910 records, followed by G06F (electric digital data processing) at 31.0% and G06T (image data processing) at 14.9%. Because a single patent can carry multiple IPC classes, these percentages overlap rather than summing to 100%. The pattern indicates that claim density sits heavily in model-level computing and data processing, not solely in chip architecture.
Several adjacent sub-areas show comparatively thin claim density relative to the core accelerator and model-computing classes, including region-specific inference processing, wireless-network-integrated accelerator scheduling, and video-codec-coupled neural processing. These are not guaranteed to be open — they simply show less concentrated activity in this dataset than the dominant G06N and G06F categories. A proper freedom-to-operate search against the specific claim language is still needed before filing.
Not necessarily. The most-cited records in this dataset date back over a decade, including filings from 2007 and 2008, and older patents accumulate citations simply by being available longer in the corpus that gets searched. High citation counts are a useful signal of historical influence on later filings, but they should not be read as a ranking of current technical importance or commercial relevance.
Research AI Accelerators Patent Landscape in depth with Eureka
Go past this page: query the whole ai accelerators 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.