Deep Learning Compiler Stacks Patents: Top Companies & Trends 2026
Deep Learning Compiler Stacks Patents: Who Leads and Where the Claims Sit
Concentrated but not locked. The top 5 assignees hold 33.5% of all 4,628 records in scope, and the top 10 hold 45.2% — leaving over half the field to a long tail of single- and few-filing entrants. Filings rose from 154 in 2017 to a peak of 322 in 2025, with 2021's 225 climbing to 253 by 2024 — a +12% gain over that three-year span, the last window unaffected by publication lag. The 2026 figure of 62 is a partial year and should not be read as a decline.
- 1MATHWORKS INC10.2%
- 2INTERNATIONAL BUSINESS MACHINE CORPORATION8.5%
- 3INTEL CORP5.7%
- 4MICROSOFT TECHNOLOGY LICENSING LLC5.4%
- 5NVIDIA CORP3.7%
See the full deep learning compiler stacks analysis in Eureka
- The complete ranking, not just the top five
- Every IPC branch with its share of the corpus
- The most-cited records, and where claim space is still thin
Common questions about this landscape
Who holds the most patents in deep learning compiler technology?
One assignee leads the ranked field with 473 records, well ahead of the fifth-place holder at 172. The top 5 assignees combined account for 33.5% of all 4,628 records in scope, and the top 10 account for 45.2%. That still leaves more than half the landscape spread across the remaining ranked companies and unranked filers, so the field is concentrated at the top without being closed.
Is patent filing for deep learning compilers still growing?
Yes, through the last complete filing year. Filings rose from 225 in 2021 to 253 in 2024, a +12% increase, and 2025 recorded the highest count in the dataset at 322. Figures for 2025 and especially 2026 are still incomplete because publication typically lags filing by around 18 months, so treat the most recent one to two years as a floor rather than a final number.
What does ‘operator fusion’ cover in these patent claims?
Operator fusion refers to combining multiple computation steps in a neural network graph into a single executable kernel, reducing memory traffic and execution overhead. It is one of the core technical terms used to define this landscape alongside intermediate representation, layout transformation and graph lowering. Claims in this area tend to sit within the broader G06F electric digital data processing class, which covers 89.6% of the 4,628 records in scope.
Disclaimer. This analysis is based on Patsnap Eureka data drawn from a limited snapshot of global patent records and is provided for general information and reference only. Patent data carries inherent limitations — recent filings are under-counted because of publication lag, counts may be on a record or family basis, classification and applicant-name data may contain errors or duplicates, and the underlying search query defines the scope shown — so the analysis may be incomplete or inaccurate and may not reflect the full technology landscape.
Nothing here is an exhaustive prior-art, novelty, freedom-to-operate or validity search, nor does it constitute legal, financial or professional advice, and it should not be relied upon as such. Verify independently and review with qualified patent and legal professionals before acting on it.
Method: Filing trend and technology composition. Derived from a Patsnap search on Deep Learning Compiler Stacks covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish. Every share divides by all records in scope. Data: Patsnap Eureka. See the full landscape report.