Sparse MoE Routing Patents: Top Companies & Filing Trends 2026
- 56.1% of all 619 records sit with just five assignees, with one leader alone holding 143 filings.
- +2767% filing growth from 2021 (3 filings) to 2024 (86 filings), the most recent year that can be read as complete.
- Filings span eight IPC subclasses from core AI computing (G06N, 34.2%) into healthcare informatics and body-fluid devices — routing logic is migrating into domain-specific hardware.
Filing growth compares 2021 (3 records) with 2024 (86) — 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 619 records in scope (CR5), not by the ranked leaders only.
What the sparse MoE routing filing record shows
Sparse mixture-of-experts routing sits at the centre of how foundation models scale compute without scaling cost proportionally. The 619 records in scope span 2015 through the mid-2026 cut-off, but the shape of the field only becomes visible after 2021: filings grew from 3 in 2021 to 86 in 2024, a run documented in this dataset at +2767% over that three-year span. That is a field moving from research curiosity to contested claim territory in the space of three filing years.
Because publication trails filing by roughly 18 months, the 2025 and 2026 figures in the trend chart are still filling in and should not be read as a slowdown. The concentration figures are the more durable signal: 56.1% of all 619 records belong to five assignees, and 68.7% belong to ten. That is a field with a defined set of incumbents rather than one still being staked out by newcomers.
Let an AI agent run this analysis on your own technology
Pick a task. Every answer cites the patents behind it.
Filing trend and technology composition
Two views of the same 619 records: how filing activity moved year over year, and which IPC subclasses the claims actually sit in.
Filings accelerated sharply after 2021
From a single filing in 2017, activity stayed low through 2020 before climbing to 3 in 2021 and 86 in 2024 — the +2767% span the hero figure reflects. 2025 shows 380 filings and 2026 shows 23 so far, but both years are still incomplete in the publication record and should be read as provisional rather than as a peak.
Routing claims cluster in core AI computing but reach into hardware domains
G06F (electric digital data processing) covers 38.0% of the 619 records and G06N (AI-model computing) covers 34.2%, the expected core for routing logic. But G16H (healthcare informatics, 10.3%) and A61M (body-fluid devices, 9.5%) both appear at meaningful share, showing sparse routing architectures being claimed inside domain-specific medical and imaging hardware rather than staying confined to general-purpose compute.
Shares are the percentage of the 619 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Go deeper on Foundation Models: Sparse MoE Routing Patent Landscape with Eureka
This page is one run against one query. Ask Eureka your own question about foundation models: sparse moe routing patent landscape and every answer comes back with the patent numbers behind it.
Try EurekaA representative filing and the most-cited prior art
Sparse encoding and decoding at mixture-of-experts layer
A computing system executes a Mixture-of-Experts (MoE) layer by receiving an input tensor of input tokens, computing a gating function output vector from that tensor, and computing a sparse encoding of the tensor and gating output. The sparse encoding identifies one or more destination expert sub-models. The system dispatches the input tensor to those destination experts for processing and computes an expert output in response.Filed by Microsoft Technology Licensing, published 2024-03-14. This is the clearest single claim in the dataset to the routing step itself — the gating computation plus sparse-encoding dispatch — rather than to a downstream application of MoE.
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US20200067861A1 | SCAM evaluation system | 475 |
| 2 | US20140310243A1 | Heart beacon cycle | 337 |
| 3 | US20250259085A1 | Convergent Intelligence Fabric for Multi-Domain Orchestration of Distributed Agents with Hierarchical Memory … | 213 |
| 4 | WO2021220008A1 | Image compression and decoding, video compression and decoding: methods and systems | 97 |
| 5 | WO2024155584A1 | Systems, methods, devices, and platforms for industrial internet of things | 95 |
| 6 | US20230154055A1 | Image compression and decoding, video compression and decoding: methods and systems | 78 |
| 7 | US20260046317A1 | AI-Optimized Memory Fabric for Large Contexts and Multimodal Workloads | 67 |
| 8 | US20130135008A1 | Method and system for a run-time reconfigurable computer architecture | 67 |
| 9 | WO2024226801A2 | Systems, methods, kits, and apparatuses for generative artificial intelligence, graphical neural networks, tr… | 66 |
| 10 | US20250259041A1 | Ai agent decision platform with deontic reasoning | 62 |
Citation counts favour older, earlier-filed documents inside this corpus — read them as a measure of influence on later filings, not as a signal of which patents matter most today.
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 numbers mean for a filing decision
Three read-throughs of the concentration, growth and classification figures above, aimed at where a new filing would actually land.
The core routing claim space is already staked out
Five assignees hold 347 of the 619 records in scope, with the leader alone accounting for 143. A new entrant filing on core gating-and-dispatch mechanics is filing into dense prior art held by a small number of parties, not into open ground.
The field only became a filing priority in the last three complete years
Filings sat at single digits through 2020 and only reached 3 in 2021. By 2024 that had grown to 86 — the sharpest sustained climb in the dataset. Anyone benchmarking against older MoE literature is comparing against a much thinner prior-art base than exists today.
Routing logic is leaking into domain hardware, not staying in general compute
G16H (healthcare informatics) and A61M (body-fluid devices) together touch roughly a fifth of the records, alongside the expected G06F/G06N core. That is a sign that sparse routing is being claimed as a component inside medical and sensing systems, a track separate from the foundation-model core.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to foundation models: sparse moe routing patent landscape, with the prior art for and against each one.
Who is filing, and where the activity is thinning out
The ranking behind this page covers 100 companies, counted in records — not a top-50 or top-100 filtered list, but the full set the data endpoint returns for this search.
One assignee holds close to a quarter of all records
The leading assignee's 143 filings sit well above the fifth-place figure of 21, meaning the gap between first and the rest of the leaderboard is wide rather than gradual.
The drop-off from top five to top ten is steep
Tenth place holds 12 records against the fifth-place figure of 21, and the top ten combined reach 68.7% of all 619 records. Past that point the ranking thins quickly into a long tail of smaller filers.
Even active recent filers are showing sharp swings, not steady growth
One of the more recently active assignees moved from a stronger prior year to 8 filings in the latest year, an 85% year-on-year drop. Several other assignees in the recent-momentum data show filings falling to zero. Read these latest-year figures cautiously given the 18-month publication lag.
| Assignee | Recent year | YoY |
|---|---|---|
| DATADOSE LLC | 8 | -85% |
| Microsoft Technology Licensing, LLC | 1 | — |
| Innatera Nanosystems B.V. | 1 | 0% |
| Intel Corporation | 0 | -100% |
| SILVEBROOK KIA | 0 | -100% |
| Dipu Rende Technology | 0 | -100% |
| Strong Force VCN Portfolio 2019, LLC | 0 | -100% |
| STRONG FORCE TX PORTFOLIO 2018 LLC | 0 | -100% |
Where to take this next
The figures above describe the field as filed. Turning that into a filing or freedom-to-operate decision means going claim by claim.
Map your own architecture against the leader's claim set
With 143 records from a single assignee, a new gating-and-dispatch filing needs a clause-level comparison against that portfolio before drafting, not just a keyword search.
Run a claim comparison in EurekaCheck the under-claimed branches before committing scope
Fault-tolerant routing and cross-modal gating show thinner coverage in this dataset than core dispatch mechanics — worth confirming with a fresh search before scoping a new application.
Explore white space in EurekaCommon questions about sparse MoE routing patents
One assignee leads this dataset with 143 records out of 619 in scope, well ahead of the fifth-place holder at 21 records. The top five assignees combined hold 56.1% of all records, and the top ten hold 68.7%. That means the field has a small set of incumbents with substantial portfolios rather than activity spread evenly across many filers, so a freedom-to-operate review should prioritise the leading holders first.
Filings grew from 3 in 2021 to 86 in 2024, a rise the dataset documents as +2767% over that three-year span. This is the clearest complete-year trend available, since publication typically lags actual filing by around 18 months, meaning the 2025 and 2026 figures in the raw trend data are still being filled in and understate real activity. Treat 2024 as the most recent year you can compare against with confidence.
The largest classes are G06F (electric digital data processing, 38.0% of the 619 records) and G06N (computing based on AI models, 34.2%), which together form the expected core for routing and gating logic. Beyond that core, G06T (image data processing, 20.5%), G16H (healthcare informatics, 10.3%) and A61M (body-fluid devices, 9.5%) show routing concepts being claimed inside domain-specific hardware and imaging systems. Since records can carry multiple classes, these shares add up to more than 100% and should not be summed.
This Microsoft filing, published 2024-03-14, claims a computing system that executes an MoE layer by computing a gating function output from an input tensor, deriving a sparse encoding that identifies destination expert sub-models, and dispatching tokens to those experts for processing. It is a claim on the routing mechanism itself — the gating-plus-sparse-encoding-plus-dispatch sequence — rather than on any particular downstream application of mixture-of-experts. Anyone building a gating function that selects experts from a sparse encoding of an input tensor should review this filing's claim scope directly before finalising an architecture.
The IPC composition and citation data in this dataset point to thinner coverage in fault-tolerant expert fallback, cross-modal or multi-domain gating, and quantisation-aware dispatch, compared with the dense coverage of core gating-and-dispatch mechanics. These are not claims that no one has filed, but they show up less densely than the core routing claims held by the leading assignees. A first claim in one of these branches would need to specify the mechanism precisely — for example, how fallback routing behaves when a selected expert exceeds capacity — rather than restating general MoE dispatch.
Research Foundation Models: Sparse MoE Routing Patent Landscape in depth with Eureka
Go past this page: query the whole foundation models: sparse moe routing 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.