Mixture-of-Experts Routing Patents: Who Leads, Where Gaps Are 2026
- Flat, not booming. filings ran 737 in 2017, peaked at 777 in 2021, and sit at 694 by the 2022 midpoint — this is a mature, saturated filing pattern, not a hockey stick.
- No single gatekeeper. the top 5 assignees together hold just 4.4% of all 899,711 records in scope, and the top 10 only 6.9% — concentration is low across a very long tail.
- Classification is scattered. the leading IPC subclass, H04L, covers only 0.5% of records, meaning routing and load-balancing claims are dispersed across networking, compute and even chemistry-adjacent classes rather than sitting in one bucket.
Filing growth compares 2021 (777 records) with 2024 (523) — 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 899,711 records in scope (CR5), not by the ranked leaders only.
What this landscape covers
This dataset tracks patent activity around mixture-of-experts routing, load balancing and inference serving — the mechanisms that decide which compute node, server or expert sub-model handles a given request. The search spans terms from classic distributed-systems load balancing through to token-level routing and expert placement, which means the corpus mixes decades-old server-farm patents with recent AI-inference filings under the same technical umbrella. Coverage runs from 2015 through the 2026-07-31 cut-off, with 899,711 records in scope.
Because publication lags filing by roughly 18 months, the 2026 figure in any trend line is necessarily incomplete and should be read as a floor, not a ceiling. The most-cited records in this corpus date back to the late 1990s web-server era, which says more about how long a foundational load-balancing patent accumulates citations than about which techniques matter today.
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 899,711-record corpus: how filing volume has moved year over year, and which IPC subclasses carry the routing and load-balancing claims.
A flat filing curve since 2017
Volume moved from 737 filings in 2017 to a peak of 777 in 2021, then eased to 694 by 2022 — a plateau rather than sustained growth, before the incomplete 2026 count of 30 comes in below trend as expected from publication lag.
Claims spread thin across IPC subclasses
H04L (digital information transmission) leads at 0.5% of all records, with G06F and H04W following at 0.4% and 0.2%. No single subclass concentrates the field, and several unrelated classes — B01D, C07C, A61K, H01M — appear in the top ranks, reflecting the search terms' overlap with chemical and battery-related "load balancing" and "mixture" usage outside computing.
Shares are the percentage of the 899,711 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Go deeper on Mixture-of-Experts Routing and Serving with Eureka
This page is one run against one query. Ask Eureka your own question about mixture-of-experts routing and serving and every answer comes back with the patent numbers behind it.
Try EurekaMost-cited prior art and a recent filing
US20250373684A1 — Load balancing method and system for providing AI service
A load balancing method in an AI service providing system that obtains load balancing information from a plurality of servers, builds a load balancing table, receives an inference task request from a user device, derives a target server using the table, and performs load balancing for that inference task based on a preset algorithm.Filed by REBELLIONS INC., published 2025-12-04 — illustrates how classic server load-balancing claim structure is being re-applied directly to AI inference serving.
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US5774660A | World-wide-web server with delayed resource-binding for resource-based load balancing on a distributed resour… | 2,303 |
| 2 | US20030018927A1 | High-availability cluster virtual server system | 2,036 |
| 3 | US5774668A | System for on-line service in which gateway computer uses service map which includes loading condition of ser… | 1,733 |
| 4 | US6351775B1 | Loading balancing across servers in a computer network | 997 |
| 5 | US6195680B1 | Client-based dynamic switching of streaming servers for fault-tolerance and load balancing | 858 |
| 6 | US4935326A | Electrophotographic carrier particles coated with polymer mixture | 834 |
| 7 | US20090300210A1 | Methods and systems for load balancing in cloud-based networks | 790 |
| 8 | US5672097A | Abrasive article for finishing | 777 |
| 9 | EP0363608A1 | Polymer mixture comprising an aromatic polycarbonate, a styrene-containing copolymer and/or graft polymer and… | 743 |
| 10 | US5679526A | Threshold ligand-receptor assay | 712 |
Citation counts favour older filings simply because they have had longer to accumulate citations inside this searched corpus — treat them as a signal of foundational influence, not of current commercial 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 signals stand out once the filing trend, IPC spread and citation pattern are read together.
Volume has plateaued
Filings rose from 737 in 2017 to a peak of 777 in 2021 before easing to 694 in 2022. That is a flat-to-declining curve, not an accelerating one — the field is not in a land-grab phase, and a new entrant is not racing a rising tide of filings.
No dominant gatekeeper
The five largest assignees together account for only 4.4% of all records in scope, and the ranked leaders' top 10 reach just 6.9%. Ownership of core routing and load-balancing claims is fragmented across a long tail rather than locked up by a handful of firms.
Claims sit outside a single technical home
The leading subclass, H04L, covers only 0.5% of the 899,711 records, with G06F, H04W and several non-computing classes close behind. Prior-art search for a new routing claim needs to cross multiple IPC subclasses rather than assume one dominant classification.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to mixture-of-experts routing and serving, with the prior art for and against each one.
| Assignee | Co-assignee | Shared families |
|---|---|---|
| International Business Machines Corporation (IBM) | IBM UNITED KINGDOM LTD INTELLECTUAL PROPERTY DEPARTMENT | 13 |
| Citrix Systems, Inc. | SUGANTHI JOSEPHINE | 9 |
| Citrix Systems, Inc. | RAJA MURALI | 9 |
| Citrix Systems, Inc. | VERZUNOV SERGEY | 7 |
| Citrix Systems, Inc. | KAMATH SANDEEP | 7 |
| Citrix Systems, Inc. | SHETTY ANIL | 5 |
| Citrix Systems, Inc. | NARAYANA RAGHAV SOMANAHALLI | 5 |
| Qualcomm Incorporated | CHENG PENG | 5 |
Only 10 co-assignee pairs appear in the corpus, the strongest linking IBM's US and UK IP entities — cross-company collaboration on these claims is the exception, not the norm.
Assignee landscape and recent momentum
The assignee ranking covers 100 companies counted in records — the full set the data endpoint returns, not a curated top list. Recent-year momentum across the largest names is negative across the board.
A fragmented lead position
The top-ranked assignee holds 16,034 records, well ahead of fifth place at 5,434 and tenth place at 3,644 — a real gap exists at the very top, but it does not translate into field-wide control given the top 5's 4.4% share of all records.
Recent-year filings are cooling across the board
Every tracked leading assignee shows flat or negative year-over-year momentum in the latest year, several dropping to zero filings. This is consistent with the plateaued overall filing trend rather than a single company retreating.
Filing is mostly solo
With only 10 co-assignee pairs identified across the entire corpus, joint filings are rare. The strongest pairing links a single company's domestic and international IP-holding entities rather than two independent firms.
| Assignee | Recent year | YoY |
|---|---|---|
| Huawei Technologies Co., Ltd. | 2 | -60% |
| Cisco Technology, Inc. | 1 | -67% |
| International Business Machines Corporation (IBM) | 0 | -100% |
| Telefonaktiebolaget LM Ericsson | 0 | -100% |
| BASF SE | 0 | -100% |
| Microsoft Corporation | 0 | — |
| Citrix Systems, Inc. | 0 | — |
| Qualcomm Incorporated | 0 | -100% |
Where to take this analysis
The landscape data points to specific next steps depending on what you are trying to protect or clear.
Run a freedom-to-operate check on token-routing claims
With ownership this fragmented, a proposed routing or expert-placement claim needs checking against a long tail of small assignees, not just the ranked leaders.
Explore in EurekaMap claim language across IPC subclasses
Because relevant claims spread across H04L, G06F, H04W and beyond, a single-class search will under-count prior art for any new filing in this space.
Explore in EurekaWatch the under-claimed branches before they fill in
Token-dropping mitigation and memory-tier-aware placement show thin coverage today; that gap narrows as inference-serving filings continue.
Explore in EurekaCommon questions on this landscape
The leading assignee in this corpus holds 16,034 records, compared with 5,434 at fifth place and 3,644 at tenth. However, the top 5 assignees combined account for only 4.4% of all 899,711 records in scope, and the top 10 reach just 6.9%. That means even the largest holder controls a small slice of a very fragmented field, so a single-company clearance check is not sufficient for freedom-to-operate purposes.
No — filings ran at 737 in 2017, peaked at 777 in 2021, and had already eased to 694 by the 2022 midpoint, a flat-to-declining pattern rather than sustained growth. The 2026 count of 30 looks low but that is expected: publication lags filing by roughly 18 months, so the most recent year always understates true activity. Read the trend through 2022-2023 as the more reliable recent signal.
No single IPC subclass dominates. H04L (digital information transmission) leads but covers only 0.5% of the 899,711 records, with G06F (electric digital data processing) at 0.4% and H04W (wireless communication networks) at 0.2% close behind. Because a record can carry several classes, a prior-art search limited to one subclass will miss a meaningful share of relevant filings — search across at least H04L, G06F and H04W together.
Filed by REBELLIONS INC. and published 2025-12-04, this filing describes a load balancing method for an AI service system: it builds a load balancing table from server load information, receives an inference task request, derives a target server from the table, and applies a preset load balancing algorithm to route the task. It is a direct re-application of classic server load-balancing claim structure to AI inference serving rather than a novel routing algorithm, which narrows what it can block to that specific table-and-derivation architecture.
The corpus shows thin representation in token-dropping mitigation at expert capacity limits, dynamic expert placement across heterogeneous memory tiers, routing-stability guarantees under load variance, and latency-variance-aware inference scheduling. These sit adjacent to the heavily cited classic load-balancing claims but are not yet crowded, based on the current IPC and filing composition. As inference-serving patenting continues, these branches are the ones most likely to fill in first.
Research Mixture-of-Experts Routing and Serving in depth with Eureka
Go past this page: query the whole mixture-of-experts routing and serving 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.