Large Language Model Serving Patent Landscape 2026
Large Language Model Serving Patent Landscape in 2026
The large language model serving patent space is nascent and highly concentrated, with virtually all activity compressed into 2024–2025 and a small group of hardware and cloud specialists holding the dominant positions. DeepX leads the field, and the United States is the primary filing jurisdiction, signaling an early-stage land-grab by inference hardware and cloud infrastructure players.
DeepX leads a tightly held, early-stage field
DeepX Co Ltd holds the top position in this corpus, followed by Salesforce and Microsoft Technology Licensing, each with two patent families. The field is nascent, with the vast majority of filings appearing only from 2024 onward.
The top five filers account for 53% of the combined output of the hundred largest filers, indicating a notably concentrated competitive structure at this early stage. The gap between the leader and the second tier is already visible.
| # | Applicant | Patent families | Share |
|---|---|---|---|
| 1 | DeepX Co Ltd | 4 | |
| 2 | Salesforce Inc | 2 | |
| 3 | Microsoft Technology Licensing LLC | 2 | |
| 4 | Huawei Cloud Computing Technologies Co Ltd | 1 | |
| 5 | Beijing Shenzhou Digital Cloud Computing Co Ltd | 1 | |
| 6 | Soteria Inc | 1 | |
| 7 | Groq Inc | 1 |
| # | Applicant | Patent families | Share |
|---|---|---|---|
| 8 | Jiangnan University | 1 | |
| 9 | GDM Holding LLC | 1 | |
| 10 | SOGANG UNIV RES & BUSINESS DEV FOUND | 1 | |
| 11 | Fudan University | 1 | |
| 12 | Intel Corporation | 1 | |
| 13 | Taiping Technology Co Ltd | 1 | |
| 14 | CentML AI Inc | 1 |
The leaders’ positions reflect a mix of inference hardware specialists (DeepX, Groq), hyperscale cloud providers (Huawei Cloud, Microsoft), and enterprise software platforms (Salesforce), suggesting that serving-layer IP is being built simultaneously from the chip up and the application down.
Filings from the most recent 18–24 months are subject to publication lag and are likely undercounted; the corpus should be treated as a floor, not a ceiling, for recent activity. Longer-window growth, applicant concentration, and technology-route coverage are therefore more reliable signals than the latest-year bar alone.
Explosive emergence in 2024–2025 anchored in AI-model computing
The annual filing trend and the technology branch composition together reveal a field that barely existed before 2024 and is overwhelmingly focused on AI-model computing foundations.
Annual filing trend
Recorded filings were zero across 2017–2023, then jumped sharply in 2024 and again in 2025. The 2026 figure reflects only the earliest published applications and will grow substantially as the publication pipeline clears — treat the recent uptick as a lower bound.
↗ Hover for values · click a bar to ask EurekaTechnology composition
G06N (Computing based on AI models) dominates the branch mix, with G06F (Electric digital data processing) as a secondary but meaningful strand covering system-level and data-handling aspects. G06Q (Business, commerce and admin data processing) is represented by a single patent family, making it the most sparsely covered branch.
↗ Hover for values · click a bar to ask EurekaHighly cited patent families surfaced by the query
Citation-heavy patent families returned by the query. Use this section as citation context, not as a curated list of the most topic-specific patents.
Processor-implemented methods and systems for mode…
Processor-implemented methods and systems are disclosed for optimizing pre-existing large language models for model inference. Different types of optimization techniques are provided to an offline optimization program. Within a generic model framework, different combinations of large language model serving configurations are generated. An automated online… (excerpt from the patent abstract)
| # | Patent | Citations |
|---|---|---|
| 1 | 一种基于低碳贡献反馈的强化学习绿色大模型训练方法 | 1 |
| 2 | Ai system comprising a plurality of ai modules wit… | 1 |
| 3 | Systems and methods for a reasoning-intensive rera… | 1 |
| 4 | 基于参数高效微调的大型语言模型训练方法 | 1 |
Ranked by total forward citations. Citation counts favour older and broadly cited patent families, and broad or adjacent patents may appear when they match the search scope. Treat this section as citation context, not as a curated list of the most topic-specific patents. Some patent titles may be shown in their original, non-English language where an accurate translation could not be guaranteed.
What the structure means for R&D investment
The combination of a nascent lifecycle, high concentration, absence of recorded co-filing partnerships, and US-centric jurisdiction points to a field where first-mover IP positions are still being established and entry windows remain open.
Emerging field with no established lifecycle stage confirmed
The lifecycle stage cannot be formally assigned from current evidence. However, the filing timeline — zero activity before 2024, then a sharp rise — is consistent with a field at or near inception. R&D teams should expect rapid claim-space evolution and frequent Freedom-to-Operate reassessments as the corpus grows.
Early-stageTop five filers hold majority share among leading applicants
The top five applicants account for 53% of the combined output of the hundred largest filers. With only 18 patent families in scope in total, individual filings carry outsized weight. A single new entrant with a focused filing program could meaningfully shift the competitive balance within one to two years.
High concentrationNo co-applicant partnerships recorded
The collaboration evidence shows no co-filed patent families in this corpus. This is consistent with the field’s very early stage, where players are still defining proprietary positions rather than pooling resources. The absence of consortia or joint filings leaves open the possibility of ecosystem-building through licensing rather than co-development agreements.
Solo filers onlyUnited States leads; China and PCT routes provide secondary coverage
The United States is the leading jurisdiction, followed by China and WIPO PCT filings, with Europe and South Korea each registering limited coverage. The PCT route’s presence alongside national filings suggests some filers are pursuing international protection early, but the European footprint remains thin — a gap that could matter as EU AI regulation tightens.
US-centricGo beyond the landscape: Eureka’s TRIZ Solution agent breaks down an R&D problem and returns patented concept solutions, each with a technical approach and cited patent & literature evidence.
Co-filing pairs, ranked by the number of jointly-filed patent families.
DeepX anchors hardware inference; Microsoft and Salesforce represent software serving
The top two tiers reflect structurally different approaches: dedicated inference silicon (DeepX, Groq) versus cloud and enterprise software platforms (Microsoft, Salesforce, Huawei Cloud). New academic and startup entrants signal broadening participation.
DeepX Co Ltd
DeepX holds 4 patent families — the largest single position in this corpus — concentrated in G06N 3 (neural-network computing) with a secondary G06F 17 (digital data processing) component. This profile is consistent with on-device and edge inference hardware. Momentum data is not available for DeepX in the applicant momentum evidence, but its lead position was established within the 2024–2025 emergence window.
families: 4Salesforce Inc
Salesforce holds 2 patent families, tying with Microsoft Technology Licensing for second place. Its filings reflect enterprise software serving use cases. Huawei Cloud Computing Technologies, Groq, and CentML AI each hold 1 patent family; Huawei Cloud, Fudan University, and Salesforce’s counterpart entity (Shuodongli / CentML) are all flagged as new entrants in the most recent filing period, indicating the challenger tier is actively forming.
families: 2| Applicant | Recent (3 yrs) | Trend |
|---|---|---|
| CentML AI Inc | 1 | ▲ new entrant |
| Fudan University | 1 | ▲ new entrant |
| Huawei Cloud Computing Technologies Co Ltd | 1 | ▲ new entrant |
Business-process serving is an under-served adjacent branch
The technology composition reveals one IPC branch that is notably sparse relative to the dominant AI-computing cluster, representing an area where the current patent corpus offers limited coverage.
G06Q · Business, commerce and admin data processing
With only 1 patent family recorded under G06Q — a 3% share among the top-100 filers’ combined output — business-process and commercial-workflow integration of LLM serving is the least-covered branch in this corpus. The technical gap is plausible: applying serving-layer optimizations (batching, caching, routing) directly to enterprise transaction and decision workflows requires domain-specific claim strategies that go beyond pure G06N coverage. Filers with enterprise software expertise (e.g., CRM, ERP, financial services) may find fewer blocking positions here than in the core AI-model branch.
Search this in Eureka →G06F · System-level inference infrastructure beyond AI models
G06F appears as a secondary branch with 10 patent-record occurrences, but its sub-classes span a wide range — from operating-system scheduling (G06F 9) to fault tolerance (G06F 11) and data retrieval (G06F 16, 18). Coverage is uneven: Groq’s filings concentrate on arithmetic and data-format units (G06F 5, 7, 17), while most other applicants touch G06F only incidentally. System-level serving concerns such as memory management, inter-node communication, and fault-tolerant inference at scale remain relatively sparse, suggesting entry room for players with operating-systems or distributed-systems expertise.
Search this in Eureka →How leaders differ by technology route
Strength of each leader across the main technology routes.
| Player | G06N 3 · Computing based on AI models | G06F 9 · Electric digital data processing | G06N 20 · Computing based on AI models | G06F 11 · Electric digital data processing | G06F 16 · Electric digital data processing |
|---|---|---|---|---|---|
| DeepX Co Ltd | Strong · 4 | Absent | Absent | Absent | Absent |
| Microsoft Technology Licensing LLC | Strong · 2 | Moderate · 1 | Absent | Absent | Absent |
| Sogang University Research & Business Development Foundation | Strong · 1 | Strong · 1 | Absent | Strong · 1 | Absent |
| Soteria Inc | Strong · 1 | Strong · 1 | Absent | Strong · 1 | Absent |
| GDM Holding LLC | Strong · 1 | Absent | Strong · 1 | Absent | Absent |
| Beijing Shenzhou Digital Cloud Computing Co Ltd | Strong · 1 | Absent | Absent | Strong · 1 | Absent |
Frequently asked questions
The current corpus contains 18 patent families in scope. Given that virtually all activity is concentrated in 2024–2025, this number is likely to grow as more recent filings complete the publication process.
DeepX Co Ltd leads with 4 patent families, focused on neural-network computing (G06N 3) and digital data processing (G06F 17). Salesforce and Microsoft Technology Licensing each hold 2 patent families in the next tier.
Recorded filings were zero from 2017 through 2023. Activity first appeared in 2024 and accelerated sharply in 2025. The 2026 figure is minimal due to publication lag and does not represent the true filing pace.
The United States leads with 8 patent records, followed by China and WIPO PCT routes with 4 each. Europe (EPO) and South Korea each have 1 record, indicating that international coverage outside the US and China is still limited.
G06N (Computing based on AI models) is the dominant branch, with G06F (Electric digital data processing) as a secondary strand. G06Q (Business, commerce and admin data processing) is represented by only 1 patent family, making it the most sparsely covered branch.
No co-applicant partnerships are recorded in the current evidence. All identified filings are single-applicant, which is consistent with the field’s very early stage where players are still staking out proprietary positions.
Ready to map your own LLM serving patent landscape?
Join 18,000+ innovators using PatSnap Eureka to map any technology landscape: search 2B+ patents and papers, surface key assignees, and generate a report like this in minutes.
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.