This article was first published on IAM

• Patent applications demonstrate a distinct technological focus, primarily distributed across cluster training, financial technology applications, and AI models
• DeepSeek’s early patents include applications in the financial systems domain, laying the groundwork for future expansion into other highly regulated fields
• The company’s equity structure is characterised by multiple entities, multiple levels, and decentralisation

Chinese AI startup DeepSeek this year surprised everyone when it publicly launched a series of AI innovations. On 20 January, DeepSeek released the new generation language model R1, which enhanced reasoning capabilities through reinforcement learning technology. The model achieved an accuracy rate of nearly 80% (79.8%) in the AIME 2024 mathematical benchmark test, surpassing 96.3% of human participants in the Codeforces evaluation, according to DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.

Its associated V3 model (released in December 2024) achieved performance similar to GPT-4o at a low cost of $5.58 million,the DeepSeek-V3 Technical Report shows.

DeepSeek's breakthroughs in the AI field have garnered global attention. Behind the technical and commercial success, what is the logic, current status, and future direction of DeepSeek's patent layout? We searched all 19 publicly disclosed patents of DeepSeek and its affiliated companies (by February 2025) at the CNIPA. We also analysed the evolution of patented technologies and patent filing strategies to outline the direction of DeepSeek's patent layout and provide an in-depth interpretation of the company's patent strategy.

We found that DeepSeek's patent portfolio exhibits the following five core characteristics, which will be explained in detail:

• highly-focused patent technologies;
• clear patent technology layout route;
• patent applications primarily based in China;
• complex equity structure and joint patent ownership arrangement; and
• gaps between patent technologies and open-source technologies.

Highly-focused patent technologies

Based on a technical analysis of DeepSeek's patents (as shown in the Figure 1 below), the company's patent applications demonstrate a distinct technological focus, primarily distributed across three branches: cluster training, financial technology applications, and AI models.

Specifically:

• Cluster training has been the central focus throughout the company's development trajectory, emphasising large-scale computing infrastructure optimisation.
• Financial technology applications began to take shape in 2019, initially concentrated on financial data processing and analysis.
• AI model-related patents have achieved significant progress in the past year, reflecting breakthroughs in artificial intelligence research and development.

Source: Patent search conducted in the database of “www.zhihuiya.com” as of 28 February 2025.

To further explore the secondary branches of DeepSeek's patented technology, the figure below provides detailed insights into the specific technical points of DeepSeek's patent technology layout.

Source: Patent searches conducted in the database of “www.zhihuiya.com” as of 28th February 2025.

Full-chain layout of cluster training technology (79% of the patent portfolio)

DeepSeek has positioned multiple core technologies in the field of cluster training, covering various interrelated aspects such as resource allocation, task scheduling, and communication optimisation.

• Resource allocation optimisation: Patent CN114780203A introduces a ‘preset seat mechanism’, decoupling the scheduling module from the container creation and deployment module. This mechanism enhances scheduling efficiency by forcibly removing interrupted and suspended tasks, addressing the issue of GPU resource wastage.
• Innovative task scheduling: Patent CN112925640B constructs node operation status scores based on historical error rates, dynamically sorting and prioritising high-priority nodes. Combined with a locking mechanism, this achieves adaptive cluster resource allocation, reducing failure rates and improving resource utilisation.
• Cluster communication enhancement: Patent CN118503194A employs primary/auxiliary GPU collaborative slicing, combined with NVLink and dynamic congestion control to achieve multi-network card RDMA parallel transmission, enhancing cluster network bandwidth utilisation and throughput performance.

Early technological extensions in financial scenarios (16% of the patent portfolio)

DeepSeek's early patents include three applications in the financial systems domain. The financial sector's high data sensitivity provides a ‘stress test’ environment for technology validation, laying the groundwork for future expansion into other highly regulated fields.

Some representative patents in the financial systems area include:

• Securities data compression: Patent CN109787638B uses a difference replacement algorithm to separate sign bits, combined with a dynamic three-tier storage allocation mechanism to achieve lossless compression of mixed trading data, enhancing the efficiency of financial data storage and real-time transmission.

DeepSeek's patent layout currently focuses on underlying technologies for cluster training, which seems to align with its ‘geek culture’ ethos.

Artificial intelligence models (5% of the patent portfolio)

In 2024, DeepSeek filed its first patent application related to the construction of AI model training sets, marking a step towards a comprehensive patent layout from underlying technologies.

• AI Model training set construction optimisation: Patent CN118246542A is based on a data sequence indexing mechanism, achieving efficient construction of large language model datasets through asynchronous IO batch reading and virtual hybrid ratio adjustment, reducing storage and computational resource consumption.

Clear patent technology layout route

The diagram below shows the evolution of DeepSeek's patent technology themes. Overall, DeepSeek's patent layout presents a clear evolution path from solidifying infrastructure to deepening technology and expanding scenarios. This process is closely related to model iteration, engineering optimisation, and industry applications.

Table 1: Detailed evolution of patented technologies

Source: Patent searches conducted in the database of “www.zhihuiya.com” as of 28 February 2025.

Phase one (2018 – 2023): Infrastructure construction

The period from 2018 to 2023 marked a critical phase for breakthroughs and large-scale applications of global large language model technology. This phase, based on the Transformer architecture, initiated the explosion of pre-trained models. Meanwhile, the open-source ecosystem grew rapidly, and industry applications penetrated fields such as finance and healthcare.

DeepSeek broke through computational bottlenecks through software layer optimisation, increasing training efficiency by 2.7 times (according to a report by China International Capital Corporation) under the same computational conditions, paving the way for a low-cost technical route.

DeepSeek focused on cluster training technology during this phase, implementing a very clear progressive patent layout strategy. It started with core technologies such as resource allocation and task scheduling, then expanded to optimisation technologies like asynchronous I/O and fault recovery and peripheral technologies like virtual development environments.

Representative patents at this stage include:

• Elastic resource allocation optimisation (core technology): DeepSeek's first patent CN109165093B is based on historical task clustering prediction and dynamic threshold warning mechanisms, achieving on-demand elastic scaling of cloud computing nodes, enhancing GPU resource utilisation and task response efficiency.
• Asynchronous/input-output (optimisation technology): Patent CN117707416A proposes a fixed block batch reading mechanism, integrating storage tags and adopting multi-thread asynchronous IO, achieving unified memory management and RDMA transmission, improving distributed storage reading efficiency and bandwidth utilisation.
• Virtual development environment expansion optimisation (peripheral technology): Patent CN115061725B proposes hierarchical recursive installation of extension packages, achieving intra-group shared development environments through dedicated channel data tables, and synchronising containers to training clusters, addressing resource waste as well as collaboration efficiency issues.

Phase Two (starting from 2024): Technology deepening and expansion

Here, DeepSeek's technology deepening and open-source strategy create a synergistic effect, proposing a series of patent applications related to GPU cluster network RDMA communication optimisation, and as mentioned above, the first patent application related to AI model training set construction.

Representative patents at this stage include:

• Multi-GPU communication optimisation: Patent CN117707416A proposes an in-transit flow control technology based on fragmentation. It dynamically adjusts packet fragment size, proportionally polls multiple GPUs for sending, and controls the number of in-transit fragments. This addresses network congestion and uneven load caused by multiple GPUs competing for RDMA network cards, improving bandwidth utilization.

Patent applications based in China: potential for globalisation

DeepSeek's current patent applications exhibit localisation characteristics in terms of application regions and drafting practices.

Application regions

DeepSeek's patents are only filed within China, without any applications for PCT international stage or overseas countries. While this strategy reduces initial legal costs, in the context of global competition and the rapid global promotion and application of AI technology, the patent layout may need to expand and adjust regionally as the company develops.

Early Publication

Most of DeepSeek's patent applications adopt an ‘early publication’ model, being published within six months after application, with some of the recent applications being published in as little as two months. This early publication strategy accelerates patent examination and quick authorisation but may also prematurely expose ongoing R&D projects or directions to competitors.

Drafting Practices

Claims are often drafted to better meet China's patent examination requirements. For example, in patent application CN117669701A, the software system architecture is defined through virtual modules. If expanding overseas, it is necessary to consider the patent examination rules of the target countries and compatible drafting practices to improve the chances of authorisation and obtain better patent protection.

Complex equity structure and joint patent ownership arrangement

We searched and reviewed DeepSeek's equity structure and patent ownership arrangements.

As shown in the equity penetration diagram below, DeepSeek's equity structure is characterised by multiple entities, multiple levels, and decentralisation.

This equity arrangement helps achieve concentrated control through penetration, maintaining dominance over the company. In the diagram, entities within DeepSeek that act as patent applicants are marked with a purple background.

Source: Searches in the licensed database of www.qcc.com as of 28 February 2025.

From the perspective of patent ownership arrangements, it has been observed that since mid-2022, DeepSeek's patent applications have shown a pattern of multiple legal entities being listed as joint applicants and joint patent owners. The diagram below lists the six entities within DeepSeek that appear as patent applicants in 19 published patents, along with the number of times each entity appears as a patent applicant.

Source: Patent searches conducted in the database of “www.zhihuiya.com” as of 28 February 2025.

It can be seen that the three entities under control in DeepSeek's equity control diagram, namely,

• Hangzhou DeepSeek Artificial Intelligence Basic Research Co., Ltd;
• Hangzhou DeepSeek Technology Co., Ltd; and
• Ningbo Jimi Information Technology Co., Ltd.

are the main patent applicants and owners.

From a patent management perspective, having multiple co-applicants may increase the legal costs associated with patent management, operation, and enforcement. In matters involving patent ownership transfer or other major issues, certain key procedures require consent from all the joint owners, which may create structural conflicts with the decision-making efficiency of a founder-controlled company.

Gaps between patented technology and open-source technology

DeepSeek adopts a dual-track strategy of ‘patent moat + open-source ecosystem’, building technical barriers based on core algorithm and architecture patents while open-sourcing core model architecture, training and inference toolchains, multimodal middleware, and developer kits to attract developers to co-build the ecosystem. This creates a synergistic effect between technology standardisation and commercial monetisation, accelerating the penetration of the AI industry.

We have also studied DeepSeek's published papers. In these papers, DeepSeek's core technology focuses on efficient architecture and training optimisation, adopting the finely grained expert division DeepSeek-MoE architecture, combined with multi-head implicit attention (MLA), and using FP8 mixed precision and DualPipe pipeline to achieve low-cost training.

These techniques support the model in achieving top-tier closed-source levels in inference tasks. However, DeepSeek does not seem to have corresponding patent applications for these key technologies disclosed in the papers, especially for improvements to the MoE architecture and other AI model-related technologies.

From a patent protection strategy perspective, technological improvements related to AI large language models are not entirely patent-prohibited. If the eligibility and inventiveness challenges can be addressed according to the patent examination rules of the target country, an opportunity still exists to obtain exclusive patent rights, enhancing technical competitiveness and market control.

As a reference, Google's patent application CN115699041A (see Figure 5 below) proposes a transfer learning framework based on a mixture of experts (MoE), solving traditional transfer computation redundancy issues through pre-training multidomain expert models, dynamic performance evaluation, and adapter fine-tuning, improving cross-task generalisation efficiency and meeting the needs of visual tasks in low-data scenarios.

Although Google does not have a C-end commercial layout for AI large language models in China, this patent is laid out in China and the US, emphasising the Chinese market and technological competition.

Source: Search in the licensed database of www.zhihuiya.com as of 28 February 2025.

The above analysis shows that DeepSeek has initially formed technical barriers in the AI field through architectural innovation and patent applications. In the future, through further in-depth layout, DeepSeek is expected to form a highly flexible patent asset pool, enhance cross-licensing negotiation capabilities in technical games, and build a strategic support system for participating in global competition.

As AI technology falls within the scope of national security reviews in various jurisdiction, patent layout has surpassed the scope of corporate business strategies and become a key element affecting technological sovereignty competition. DeepSeek's patent strategy upgrade concerns the company's development and has exemplary significance for China's participation in the construction of the global AI governance system.