In the AI era, a chip company lacking a systematic intellectual property layout can hardly be considered a true master of core technologies. Recently, domestic GPU manufacturer Moore Threads Technology Co.,Ltd. (688795.SH) announced the official launch of its AI Coding Plan intelligent programming service. This new product leverages the full-precision computing capabilities of the MTT S5000, achieving a multiplication of computing efficiency through a software-hardware collaborative architecture. It transforms the previously lower-level GPU capabilities directly into productivity tools for developers. This step signifies that domestic GPU companies are beginning to evolve from pure hardware suppliers into builders of complete computing platforms.
Founded in June 2020, Moore Threads went public on the STAR Market at the end of 2025, attracting market attention with a record-fast IPO approval process of just 88 days. Its ambition extends beyond merely filling computing power gaps; it aims to build a global acceleration computing infrastructure and one-stop solutions, providing foundational support for the digital and intelligent transformation of various industries. If computing power is the entry ticket, then systematic capability is the key to staying at the table long-term. Focusing on this deeper competitive dimension, insights from Moore Threads' relevant personnel and industry experts reveal a less visible aspect: how domestic GPU companies are building their own "moats" at the level of patent systems and underlying architectures.
From a product perspective for patent strategy, Moore Threads does not simply pursue quantity growth but emphasizes the cultivation of high-value patents. As of June 2025, Moore Threads has accumulated 514 authorized patents, including 468 invention patents, placing it among the top domestic GPU companies in terms of quantity. Notably, these patents are not scattered but are highly focused on the core chain of AI computing. They cover key areas such as processor architecture design, AI application acceleration and parallel computing optimization, driver and underlying software systems, and GPU computing cluster and high-performance interconnection, gradually forming a systematic layout. High-value patents refer to those concentrated in critical aspects of GPU R&D, including processor architecture design, parallel computing optimization, memory management mechanisms, high-speed interconnection protocols, compiler and driver optimization, energy efficiency control, and AI computing acceleration. These technologies directly impact performance, power efficiency, and compatibility, easily forming technical barriers. In 2024, Moore Threads achieved significant results in two national high-value patent competitions: its Kua'e Intelligent Computing Cluster project won the first prize at the China Haidian High-Value Patent Cultivation Competition, and its Full-Function GPU project received the gold award at the China Xiong'an High-Value Patent Competition, demonstrating its deep accumulation in core technology R&D and patent management.
A senior industrial economic observer pointed out that possessing high-value patents enhances a company's technological voice, supports ecosystem cooperation such as hardware adaptation and software open-source, defends against infringement risks in global competition, improves licensing bargaining power, and aids in building a self-controlled industrial chain. The observer believes that GPU companies should shift their focus on patents towards the overall support capability for industrial implementation and commercial application—that is, whether the patents can support a complete technological system and industrial solution. In terms of patent portfolio evaluation and commercial implementation, companies need to comprehensively consider technical coverage, legal stability, market relevance, and the difficulty of circumventing competitors, analyzing the practical support role patents play in product functionality, cost control, and ecosystem compatibility. Key metrics include patent citation rates, scope of claims, geographical coverage, litigation history, and contributions to standardization. Companies should balance quantity and quality, deploying high-value patents in core areas while supplementing them with peripheral patents to form a protective network, avoiding a blind pursuit of quantity.
In this context, Moore Threads' intellectual property management has also undergone a systematic transformation. Through national-level intellectual property standard certification, the company has expanded its management focus from single legal affairs to a full-process system encompassing R&D, management, and technology transfer. Patent layout now proactively participates in product planning, technology roadmap selection, and market strategy formulation, becoming an infrastructure for long-term corporate development rather than a post-hoc remedial measure. Behind these actions is a shift in mindset at Moore Threads. In the AI era, GPU is no longer just a single hardware product but a complex computing platform composed of instruction sets, compilers, drivers, operator libraries, scheduling systems, and cluster architectures. Each link corresponds to technological innovations that can be solidified into patents, and each technological node may become a crucial bargaining chip in future commercial cooperation and industrial competition.
The space for hardware differentiation is being compressed. The cost of increasing transistor density grows exponentially, while performance gains continue to narrow, indicating a clear decline in the benefits of process node advancement. For chips like GPUs, which are highly dependent on transistor density, power efficiency, and frequency, the manufacturing process remains a key factor determining performance density and power consumption. Currently, many of Moore Threads' products primarily utilize mature mid-to-high-end process nodes. For instance, the MTT S50, launched for the Xinchuang market in 2022, achieved mass production based on a 12nm process. This node represents a stable and reliable choice for graphics products, conducive to rapid mass production and cost control. From an industry-wide perspective, mainstream high-performance GPUs are gradually transitioning to more advanced process nodes, which typically offer higher transistor density and superior power efficiency. Industry sources indicate that the current common standard is a 7nm-based process. In contrast, leading international GPU manufacturers like NVIDIA often utilize TSMC's 4nm node for their top-tier products, yielding higher performance density and energy efficiency advantages. Domestic wafer foundries face constraints from the global supply chain and equipment limitations in advanced nodes, where the maturity and yield of the most advanced processes still have room for improvement. This directly impacts the ultimate performance and low-power capabilities of domestic GPUs. At this stage, Moore Threads primarily optimizes architecture and scheduling strategies at the design level while combining them with mature process nodes. This approach allows its products to achieve a market compromise between energy consumption and performance while controlling costs and ensuring supply stability. This strategy is not only suitable for the rapid deployment of current products but also reserves iteration space for the future as domestic advanced process capabilities gradually improve.
According to a professor from the Intellectual Property Research Center at Zhongnan University of Economics and Law, this change stems from the physical limits being approached at the hardware level. As advanced processes move towards 3nm and even 2nm nodes, the cost of increasing transistor density grows exponentially, while performance gains continue to narrow. Simultaneously, global advanced process capacity is highly concentrated in a few manufacturers. Companies like NVIDIA, AMD, and Qualcomm generally rely on foundry systems like TSMC, Samsung, and Intel for manufacturing, further compressing the space for hardware differentiation.
The focus of competition in the GPU industry is undergoing a structural shift, with IP systems and developer ecosystems becoming key. If hardware physical limits compress the space for differentiation through process nodes, then the maturity of the software ecosystem significantly raises the industry's competitive threshold. The professor pointed out that the competitive focus of the GPU industry is experiencing a structural migration. Process technology (Moore's Law) and raw computing power metrics (FLOPS) remain the basic entry barriers for the industry. However, factors that truly determine a company's long-term profitability and industrial position increasingly depend on the IP system constituted by patent barriers and the developer ecosystem supported by the software stack. According to information from Moore Threads' leadership, the company has built its core technological and ecological support around its self-developed MUSA (Meta-computing Unified System Architecture). MUSA is a unified system architecture for full-function GPU computing acceleration, integrating GPU hardware and software, developed independently by Moore Threads. This architecture encompasses a full-stack technology system ranging from chip architecture, instruction set, and programming model to software runtime libraries and driver frameworks. It aims to provide high-performance computing capabilities for various parallel computing scenarios. Based on the MUSA architecture, it can efficiently support diverse high-performance computing scenarios such as AI computing, graphics rendering, physical simulation, scientific computing, and ultra-high-definition video codec. After five years of deep R&D and continuous iteration, the newly upgraded MUSA 5.0 marks the entry of this unified architecture into a relatively mature new stage, achieving key breakthroughs in full-stack unity, computational efficiency, and ecosystem openness. In terms of programming ecosystem, MUSA not only natively supports MUSA C but also offers deep compatibility with modern parallel programming languages like TileLang and Triton, providing developers with a more flexible and efficient full-stack development experience, thereby reducing migration and adaptation costs. In computational performance, the core computing library muDNN achieves efficiency close to the theoretical limit on key operators like GEMM and FlashAttention. Communication efficiency has also been significantly improved, while compiler performance has been optimized manifold, integrating high-performance operator libraries to significantly accelerate the entire model training and inference pipeline. Meanwhile, MUSA's ecosystem strategy further extends to the construction of an open system. It is reported that Moore Threads plans to gradually open-source core components including computing acceleration libraries, communication libraries, and system management frameworks, opening its deeply optimized underlying capabilities to the developer community to attract more partners for co-building the ecosystem. Furthermore, the professor believes that behind the software ecosystem, patents are playing an increasingly fundamental institutional support role. The GPU software stack is not a simple engineering pile-up but contains a large amount of underlying innovation, such as compilation optimization techniques, parallel computing scheduling strategies, driver-hardware协同 mechanisms, and deep adaptation capabilities for mainstream AI frameworks. These all involve core technologies that can be patented. Software capabilities lacking the support of an intellectual property system find it difficult to gain full trust in cross-enterprise cooperation and industrial division of labor, and it is challenging to form stable ecological alliances.
For latecomers in the GPU industry, the most obvious change is the significantly raised entry barrier. Latecomers often must invest huge sums of money in chip R&D and tape-out without seeing market returns beforehand. However, producing the hardware is only the first step. If mature compilation tools, driver support, and adaptation capabilities for mainstream AI frameworks are lacking, developers will find it difficult to use the chip efficiently, and users will not easily migrate platforms. Without a user base, an ecosystem cannot form; without an ecosystem, products struggle to gain traction. This cycle makes starting in the GPU industry far more difficult than in most semiconductor sub-sectors. Today's developers are deeply tied to NVIDIA's CUDA and its toolchain. Training workflows, operator optimization, and engineering experience are all built on existing platforms. Unless a new platform can offer an order-of-magnitude advantage in performance or energy efficiency, it is difficult to persuade developers to rewrite code and rebuild processes. This is why many AI chip companies do not choose to build their own ecosystem entirely from scratch but instead prioritize compatibility with CUDA or mainstream frameworks, essentially leveraging existing systems to reduce the difficulty of entry. Moore Threads' strategy also reflects this pragmatic orientation. On one hand, it advances the construction of its own programming models and underlying libraries, attempting to form a controllable technological foundation. On the other hand, it emphasizes adaptation to mainstream graphics interfaces and AI frameworks, finding a balance between an "autonomous system" and "practical compatibility."
Parallel to ecosystem barriers are increasingly unavoidable patent and legal risks. The professor analyzed that after years of accumulation, core areas of GPU key technologies are already densely populated with patent layouts. Before a new entrant's product reaches the market, it often requires complex and expensive patent clearance searches to confirm whether its solutions pose infringement risks. Once卷入ed in patent disputes with leading companies, the lengthy litigation cycles and high costs can impose heavy pressure on companies with limited funding. The professor believes that in such an environment, a more realistic path for latecomers is to focus on specific vertical scenarios, establishing technically optimized stacks tailored for relatively closed application environments, such as autonomous driving inference, edge computing, or industrial vision. In these scenarios, factors like system power consumption, latency, or environmental adaptability are often more critical than general-purpose computing power. Latecomers have the opportunity to form differentiated capabilities through "scenario deep cultivation" rather than attempting to fully replicate the general-purpose GPU ecosystem. Looking further ahead, the GPU industry landscape is not entirely solidified. The rise of open-source hardware and software is introducing new variables to the industry. Routes like RISC-V vector extensions provide a technical foundation for building high-performance computing platforms different from traditional systems. The open-sourcing of frameworks like PyTorch and TensorFlow also gives hardware vendors the opportunity to optimize around a common software ecosystem without relying entirely on a single vendor's proprietary platform. While these changes are unlikely to shake the existing格局 in the short term, they preserve some entry space for latecomers in the long run. The senior observer believes that future GPU patent competition will focus more on heterogeneous computing, AI integration, software-hardware协同 optimization, and emerging application scenarios like the metaverse and autonomous driving. Simultaneously, the trend in software ecosystem construction will develop towards openness and open-source. Companies can combine standard-essential patents with open-source licenses to expand their technological and market influence. For latecomers, viable strategies include focusing on breakthroughs in细分 technologies to form patent advantages, actively participating in open-source communities to accumulate ecological influence, rapidly acquiring key technologies through cooperation and licensing, and emphasizing international layout and risk预警 to secure a foothold in the intense global competition.
Comments