Moore Threads: Breaking Ecosystem Competition Barriers with Unified Architecture

In the AI era, chip companies that lack a systematic approach to intellectual property layout will find it difficult to be recognized as truly mastering core technologies.

Text | Hu Jiaqi

ID | BMR2004

Recently, domestic GPU manufacturer Moore Thread (688795.SH) announced the launch of its AI Coding Plan intelligent programming service. This new product leverages the full-precision computing power of MTT S5000, achieving exponential increases in computing efficiency through a hardware-software co-design architecture. It transforms the traditionally low-level GPU capabilities into productivity tools for developers. This step marks the beginning of domestic GPU companies shifting from pure hardware suppliers to builders of complete computing platforms.

Moore Thread was founded in June 2020 and listed on the STAR Market at the end of 2025, setting a record for the fastest IPO approval in 88 days, attracting market attention. Its goal is not just to fill the computing power gap but to build a global infrastructure for accelerated computing and a one-stop solution, providing fundamental support for digital and intelligent transformation across industries.

If computing power is the entry ticket, then a systematic capability is the key to staying at the table long-term. Regarding this deeper competitive dimension, the Business School interviewed Moore Thread executives and industry experts to reveal a less visible part: how domestic GPU companies are building their own “moats” in patent systems and underlying architectures.

01

Product-oriented patent layout

In patent strategy, Moore Thread does not pursue quantity alone but emphasizes cultivating high-value patents.

According to a senior executive at Moore Thread, as of June 2025, the company has been granted 514 patents, including 468 invention patents, ranking among the top domestic GPU companies. Notably, these patents are not scattered but highly focused on the core chain of AI computing, covering processor architecture design, AI application acceleration and parallel computing optimization, drivers and underlying software systems, as well as GPU computing clusters and high-performance interconnects, gradually forming a systematic layout.

In patent strategy, Moore Thread emphasizes cultivating high-value patents, which focus on key areas such as processor architecture, parallel computing, memory management, high-speed interconnect protocols, compiler and driver optimization, energy efficiency control, and AI acceleration. These technologies directly impact performance, energy consumption, and compatibility, and are more likely to form technological barriers.

In 2024, Moore Thread achieved notable results in two national high-value patent competitions: the Quwa Smart Computing Cluster project won first prize at the China Haidian High-Value Patent Cultivation Competition, and the full-feature GPU project received the Gold Award at the China Xiong’an High-Value Patent Competition, demonstrating deep accumulation in core technology R&D and patent management.

Industry veteran Liang Zhenpeng told Business School that possessing high-value patents can enhance a company’s technological discourse power, support ecosystem collaborations such as hardware adaptation and open-source software, and defend against infringement risks globally while increasing licensing bargaining power, helping to build an autonomous and controllable industrial chain.

Liang believes that GPU companies should shift their patent focus toward supporting industry deployment and commercial applications—whether patents can underpin a complete technical system and industry solutions.

In evaluating patent portfolios and commercialization, companies need to consider technical coverage, legal stability, market relevance, and the difficulty for competitors to circumvent. They should analyze how patents support product functions, cost control, and ecological compatibility, using key metrics like citation rates, claim scope, geographic coverage, litigation history, and standardization contributions. Balancing quantity and quality, companies should deploy high-value patents in core areas and supplement with peripheral patents to form a protective network, avoiding blind pursuit of quantity.

Against this backdrop, Moore Thread has also systematized its intellectual property management. Through national-level IP standardization certification, the company has expanded its management focus from legal affairs to the entire R&D, management, and technology transfer processes. Patent layout now participates early in product planning, technology roadmaps, and market strategies, becoming a foundational infrastructure for long-term development rather than a remedial measure.

Behind these actions is a shift in cognition. In the AI era, chip companies lacking a systematic IP layout are unlikely to be seen as truly mastering core technologies. GPUs are no longer just hardware products but complex computing platforms composed of instruction sets, compilers, drivers, operator libraries, scheduling systems, and cluster architectures. Each link involves patentable technological innovations, and each technological node could become a key bargaining chip in future commercial cooperation and industry competition.

02

The space for hardware differentiation is shrinking

The cost of increasing transistor density grows exponentially, while performance gains are narrowing, and process dividends are diminishing.

For chips like GPUs that rely heavily on transistor density, energy efficiency, and frequency, process technology remains the key factor determining performance density and power consumption.

Moore Thread’s current products mainly use mature mid-to-high-end process nodes. For example, the 2022 MTTS50 for the Xinchuang market is based on a 12nm process, which is a reliable choice for graphics products, facilitating rapid mass production and cost control.

Industry-wide, mainstream high-performance GPUs are gradually moving toward more advanced process nodes. Typically, these advanced nodes offer higher transistor density and better energy efficiency. Industry insiders told Business School that most current high-performance GPUs are based on 7nm processes.

In comparison, leading international GPU manufacturers like NVIDIA mainly use TSMC’s 4nm process for top-tier products, delivering higher performance density and efficiency. Domestic foundries face constraints from global supply chains and equipment, and the maturity and yield of the most advanced nodes still have room for improvement, directly affecting the performance limits of domestically produced high-performance, low-power GPUs.

At this stage, Moore Thread mainly optimizes architecture and scheduling strategies at the design level, combined with mature process nodes, to strike a balance between energy consumption and performance, while controlling costs and supply stability. This strategy not only enables rapid deployment of current products but also leaves room for future iterations as domestic advanced process capabilities improve.

Professor Cao Xinming from the Intellectual Property Research Center at Zhongnan University of Economics and Law believes that this change is primarily driven by physical limits at the hardware level. As advanced nodes approach 3nm and 2nm, the cost of increasing transistor density grows exponentially, while performance gains diminish. Additionally, global capacity for advanced processes is highly concentrated among a few manufacturers like TSMC, Samsung, and Intel, further compressing hardware differentiation space.

03

Unified architecture and software ecosystem

Today’s GPU industry competition is undergoing a structural shift, with IP systems and developer ecosystems becoming key.

If physical hardware limits reduce the space for differentiation through process technology, then the maturity of software ecosystems significantly raises the industry’s entry barriers.

Cao Xinming pointed out that the focus of GPU industry competition is shifting structurally. While process technology (Moore’s Law) and raw computing power (FLOPS) remain basic thresholds, long-term profitability and industry standing increasingly depend on patent barriers within the IP system and developer ecosystems supported by software stacks.

According to Moore Thread executives, the company has built its core technology and ecosystem around its自主研发的MUSA（Meta-computing Unified System Architecture）.

MUSA is an independently developed, fully integrated GPU computing acceleration architecture that combines hardware and software. It covers the entire stack from chip architecture, instruction set, programming models, to software libraries and driver frameworks, aiming to provide high-performance computing for various parallel computing scenarios. Based on MUSA, the platform can efficiently support AI computing, graphics rendering, physics simulation, scientific computing, and ultra-high-definition video encoding/decoding.

After five years of intensive R&D and continuous iteration, the upgraded MUSA 5.0 marks a relatively mature stage, with breakthroughs in full-stack integration, computing efficiency, and ecosystem openness. In programming, MUSA natively supports MUSA C and is deeply compatible with modern parallel languages like TileLang and Triton, offering developers a flexible and efficient full-stack development experience, reducing migration and adaptation costs. In performance, the core library muDNN achieves near-theoretical efficiency in key operators like GEMM and FlashAttention, with significantly improved communication efficiency, optimized compiler performance, and integrated high-performance operator libraries, greatly accelerating training and inference workflows.

Meanwhile, MUSA’s ecosystem strategy extends to open system construction. Moore Thread plans to gradually open-source core components such as acceleration libraries, communication libraries, and system management frameworks, inviting more partners to co-build the ecosystem.

Cao Xinming also believes that behind the software ecosystem, patents are playing an increasingly fundamental role in supporting the system.

GPU software stacks are not just engineering assemblies but involve many underlying innovations, such as compiler optimization techniques, parallel computing scheduling strategies, driver-hardware coordination mechanisms, and deep integration with mainstream AI frameworks. These involve core technologies that can be patented. Without a robust IP system, it’s difficult to gain full trust in cross-company collaborations and industry division of labor, and to form stable ecosystem alliances.

04

The competition from latecomers

Future GPU patent competition will focus more on heterogeneous computing, AI integration, hardware-software co-optimization, and emerging application scenarios.

For latecomers, the most obvious change in the industry is the significantly higher entry barrier.

They often need to invest huge sums upfront in chip development and fabrication without seeing immediate market returns. But producing hardware is only the first step. Without mature compiler tools, driver support, and compatibility with mainstream AI frameworks, developers will find it hard to use the new chip efficiently, and users will be reluctant to migrate platforms. Without a user base, ecosystems cannot form; without ecosystems, products struggle to land. This cycle makes entering the GPU industry far more difficult than many other semiconductor segments.

Today’s developers are deeply tied to NVIDIA’s CUDA and its toolchain, with training workflows, operator optimization, and engineering experience built on this platform. Unless a new platform offers performance or efficiency advantages by an order of magnitude, convincing developers to rewrite code and rebuild workflows is challenging. That’s why many AI chip companies prioritize compatibility with CUDA or mainstream frameworks rather than building entirely new ecosystems—reducing entry barriers by leveraging existing systems.

Moore Thread’s strategy reflects this pragmatic approach. It promotes the development of its own programming models and low-level libraries to establish a controllable technical foundation, while also emphasizing compatibility with mainstream graphics interfaces and AI frameworks, balancing “independent systems” with “practical compatibility.”

Alongside ecosystem barriers, legal and patent risks are becoming increasingly unavoidable. Cao Xinming explained that key GPU technologies have been heavily patented over the years. New entrants must conduct complex and costly patent clearance before product launch to avoid infringement. Once involved in patent disputes with industry leaders, long litigation cycles and high costs can impose heavy financial burdens on less-funded companies.

He suggests that a more practical path for latecomers is to focus on specific vertical scenarios—such as autonomous driving inference, edge computing, or industrial vision—and develop targeted, optimized technology stacks within relatively closed environments. In these scenarios, power consumption, latency, and environmental adaptability are often more critical than general-purpose computing power. By “deeply cultivating” these niches, latecomers can create differentiated capabilities rather than trying to replicate the entire GPU ecosystem.

Long-term, the GPU industry landscape is not entirely fixed. The rise of open hardware and open software, such as RISC-V vector extensions and open-source frameworks like PyTorch and TensorFlow, provides new variables. These developments enable hardware vendors to optimize around general software ecosystems without relying solely on proprietary platforms. While these changes may not immediately overturn the existing order, they leave room for new entrants over the long term.

Liang Zhenpeng believes that future GPU patent competition will focus more on heterogeneous computing, AI integration, hardware-software co-optimization, and emerging applications like the metaverse and autonomous driving. The trend toward open-source software ecosystems will continue, with companies leveraging standard-essential patents and open licenses to expand their influence. For latecomers, strategies include focusing on niche technological breakthroughs to build patent advantages, actively participating in open-source communities to influence ecosystems, forming partnerships for rapid technology acquisition, and international expansion to mitigate risks—aiming to secure a foothold in the fiercely competitive global market.

Source | Business School, March & April Issue