Analysis of Commercialization Progress and Domestic Substitution Paths for China's Top Four GPU Players

Based on public information in December 2025, China’s computing power track is moving from “usable” to “good-to-use”. The four leading companies (Moore Threads, Biren Technology, Enflame, Muxi) have formed differentiated positioning and are accelerating capital and commercialization through IPOs or planned IPOs. Achieving “full substitution” of NVIDIA still takes time, but under the combined effect of structural opportunities in “training-inference”, cost and policy guidance, and local ecosystem building, domestic GPUs are expected to form a “multi-route coexistence” pattern in the domestic market, gradually narrowing the gap with international giants in training scenarios and occupying a larger share in inference and specific industry scenarios.

Key Judgments (All Based on Public Search Results):

• Ecosystem Aspect: NVIDIA’s CUDA ecosystem remains mainstream, but domestic manufacturers reduce barriers through compatibility and migration tools, and the penetration rate of domestic substitution in inference scenarios increases [15][19][18].
• Commercial Implementation: The four leading players have large-scale orders or deployments in scenarios such as operators, intelligent computing centers, and Internet/industry AI. Revenue and hand orders confirm the breakthrough from “technology to business” (but enterprises are generally in loss and rely on high-intensity R&D investment) [1][3][6][7][8][9][12].
• Technical Route: General-purpose GPU (GPGPU) and dedicated architecture (DSA/Yunsui-style customization) coexist, and low precision, interconnection, packaging and other engineering capabilities become key grasp for catching up [5][6][8][12][19].

1. Moore Threads: Targeting “Full-Function GPU”, MUSA Ecosystem Compensates for CUDA

• Technical Route: Following NVIDIA’s “AI + Graphics + Simulation” full-function route; emphasizing unified architecture and multi-scenario coverage; Kua E 10,000-card cluster evolving from 1,000-card to 10,000-card and then to 100,000-card [1][2][18][19].
• Architecture and Products: New generation “Hua Gang” architecture; AI training and inference integrated “Hua Shan” and graphics rendering “Lu Shan” chips; Kua E 10,000-card cluster with MFU (Model Computing Power Utilization) of over 60%; collaborating with Silicon Flow to achieve single-card inference throughput improvement on DeepSeek-R1 [1][2][19].
• Ecosystem and Compatibility: MUSA 5.0 upgrade, emphasizing compatibility and migration; MUSA Code and other toolchains reduce migration barriers, and compatibility and compiler technology are emphasized multiple times [18][19].
• Business and Capital: Sci-Tech Innovation Board “Domestic GPU First Stock”; cumulative revenue and loss coexist from 2022 to 2024 (2024 revenue 438 million yuan), relying on high-intensity R&D investment (cumulative over 4.3 billion yuan in three years) [1][3][8].
• Advantages: Covering training and inference, cloud and end; ecosystem compatibility advancing rapidly; highest capital and market attention [1][2][19].

2. Biren Technology: Focusing on GPGPU, Cloud Computing Power and Packaging First

• Technical Route: Focusing on GPGPU and intelligent computing solutions; adopting cutting-edge integration schemes such as Chiplet and 2.5D CoWoS packaging, optical interconnection [6][7].
• Products and Performance: BR100 series 7nm+CoWoS, some performance indicators claimed to be more than 3 times higher than mainstream products; advancing the development of the second-generation architecture [6][7].
• Commercialization and Orders: Revenue jumped from 499,000 yuan to 337 million yuan from 2022 to 2024, with significant compound annual growth rate; outstanding orders about 822 million yuan, and framework agreements/contracts about 1.241 billion yuan; cooperative customers include China Mobile, China Telecom, SenseTime, etc. [6][7].
• Capital Process: Sprinting for Hong Kong Stock Exchange “GPU First Stock”, planning to raise funds to focus on next-generation GPU R&D, Chiplet and global ecosystem; cumulative loss but still recognized by the capital market [6][7].

3. Enflame: Deeply Cultivating Cloud AI Training and Inference, “Yunsui” Dual Line and Ecosystem Engineering

• Technical Route: Adopting DSA dedicated architecture (GCU-CARA), focusing on cloud training and inference; dual-line products (Yunsui T series training/i series inference) [9][12].
• Products and Performance: Fourth-generation L600 (7nm), native FP8, 144GB storage, 3.6TB/s bandwidth and 800GB/s interconnection; 1,000-card level expansion; Yunsui i20 liquid cooling inference card achieves high computing power utilization in 10,000-card inference cluster (reported to be about 92%) [9][12].
• Commercial Implementation: Tencent is the largest shareholder and core cloud customer, with cumulative delivery of tens of thousands of cards; supporting mainstream cloud vendors such as Tencent Cloud and inference scenarios in finance, energy and other industries; landing in projects such as Gansu Qingyang 10,000-card inference cluster (50,000P computing power) [9][12].
• Capital Process: Financing nearly 7 billion yuan, valuation about 20.5 billion yuan; S60 landing 70,000 cards, total orders over 100,000 pieces; restarting Sci-Tech Innovation Board listing guidance (broker changed to CITIC Securities) [9][12].

4. Muxi: “China AMD” Narrative, Focusing on AI Training and Inference and Domestic Manufacturing

• Technical Route: Focusing on AI computing (integrated training and inference), emphasizing low precision and inference efficiency optimization; Xiyun C500 series is the main revenue source [8][10].
• Products and Manufacturing: Xiyun C500 (12nm domestic foundry + domestic packaging and testing) mass-produced in 2024 and became the revenue pillar; next-generation Xiyun C700 positioned to target H100 (under R&D) [8][10].
• Business and Orders: Three-year revenue grew rapidly from 2022 to 2024, with 2024 revenue 722 million yuan (accounting for 97.28% of main business revenue); cumulative sales over 25,000 units, outstanding orders about 1.43 billion yuan; entering key customer lists such as China Telecom’s centralized procurement [8][10].
• Capital and Loss: Sci-Tech Innovation Board “Domestic GPU Second Stock”, significant increase on the first day of listing; although cumulative loss exceeds 3.2 billion yuan, revenue growth is strong, and the market expects to reach break-even earliest in 2026 [8][10].

1. General-Purpose GPU (GPGPU) vs Dedicated Architecture (DSA/ASIC/NPU/TPU)

• GPGPU (e.g., Moore Threads, Biren, some Muxi products): Strong versatility, smoother migration and compatibility paths; but under process constraints, more dependent on engineering breakthroughs in architecture, interconnection and packaging [1][6][8][19].
• DSA/Yunsui-style customization (e.g., Enflame): Focus on energy efficiency and cost in specific scenarios, suitable for cloud vendors and batch inference; faster engineering and ecosystem adaptation speed [9][12][19].

2. Interconnection, Packaging and System-Level Engineering

• High-speed interconnection and super nodes: Moore Threads’ MTLink 4.0 and super nodes (MTTC256/1024-card vertical expansion) lay the foundation for large-scale clusters [19]; Enflame’s 800GB/s interconnection and high-bandwidth design also complement at the system level [12].
• Packaging yield progress: Domestic 2.5D packaging yield increased from early low level to 40%-60% (Q3 2025), supporting commercial mass production (still a gap with TSMC’s 90%+, but crossing the break-even point) [19].

3. “Cost-Effective Opportunity” in Low Precision and Inference Scenarios

• Domestic manufacturers continue to optimize FP8/FP4 and other low precision and inference throughput, and carry out adaptation with local large models (e.g., DeepSeek), forming an entry point of “usable inference, lower cost and controllable” [1][8][10][19].
• Reports show that some domestic cards have multiple advantages over products like H20 in specific inference throughput indicators (subject to specific scenarios and actual tests) [1][19].

• NVIDIA CUDA remains the industry de facto standard; but the rise of “de-bottoming” and compilers, automatic migration tools significantly reduces the migration threshold of domestic chips, and heterogeneous mixed training becomes a realistic strategy [15][19].
• Domestic substitution progresses faster in inference, industry intelligent computing and operator centralized procurement; training and large-scale pre-training scenarios still focus on high-end computing power such as H series [1][3][12][15][19].
• Driven by policy and supply chain security, local “sovereign AI” and computing power base construction accelerate, and leading cloud vendors and operators are more inclined to domestic verification and adaptation in procurement [8][10][12][18][19].

• Export and Supply: Potential supply changes of products like H200 ease the short-term computing power gap, but security and cost constraints remain [8][12][15][19].
• Track Differentiation: Some enterprises are on the fast track of capitalization through technology and commercial implementation, while some early projects face challenges in financing/mass production (Xiangdixian, Lisuan, etc. are mentioned in the search, and their current situation should be based on authoritative sources) [17].
• Capital and Valuation: The market pays for the dual narrative of “domestic substitution + AI computing power explosion” with the rise of market value of leading enterprises accompanied by high PS valuation, which needs to be兑现 through fundamentals to cross the cycle [8][12][17].

• Short-Term (1-2 Years): Inference, industry intelligent computing and cloud vendor mixed training and inference loads will be the fastest penetration scenarios for domestic GPUs; 10,000-card cluster scale deployment will verify interconnection, stability and software stack maturity [1][9][12][19].
• Medium-Term (3-5 Years): Advanced process and packaging yield, ecosystem compatibility and toolchain, large model adaptation and industry standard game will determine whether each company can move from “usable” to “good-to-use” and form a stable share in the domestic market [6][8][10][19].
• Risk Points: High-end process constraints, supply chain fluctuations; core procurement of Internet giants is still highly prudent; profit schedule depends on R&D investment and commercialization rhythm; industry integration may accelerate [8][10][12][17][19].

[0] Jinling API Data (Note: This item is only a preset placeholder; this answer does not actually call this tool, and the above discussion is all from the following public web search sources)
[1] Securities Times - Moore Threads Announces Full-Function GPU Technical Roadmap, Kua E 10,000-Card Cluster and MTTC256 Super Node, etc. (2025-12-22/21)
[2] Geek Park - Targeting NVIDIA, Moore Threads Shows Its Family Background for the First Time After Listing (2025-12-25)
[3] Sina Finance - Moore Threads Listed, “Domestic GPU First Stock” and R&D Investment (2025-12-05)
[8] Guest Media/Sina Finance - Muxi Listed and Domestic GPU “China AMD” Narrative (2025-12-23)
[6] iyiou.com/International Electronic Business News - Biren Technology Hong Kong IPO and GPGPU/Chiplet Layout (2025-12-23)
[7] Securities Times - Biren Technology Starts Prospectus and Technical Iteration Roadmap (2025-12-23)
[9] Guest Media/Zhihu Column - Enflame Cloud AI Training and Inference and 10,000-Card Inference Cluster Landing (2025-12)
[12] Securities Times - Year-End Inventory: Computing Power Localization and GPU New Stocks (Including Moore Threads Products and Interconnection Technology Progress) (2025-12)
[19] 36Kr - Domestic GPU “Siege” NVIDIA: Moore Threads, MUSA Ecosystem and Compatibility Migration (2025-12)
[15] Wenxuecheng/Communication Industry Network - Heterogeneous Mixed Training and Ecosystem Breakthrough Strategy (2025-12)
[17] Observer Network - The Embarrassment That Moore Threads Cannot Avoid; Mentioning Xiangdixian, Lisuan and Other Track Participants (2025-12-18)
[18] Phoenix Technology - Moore Threads MUSA Ecosystem Targets CUDA (2025-12-20)
[10] Oriental Fortune/Xinhua Finance - Domestic AI Chips Accelerate Catching Up and Sugon scaleX 10,000-Card Cluster (2025-12-26)
[5] Sina Finance/36Kr - Cambricon vs Domestic GPU Comparison, AI Chip Wealth Creation and Valuation (2025-12-18)
[8-10] Same as [8][10] reports on Muxi/domestic GPU “four leading players” listing and hand orders.

Note: The above analysis is entirely based on public reports and company announcements retrieved this time, and does not include content not appearing in the search results or unconfirmed inferences in the user’s “Context”. Time, performance and product progress are subject to disclosure by various sources.