Company Overview
Moore Threads was founded in October 2020 and is headquartered in Beijing. Its core founding team comes from international GPU giants such as NVIDIA and AMD, with deep expertise in GPU architecture and graphics. The company positions itself as a "full-function GPU" provider, with products covering graphics rendering, AI training and inference, scientific computing, and other use cases, committed to building an independently controllable computing power foundation amid the wave of domestic substitution. As of 2025, the company has completed multiple financing rounds, with cumulative funding exceeding RMB 5 billion and valuation surpassing RMB 20 billion, making it one of the most promising unicorns in China's GPU sector.
Key development milestones:
- 2020: Founded; launched MUSA architecture R&D.
- 2022: Released the first full-function GPU chip "Sudi" and the MTT S series graphics cards.
- 2023: Launched the second-generation architecture "Chunxiao"; the MTT S3000/S4000 went into mass production and entered the Xinchuang market.
- 2024: Fully adapted to domestic large models such as DeepSeek; signed multiple intelligent computing center projects.
- 2025: Taped out the third-generation architecture "Quping", with performance benchmarked against NVIDIA's mid-to-high-end products.
It covers two major product categories: graphics cards (for XinChuang PCs and servers) and AI acceleration cards (for data centers). Representative products:
- MTT S4000: Based on the "Chunxiao" architecture, delivering single-card FP16 computing power of 130 TFLOPS, with 48GB HBM2e memory, supporting PCIe 5.0, meeting the demands of large-scale AI model training and inference.
- MTT S3000: Designed for inference scenarios in data centers, with FP16 computing power of 80 TFLOPS and power consumption of 250W. It has passed the domestic server adaptation certification of telecom operators. The chips are manufactured using TSMC's 7nm/5nm processes, with all core IP self-developed, and a main frequency exceeding 1.8GHz. Downstream customers mainly include intelligent computing center integrators, operator cloud departments, and local government big data bureaus.
The MUSA software stack is Moore Threads' core differentiating capability, providing:
- MUSA SDK: Compatible with NVIDIA CUDA APIs, allowing users to migrate existing AI frameworks (PyTorch, TensorFlow, etc.) at zero cost.
- MUSA X: A distributed training framework for large models, supporting one-click adaptation of domestic models such as DeepSeek, Baichuan, and Zhipu.
- Drivers and graphics libraries: Full support for Vulkan/OpenGL/DirectX, meeting the needs of XinChuang office and industrial design. Revenue mainly comes from technical licensing fees, customized development services, and subsequent maintenance fees.
For large customers, it provides full-service solutions from chip selection to cluster deployment. For example, a provincial intelligent computing center adopted Moore Threads GPUs to build a thousand-card cluster for government large model training. This business has a relatively high gross margin (about 60%), but the revenue scale is limited.
| Product Line | Revenue Share | Core Customers | Gross Margin |
|---|---|---|---|
| GPU chips and boards | ~85% | Intelligent computing centers, XinChuang PC manufacturers | ~40% |
| Software stack and ecosystem | ~10% | AI enterprises, cloud vendors | ~70% |
| Customized solutions | ~5% | Government, large state-owned enterprises | ~60% |
Technical Moat
Moat 1: Self-developed MUSA Architecture
MUSA (Moore Threads Unified System Architecture) is a full-featured GPU instruction set architecture built from scratch by the company, now iterated to its third generation, codenamed "Curved Screen." Key features:
- Unified rendering architecture: simultaneously supports graphics rendering, AI computing, and scientific computing without the need for separated cores.
- Software-defined Tensor Core: flexibly configurable matrix computation units that adapt to different precisions (FP32/FP16/INT8).
- Memory consistency: reduces data movement overhead through unified virtual memory, enhancing large-model inference efficiency. This architecture has been filed for over 200 invention patents, forming a robust intellectual property barrier.
Moat 2: CUDA Ecosystem Drop-in Replacement Capability
Moore Threads is the only domestic GPU manufacturer to achieve application-level CUDA compatibility. Through the MUSA SDK, over 90% of CUDA programs can run directly without source code modification, significantly reducing user migration costs. In 2024, the company was the first to complete adaptation for the full DeepSeek-V2 model series, with inference performance reaching 70% of an NVIDIA A100 at a comparable level, along with deep optimization of Transformer operators.
Moat 3: Deep Integration with the Domestic AI Ecosystem
Moore Threads has completed adaptation with mainstream deep learning frameworks including Huawei MindSpore, Baidu PaddlePaddle, and Alibaba PAI, and has established strategic partnerships with LLM vendors such as DeepSeek and Zhipu AI. As domestic AI large models iterate rapidly, Moore Threads' closed-loop ecosystem of "chip + software + model" is accelerating its strategic positioning advantage, making it difficult for new entrants to replicate in the short term.
| Dimension | Data |
|---|---|
| Global Market Share | <1% (Top 3 in Domestic GPU Sector) |
| Industry Ranking (Domestic GPU) | Third (after Cambricon and Hygon Information) |
| Main Competitors | Cambricon (Siyuan Series), Hygon Information (DCU), Biren Technology (BR100), Jingjia Micro (JM9 Series) |
| Downstream Customers | China Mobile, China Telecom, State Grid, a provincial-level big data bureau, leading AI laboratories |
| Competitive Advantages | Full-function GPU vs. Cambricon/Hygon's AI-focused approach; dual-track in graphics + AI; most well-established software ecosystem |
The domestic GPU market is currently driven by government-led centralized procurement for intelligent computing centers and the Xinchuang market. Moore Threads captured approximately 15% of the market share in 2024 Xinchuang GPU tenders, ranking third. As U.S. sanctions on AI chips exports to China intensify and domestic substitution policies continue to gain momentum, the company is expected to secure a larger market share from 2025 to 2027.
Financials and Growth
Note: Moore Threads has not yet gone public; the following data is based on the prospectus draft (filed in 2024) and industry estimates.
| Metric | Data |
|---|---|
| Revenue (2024) | Approximately RMB 1.2 billion (YoY growth of 300%) |
| Gross Margin | Approximately 45% |
| Net Margin | Loss (heavy R&D investment, net loss of approximately RMB 800 million) |
| Core Growth Drivers | 1) Xinchuang CPU replacement drives GPU demand; 2) Intelligent computing center construction drives shipments of AI accelerator cards; 3) Large model ecosystem adaptation brings orders; 4) Expansion in military and industrial simulation markets |
Revenue Structure Evolution: Revenue was less than RMB 100 million in 2022, approximately RMB 300 million in 2023, and exceeded RMB 1 billion in 2024. Revenue is expected to reach RMB 3 billion in 2025, with breakeven possible by 2027. R&D expenses as a percentage of revenue have consistently exceeded 60%, with talent costs accounting for the majority (average annual compensation per engineer exceeds RMB 800,000).
Financing and Valuation: The company has completed 6 rounds of financing, with a post-money valuation of approximately RMB 20 billion in 2024. Institutional shareholders include Sequoia Capital China, 5Y Capital, ByteDance, Shenzhen Capital Group, and others.
Key Risks:
- Technology iteration risk: NVIDIA's next-generation architecture, Blackwell, delivers significantly higher performance. Moore Threads must sustain high-intensity R&D, but has limited financial resources.
- International supply chain risk: Core EDA tools and advanced process nodes depend on TSMC; escalating geopolitical conflicts could disrupt tape-outs.
- Ecosystem dependency risk: Despite achieving CUDA compatibility, NVIDIA may impose restrictions on CUDA licensing terms or push for proprietary ecosystem lock-in.
- Commercial deployment risk: Intelligent computing center projects have long payment collection cycles and a high proportion of accounts receivable; some customers did not make bulk purchases after trials due to insufficient performance.
- Talent attrition risk: GPU design talent is scarce, and competitors poach with high salaries.
Core Investment Thesis / Industry Value Summary:
Moore Threads is the only hard-tech enterprise in the domestic GPU field to achieve full-function coverage of "graphics + AI + scientific computing." Leveraging its proprietary MUSA architecture and a software ecosystem deeply adapted to domestic large models, it holds a critical position in the Xinchuang (domestic IT innovation) and intelligent computing center markets. Although still in the investment phase, with accelerating domestic substitution and surging demand for large-model inference, the company is expected to reach a profit inflection point in 2026, becoming a core target for autonomous and controllable AI computing power.