🕷️
SetupAI
💰Stalecollected in 46m

Moore Threads' Strategy for Domestic AI Chip Dominance

Moore Threads' Strategy for Domestic AI Chip Dominance
PostLinkedIn
💰Read original on 钛媒体
#gpu#domestic-chipsmoore-threads-gpumoore threads

💡Understand how domestic GPU players are building architectural moats to compete in the trillion-token AI era.

⚡ 30-Second TL;DR

What Changed

Focus on unified architecture as the primary technical moat

Why It Matters

This strategy highlights the shift in domestic chip competition from raw performance to ecosystem and architectural integration. It suggests that long-term viability depends on software-hardware synergy.

What To Do Next

Evaluate Moore Threads' SDK compatibility with your current PyTorch/TensorFlow workflows to assess integration feasibility.

Who should care:Developers & AI Engineers

Key Points

  • •Focus on unified architecture as the primary technical moat
  • •Positioning full-function GPUs as critical physical AI infrastructure
  • •Scaling beyond individual chip performance to cluster-level capabilities

🧠 Deep Insight

Web-grounded analysis with 29 cited sources.

🔑 Enhanced Key Takeaways

  • •Moore Threads' proprietary MUSA (Moore Threads Unified System Architecture) is a comprehensive full-stack solution, encompassing a unified programming model, software runtime libraries, driver framework, instruction set architecture, and chip architecture, designed to serve as a domestic alternative to NVIDIA's CUDA ecosystem.
  • •The company has significantly reoriented its business focus, with approximately 97% of its revenue now derived from high-end data center clusters for AI factories, indicating a strategic pivot away from the consumer GPU market.
  • •Moore Threads is actively developing a complete "cloud-edge-end" ecosystem for embodied AI, which includes the MT Lambda simulation platform and strategic partnerships aimed at generating synthetic data and facilitating strategy training for robotics and autonomous driving.
  • •Their newly unveiled Huagang GPU architecture, announced in December 2025, supports full-precision computing from FP4 to FP64, and is claimed to increase compute density by 50% while improving energy efficiency tenfold.
  • •In March 2026, Moore Threads secured a substantial sales contract valued at CNY 660 million (approximately $95.5 million) for its KUAE AI Computing Cluster, signaling a critical transition towards large-scale cluster deployment and commercialization.
📊 Competitor Analysis▸ Show
Feature/CategoryMoore Threads (MUSA GPUs)NVIDIA (e.g., A100, H100, Blackwell)Huawei (Ascend 910)Alibaba (Hanguang 800)
ArchitectureMUSA (Unified System Architecture), Huagang (next-gen)CUDA, Hopper, BlackwellDa VinciHanguang
Process Node12 nm (MTT S80, S3000, S4000)Advanced (e.g., 5nm for H100)7 nm12 nm
Key AI ChipsMTT S4000, MTT S5000, Huashan (next-gen)A100, H100, H200, B200Ascend 910CHanguang 800
FP32 Performance25 TFLOPS (MTT S4000)19.5 TFLOPS (A100), 67 TFLOPS (H100)~250 TFLOPS (Ascend 910, FP16)N/A (focus on inference)
FP16/BF16 Performance100 TFLOPS (MTT S4000)312 TFLOPS (A100), 1979 TFLOPS (H100)N/AN/A
INT8 Performance200 TOPS (MTT S4000)624 TOPS (A100), 3958 TOPS (H100)N/AN/A
Memory Capacity48GB GDDR6 (MTT S4000), up to 64GB (Lushan), 8 HBM modules (Huashan)80GB HBM2e (A100), 80GB HBM3 (H100)N/AN/A
Memory Bandwidth768 GB/s (MTT S4000), 1314 GB/s (MTLink 4.0 for Huagang)1.5 TB/s (A100), 3.35 TB/s (H100)N/AN/A
InterconnectMTLink (up to 1314 GB/s for Huagang)NVLink (900 GB/s for H100)N/AN/A
Cluster ScaleKUAE cluster (thousands to 100,000+ GPUs)DGX systems (large scale)N/AN/A
Software EcosystemMUSA SDK (CUDA 12.8 alignment, PyTorch support, vLLM-MUSA, MUSIFY for CUDA migration)CUDA (dominant, extensive libraries)Ascend AI software stackAlibaba Cloud AI platform
Performance Claims (vs. NVIDIA)MTT S4000 competitive with A100-era, Huashan comparable to Hopper/Blackwell, exceeding B200 memory bandwidthIndustry benchmark leaderAscend 910C: 60-80% of H100 performanceHanguang 800: positioned against NVIDIA P4
US Sanctions ImpactAdded to Entity List (Oct 2023), driving domestic tech stackExport controls restrict high-end chips to ChinaN/AN/A
Market FocusData center AI, embodied AI, cloud-edge-end solutionsCloud AI, HPC, data centers, gamingCloud AI, HPCCloud AI, data centers
Pricing (Consumer)MTT S80 ~$200, S70 ~$130 (Feb 2024)PremiumN/AN/A

🛠️ Technical Deep Dive

  • MUSA Architecture: Moore Threads Unified System Architecture (MUSA) is a proprietary full-stack solution encompassing a unified programming model, software runtime libraries, driver framework, instruction set architecture, and chip architecture. It supports DirectX, Vulkan, OpenGL, AI model training and inference, simulation, rendering, and multi-platform compatibility (x86, ARM, LoongArch).
  • MTT S80 GPU: Built on the Chunxiao graphics processor using a 12 nm TSMC process, it features 4096 shading units, 16 GB GDDR6 memory with a 256-bit interface (448 GB/s bandwidth), a GPU clock of 1800 MHz, and a 255 W TDP. It supports DirectX 12 Ultimate, including hardware ray tracing, and connects via PCIe 5.0 x16.
  • MTT S3000 GPU: A server-class GPU based on the Chunxiao chip and MUSA architecture, offering 4096 MUSA Cores, a 1.9GHz GPU clock, 15.2 TFLOPS (FP32), 32GB GDDR6 (256-bit, 448GB/s), PCIe Gen5 x16, and a 250W TDP. It includes MUSA Security Engine 1.0 and supports GPU Elastic Virtualization and SR-IOV.
  • MTT S4000 AI GPU: Features the third-generation MUSA architecture with 8192 MUSA Cores and 128 Tensor Cores. It has 48GB GDDR6 memory providing 768 GB/s bandwidth, 25 TFLOPS (FP32), 50 TFLOPS (TF32), 100 TFLOPS (FP16/BF16), and 200 TOPS (INT8). Connectivity includes PCIe 5.0 x16 and MTLink interconnect technology, offering 240 GB/s inter-card bandwidth for multi-GPU clusters. It also supports MUSA Security Engine 2.0 and hardware virtualization.
  • Huagang Architecture (Next-Gen): Unveiled in December 2025, this architecture supports full-precision computing from FP4 to FP64, boasts a 50% increase in compute density, and a tenfold improvement in energy efficiency. It introduces a new instruction set, asynchronous programming capabilities, and enhanced thread scheduling efficiency.
  • Huashan AI GPU: Based on the Huagang architecture, this next-generation AI chip features a dual-chiplet design and 8 HBM modules (some sources state 9 HBM modules). It is claimed to offer performance comparable to NVIDIA's Hopper and Blackwell GPUs, with memory bandwidth potentially exceeding the B200, and is scalable to clusters of over 100,000 GPUs using MTLink 4.0 (1314 GB/s).
  • Lushan Gaming GPU: Also built on the Huagang architecture, this GPU is positioned as the successor to the MTT S80 and S90. It promises significant performance gains, including up to 15x better performance in AAA games, a 50x boost in ray tracing, and a 64x increase in AI computing performance, along with a fourfold increase in memory capacity (up to 64GB).
  • MTLink Interconnect: Moore Threads' proprietary high-speed interconnect technology for multi-GPU clusters. MTLink 1.0 provides 240 GB/s inter-card bandwidth, while the next-generation MTLink 4.0 for Huagang architecture is designed for 1314 GB/s.
  • KUAE Computing Cluster: A large-scale intelligent computing cluster solution capable of supporting deployments ranging from thousands to tens of thousands of accelerator cards. It has demonstrated a 91% linear speedup ratio in a thousand-card cluster for LLM training.
  • Software Ecosystem: The MUSA software stack includes MUSA SDK 5.1.0, which aligns with CUDA 12.8 and fully supports 3,194 PyTorch operators. Moore Threads has also open-sourced vLLM-MUSA and developed a MUSIFY translation tool for seamless migration of CUDA code to the MUSA platform.

🔮 Future ImplicationsAI analysis grounded in cited sources

Moore Threads will significantly reduce China's reliance on foreign AI computing infrastructure.
The company's strategic focus on a full-stack domestic solution, encompassing proprietary hardware (MUSA GPUs, KUAE clusters) and a comprehensive software ecosystem (MUSA, AI coding tools), directly addresses the national 'sovereign AI' mandate and mitigates the impact of US sanctions.
Moore Threads' emphasis on full-function GPUs and embodied AI simulation will accelerate the development and deployment of robotics and autonomous driving in China.
By integrating AI computing, graphics rendering, and physics simulation on a single chip and developing platforms like MT Lambda for synthetic data generation, Moore Threads aims to overcome data bottlenecks and reduce real-world trial-and-error costs for embodied AI applications.
Moore Threads' next-generation Huashan AI GPU will offer performance competitive with NVIDIA's high-end Hopper and Blackwell architectures, particularly in memory bandwidth.
Company claims for the Huashan GPU indicate performance comparable to NVIDIA's Hopper and Blackwell, with memory bandwidth potentially exceeding the B200, suggesting a significant leap in its AI compute capabilities.

⏳ Timeline

2020-10
Moore Threads founded by Zhang Jianzhong.
2022-03
Announced first products, MTT S60 (desktop) and MTT S2000 (server), based on first-generation MUSA architecture.
2023-10
Added to the U.S. Department of Commerce's Entity List.
2025-12
Held initial public offering (IPO) on the Shanghai Stock Exchange, raising over US$1 billion.
2025-12
Unveiled next-generation Huagang GPU architecture, with Huashan (AI) and Lushan (gaming) GPUs, at MUSA Developer Conference.
2026-03
Secured a sales contract worth CNY 660 million for its KUAE AI Computing Cluster.
💰Read original article on 钛媒体
📰

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Related Updates

Same topic

Explore #gpu

Same product

More on moore-threads-gpu

Same source

Latest from 钛媒体

AI Video Industry Shifts from Hype to Maturity

AI Video Industry Shifts from Hype to Maturity

钛媒体•Jul 29

AI-curated news aggregator. All content rights belong to original publishers.
Original source: 钛媒体 ↗