Author | ZeR0 Editor | Moying
A domestic GPU cluster with zero NVIDIA content has just delivered a "major result."
According to ChipInsider's report on November 2, on Saturday, Moore Threads, a leading domestic GPU company, announced a strategic partnership with the Wuxi Huishan Investment Promotion Group and Neolix, a top-tier L4 autonomous driving unmanned vehicle enterprise. Neolix plans to conduct larger-scale autonomous driving model training at the Wuxi (Huishan) Domestic Intelligent Computing Center, requiring nearly 10,000-card level compute power. The Wuxi (Huishan) Domestic Intelligent Computing Center, based on Moore Threads' full-function GPUs, is planned for a scale of 10,000 cards. It was lit up in June this year, and the first phase's deployment of 2,112 Moore Threads AI training and inference integrated smart computing cards (MTT S5000) has been fully operational, achieving essentially full-load operation.
Recently, ChipInsider visited the Phase I data center of the Wuxi (Huishan) Domestic 10,000-card Level Intelligent Computing Center. Four Moore Threads GPU cluster modules are neatly arranged, each integrating approximately 500 cards, currently handling real-world tasks.
Fang Zhaoxia, Vice President and General Manager of Ecosystem Cooperation at Moore Threads, told ChipInsider that achieving full load for a 2,000-card cluster within three months places it in the first tier of domestic intelligent computing center construction levels, based on publicly available information. This high efficiency is attributed to Moore Threads' adoption of a "turnkey" engineering model, providing an end-to-end cluster solution that integrates servers, networks, and storage. Once the infrastructure is ready, it takes only one month to rack and deploy into operation.
As a benchmark project for the commercialization of domestic compute power, the Wuxi (Huishan) Domestic Intelligent Computing Center also undertakes the critical function of the Huishan Node in the Southern Jiangsu Compute Agglomeration Zone of the Yangtze River Delta Hub under the 'Eastern Data, Western Computing' Project, providing a highly referential template for the commercial implementation of domestic GPU intelligent computing clusters.
01.2,000 Domestic GPUs Running at Full Capacity, Already Attracting Physical AI Clients
The Wuxi (Huishan) Domestic Intelligent Computing Center was planned for a 10,000-card scale from its inception, adopting a modular delivery approach. Moore Threads provides full-stack compute support using its fully self-developed Ku'a Intelligent Computing Cluster System.
The Phase I project was lit up in June this year, entered trial operation in July, and began commercial operation in October. During the trial period, it connected to 33 mainstream large models. Currently, there are 264 domestic GPU servers and 2,112 computing cards. More than 20 clients have started testing or usage, and dozens of organizations are waiting to enter the cluster for testing. Its compute utilization rate reached 92.9%, with paid compute accounting for 59.7%. It currently serves 55 clients, with training business accounting for over 62%, integrated training and inference business for 14.7%, and inference business for 23.1%.
According to disclosures by the commercial operator, the Huishan Investment Promotion Group, and the co-investor, Mobile Cloud, tests conducted on the Moore Threads MTT S5000 resource pool using the DeepSeek V4 Flash model showed: Under a 7K input / 2K output metric, single-machine output throughput reached 994.97 tokens/s, effective output throughput reached 1,241 tokens/s, maximum concurrent connections meeting SLA standards were 48 channels (achieving 92% of NVIDIA H800 performance), memory bandwidth reached 1.6 Tbps, and memory unit price was less than half that of the H800. This indicates that the MTT S5000 can achieve domestic substitution comparable to the H800 in real business scenarios.
Zhang Tao, Deputy General Manager of the Huishan Investment Promotion Group, shared that the actual utilization rate of the Phase I project is approximately 92%~93%, with paying users accounting for more than 60%. Currently, it primarily serves local universities, research institutes, medical, and financial institutions, running existing base models to meet the needs of agent applications in local vertical domains.
In terms of operational mode, the Wuxi (Huishan) Domestic Intelligent Computing Center is the first to explore a new market-oriented operational model for purely domestic commercial intelligent computing centers, bridging the entire chain from hardware deployment and model adaptation to commercial delivery, with cluster service pricing referencing equivalent market rates.
Huishan is located in the Yangtze River Delta region, close to enterprises, universities, and research institutions. Compute services are easier to integrate with local industrial scenarios and R&D teams, offering a distance advantage for businesses requiring repeated adaptation and joint tuning. Currently, the intelligent computing center has signed compute business cooperation agreements with ten enterprises and universities, covering multiple frontier tracks such as model training, embodied intelligence, AI video generation, spatial intelligence, and life sciences.
02.Single Chip Supports Both AI Computing and Graphics Rendering, Building Full-Function GPUs Comparable to International Mainstream Products
For clients, a major headache when switching compute platforms is software migration. Existing code and development workflows are built around the original CUDA platform. Switching GPUs incurs migration and tuning costs that consume engineering resources and pose risks of performance and stability falling short of expectations, potentially impacting iteration progress.
Moore Threads adopts a dual-track ecosystem strategy, developing its own ecosystem based on the self-developed MUSA architecture while maintaining compatibility with international mainstream CUDA. This allows users to retain their existing development habits and reduces the cost of migrating from other platforms to Moore Threads clusters. Langjing Innovation, a company derived from technology transfer results from Peking University's Graphics Intelligence Lab, whose core business involves rendering, simulation, and 3D reconstruction, is also exploring world models and physical prior AI models. The company previously collaborated with Moore Threads to achieve the first instance of running an aircraft simulator on domestic graphics cards.
Zhou Ling, Vice President of Langjing Innovation, shared his experience using the Moore Threads cluster: On the experimental side, after migration and adaptation, both performance and stability met R&D requirements, with no differences in results caused by using domestic cards. Zhou Ling stated that in the field of physical AI, visual observation and physical state need to be completely parallel; any drift will cause the entire model to fail. Therefore, time sequence and parallel computing were the main factors for Langjing Innovation choosing the Wuxi (Huishan) Domestic Intelligent Computing Center and Moore Threads' full-function GPUs.
"Give me a Ferrari, and I won't drive it well; give Schumacher a Xiali, and he still can't win the championship," he said. He believes that successfully running a project is a bucket effect—it doesn't depend on one particularly long plank, but rather on sufficiently strong ecological barriers. This is the significance of the collaborative effort between Langjing and Moore Threads. In physical AI applications such as simulation and synthetic data, autonomous driving, robotics, and world models, stages like AI model training, virtual scene rendering, and object process simulation appear in the same workflow. This requires a general-purpose compute foundation capable of unified hosting and flexible scheduling.
Many domestic GPUs on the market are either designed specifically for AI computing or focus on image rendering, limiting application scenarios. However, the "full-function" characteristic of Moore Threads GPUs precisely meets the actual needs of these complex workflows. Its full-function GPU, based on the self-developed MUSA architecture, supports different types of computing demands simultaneously on a single chip, including AI computing acceleration, graphics rendering, physical simulation and scientific computing, and ultra-high-definition video encoding/decoding. It provides an integrated general-purpose compute foundation for "compute, render, simulate," exhibiting scarcity in the domestic GPU market.
As one of the few domestic GPUs natively supporting full-precision computing from FP8 to FP64, Moore Threads GPUs comprehensively match international mainstream products in functionality. The MTT S5000 has passed the national "Safety and Reliability Assessment" in the first batch. For the AI field, its full-function GPUs support the training and inference of large language models, embodied intelligence models, and world models. For graphics rendering, it was the first in China to support the DirectX 12 graphics acceleration engine, while also being the first to natively support AI generative rendering and hardware-level ray tracing capabilities. As the physical AI route continues to change rapidly, this "full-function" characteristic provides intelligent computing clusters with more options to adapt to diverse client businesses.
03.Connecting Three Major AI Factories, Expanding to 10,000-Card Clusters On Demand
During this year's World Artificial Intelligence Conference, Zhang Jianzhong, Founder, Chairman, and CEO of Moore Threads, proposed building three major "AI Factories": Model Training Factory, Token Production Factory, and Agent Production Factory. The Wuxi (Huishan) Domestic 10,000-card Level Intelligent Computing Center is one of the benchmark achievements of Moore Threads' AI Factory implementation.
- Model Training Factory: Model training demand accounts for more than half, with multiple embodied intelligence companies conducting training for VLA, VLM, and world models.
- Token Production Factory: The cluster supports over 30 mainstream models from series including Qwen, GLM, DeepSeek, Kimi, and MiniMax.
- Agent Production Factory: Has incubated applications covering agents in industries such as local healthcare, education, and government affairs, accelerating AI technology implementation.
Why did the intelligent computing center, planned for 10,000 cards, only build 2,000 cards in Phase I?
Zhang Tao, Deputy General Manager of the Huishan Investment Promotion Group, said that the Wuxi (Huishan) Domestic Intelligent Computing Center adopts a "step-by-step" strategy. Phase I first builds a cluster scale capable of supporting common client businesses, aiming to increase market acceptance of domestic GPUs, complete adaptation, verification, and client conversion. Then, based on actual demand, it seeks investors to properly advance the expansion of Phase II and III clusters, ultimately forming a 10,000-card scale.
Domestic GPUs generally face the challenge of gaining client trust upon entering the market. Clients want to verify and see results first before being willing to invest more. Therefore, the 2,000 cards in Phase I not only provide compute services but also serve as a platform for clients to verify domestic GPUs, tune software, and evaluate subsequent investments.
As one of the few domestic enterprises with training-grade GPU capabilities, Moore Threads builds full-stack self-developed capabilities from architecture, chips, clusters to software stack, centered on full-function GPUs. Based on the technical advantages of full-function GPUs and cluster engineering delivery capabilities, it provides a sustainable, high-performance, and highly reliable engineering foundation for the Wuxi (Huishan) Domestic 10,000-card Level Intelligent Computing Center.
According to disclosures by Moore Threads, its previous 10,000-card level intelligent computing cluster based on MTT S5000 has been deployed and launched for service, with several key metrics reaching international mainstream levels: floating-point compute power reaches 10 EFLOPS, effective training time can reach 90%, and training linear scaling efficiency reaches 95%.
To address the mismatch between prefill (compute-intensive) and decode (memory bandwidth-intensive) characteristics in large model inference, Moore Threads proposed a hybrid inference solution. It utilizes the high compute power of MTT S5000 for the prefill stage and leverages high memory bandwidth GPUs like NVIDIA H20 for the decode stage.
In relevant use cases, compared to using only H20, its solution improved overall throughput score by approximately 1.87 times and reduced Time To First Token (TTFT) latency by about 50%, achieving equivalent performance to "3 S5000s + 2 H20s ≈ 9 H20s," significantly enhancing cost-performance ratio.
Deep software-hardware optimization is also a crucial link. Fang Zhaoxia noted that after some partners migrated projects originally running on NVIDIA cards to Moore Threads GPUs, due to deep optimization done by Moore Threads at the underlying level with these projects' software, the performance after completion reached several times the previous level. From R&D verification to larger-scale training, the Wuxi (Huishan) Domestic Intelligent Computing Center is gradually taking on new demands brought by client business growth.
Neolix has completed autonomous driving model training verification on the MTT S5000 at this intelligent computing center. Next, it will expand the training scale, with compute demand approaching the 10,000-card level, which is five times the current number of operational cards in the cluster. According to Zhang Tao, in the future, the expansion of the Wuxi (Huishan) Domestic Intelligent Computing Center will consider heterogeneous modes, combining Moore Threads' PD separation heterogeneous inference solution to optimize cost-performance and shorten the break-even cycle.
Expanding to the 10,000-card level will bring a new round of challenges. Large-scale training requires frequent data exchange among GPUs, where network, switching, storage, and fault recovery all impact training output. Fang Zhaoxia said that Moore Threads has accumulated relevant collaborative experience and may adopt a more intensive supernode approach for delivery in the future.
04.Conclusion: From Infrastructure and R&D to Ecological Innovation, Moore Threads Is Forming a Complete Business Layout in Wuxi
In the process of domestic GPU marketization, what needs to be expanded is not just hardware infrastructure, but also client trust, usable software, and sustainable business. The Wuxi (Huishan) Domestic Intelligent Computing Center is providing phased answers. Moore Threads is comprehensively deepening its business layout in Wuxi through continuous investment in talent, technology, and ecosystem.
The Phase I of the Wuxi (Huishan) Domestic 10,000-card Level Intelligent Computing Center will complement the Moore Threads Wuxi R&D Center, Moore Threads System-Level Product R&D Headquarters, Moore Threads Wuxi Future Center, Moore Threads (Wuxi) Industrial Embodied Intelligence Innovation Center, and the Wuxi Huishan District AI Education Practical Training Base, forming a complete layout from technology and product R&D, compute infrastructure, and application demonstration to talent cultivation.
Whether for current large model training and inference or for future explosive scenarios like physical AI and embodied intelligence, Moore Threads can provide full-scenario, full-cycle robust compute support. It is reported that Moore Threads' next goal is 100,000 cards, and it has established deep cooperation with JD Cloud, planning to build a 100,000-card cluster.