Phase 1 of Moore Threads' Wuxi domestic intelligent computing center is running at full capacity with 2,112 GPUs

A domestic GPU cluster with zero NVIDIA content has just achieved a "big result."

Xindongxi reported on November 2 that on Saturday, domestic GPU leader Moore Threads, together with the Wuxi Huishan Investment Promotion Group and Neolix, a leading domestic L4 autonomous driving unmanned vehicle company, announced a strategic cooperation. Neolix plans to conduct larger-scale autonomous driving model training at the Wuxi (Huishan) Domestic Intelligent Computing Center, which will require nearly ten-thousand-card-level computing power.

The Wuxi (Huishan) Domestic Intelligent Computing Center, based on Moore Threads' full-function GPUs, is planned at a ten-thousand-card scale and was lit up in June this year. The first phase's 2,112 Moore Threads MTT S5000 AI training-and-inference integrated intelligent computing cards have been fully put into operation and are basically running at full load.

Recently, Xindongxi visited the first-phase machine room of the Wuxi (Huishan) Domestic Ten-Thousand-Card-Level Intelligent Computing Center. Four Moore Threads GPU cluster modules are neatly arranged, each integrating about 500 cards and carrying real tasks.

Fang Zhaoxia, Vice President of Moore Threads and General Manager of Ecosystem Cooperation, told Xindongxi that achieving a full load of a two-thousand-card cluster within three months, based on public information, ranks among the first tier in domestic intelligent computing center construction. Such efficiency is due to Moore Threads' adoption of a "turnkey" engineering model, providing an end-to-end cluster solution integrating servers, networking, and storage.

Once the infrastructure is ready, it takes only one month to complete racking and put it into operation.

As a benchmark project for the commercial deployment of domestic computing power, the Wuxi (Huishan) Domestic Intelligent Computing Center also serves an important function as the Huishan node of the Sunan computing power cluster in the Yangtze River Delta hub of the "East Data, West Computing" initiative, and provides a highly referential model for the commercial deployment of domestic GPU intelligent computing clusters.

1. Two thousand domestic GPUs running at full capacity, already attracting physical AI customers

Since the beginning of construction, the Wuxi (Huishan) Domestic Intelligent Computing Center has been planned at a ten-thousand-card scale, adopting a modular delivery approach, with Moore Threads providing full-stack computing power support and using Moore Threads' fully self-developed KUAE intelligent computing cluster system.

The first phase was lit up in June this year and began trial operation in July. During the trial operation period, 33 mainstream large models were connected. Commercial operation began in October, with a total of 264 domestic GPU servers and 2,112 computing cards. More than 20 customers have conducted testing or use, and dozens more units are waiting to enter the cluster for testing.

Its computing power utilization rate is 92.9%, of which paid computing power accounts for 59.7%; it currently has connected with 55 customers, with training business accounting for more than 62%, training-and-inference integrated business accounting for 14.7%, and inference business accounting for 23.1%.

According to disclosures by the commercial operator Huishan Investment Promotion Group and co-investor China Mobile Cloud, actual testing on the Moore Threads MTT S5000 resource pool at the intelligent computing center using the DeepSeek V4 Flash model showed: under a 7K input/2K output specification, single-machine output throughput reached 994.97 tokens/s, effective output throughput reached 1,241 tokens/s, the maximum concurrency meeting SLA reached 48 channels (92% of NVIDIA H800), memory bandwidth reached 1.6 Tbps, and memory unit price was less than half that of H800.

This shows that MTT S5000 can achieve a domestic replacement benchmarked against H800 in real business scenarios.

Zhang Tao, Deputy General Manager of Huishan Investment Promotion Group, shared that the actual utilization rate of the first phase is about 92% to 93%, with paying users accounting for more than 60%. It currently mainly serves local universities, research institutes, medical and financial institutions, running existing foundation models to meet local vertical-domain agent application needs.

In terms of operating model, the Wuxi (Huishan) Domestic Intelligent Computing Center has taken the lead in exploring a new market-oriented operating model for a purely domestic commercial intelligent computing center, connecting the entire chain from hardware deployment and model adaptation to commercial delivery, with cluster service pricing referenced to an equivalent ratio of market prices.

Huishan is located in the Yangtze River Delta region, close to enterprises, universities, and research institutions, making it easier to combine computing services with local industrial scenarios and R&D teams, and giving it a distance advantage for businesses requiring repeated adaptation and joint tuning.

2. A single chip can simultaneously run AI computing and graphics rendering, building a full-function GPU benchmarked against mainstream international products

For customers, one of the biggest headaches when switching computing platforms is software migration.

Existing code and R&D processes are built around the original CUDA platform. Switching GPUs brings migration and tuning work that consumes engineering resources, and also carries the risk that performance and stability will fall short of expectations, possibly even affecting iteration progress.

Moore Threads adopts a dual-track ecosystem strategy: it develops its own ecosystem with the self-developed MUSA architecture while also being compatible with mainstream international CUDA, allowing users to maintain their original development habits and reducing the cost of migrating from other platforms to Moore Threads clusters.

Langjing Innovation is a company originating from the commercialization of scientific and technological achievements at Peking University's Graphics Intelligence Laboratory. Its core business involves rendering, simulation, and 3D reconstruction, and it is also exploring world models and physical prior AI models. The company previously cooperated with Moore Threads to achieve, for the first time, running an aircraft simulator on domestic graphics cards.

Zhou Ling, Vice President of Langjing Innovation, shared his experience using Moore Threads clusters: on the experimental side, after migration and adaptation, performance and stability both met R&D requirements, with no differences in results caused by using domestic cards.

Zhou Ling said that in the physical AI field, visual observation and physical state need to be fully parallelized; once drift occurs, the entire model will produce errors. Therefore, timing and parallel computing are the main factors why Langjing Innovation chose 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 won't win a championship either." He believes that for a project to run successfully, it comes down to the barrel effect: it's not about any single stave being especially long, but about the ecological barriers being strong enough—and this is precisely the significance of the collaboration between Langjing and Moore Threads.

In physical AI applications such as simulation and synthetic data, autonomous driving, robotics, and world models, links including AI model training, rendering of virtual scenes, and simulation of object processes all occur within the same workflow. This requires a general-purpose computing foundation capable of unified hosting and flexible scheduling.

Many domestic GPUs on the market are either designed specifically for AI computing or focused on image rendering, which limits their application scenarios. The "full-function" characteristics of Moore Threads GPUs happen to meet the actual needs of the above complex workflows.

Its full-function GPU is based on the self-developed MUSA architecture, with a single chip simultaneously supporting different types of computing needs such as AI computing acceleration, graphics rendering, physical simulation and scientific computing, and ultra-high-definition video encoding and decoding, providing an integrated general-purpose computing foundation for "computing, rendering, and simulation" and possessing scarcity in the domestic GPU market.

As one of the very few domestic GPUs natively supporting full-precision computing from FP8 to FP64, Moore Threads GPUs fully benchmark against mainstream international products in functionality, and the MTT S5000 has passed the national "Security and Reliability Evaluation" in the first batch.

For the AI field, its full-function GPU supports the training and inference of large language models, embodied intelligence models, and world models; for graphics rendering, it is the first in China to support the DirectX 12 graphics acceleration engine, while also taking the lead in natively supporting AI generative rendering and hardware-level ray tracing capabilities.

Model training factory: model training needs account for more than half, and multiple embodied intelligence enterprises are conducting training of VLA, VLM, and world models.

Token production factory: the cluster supports more than 30 mainstream models of series such as Qwen, GLM, DeepSeek, Kimi, and MiniMax.

For an intelligent computing center planned at the ten-thousand-card scale, why build only two thousand cards in the first phase?

Zhang Tao, deputy general manager of Huishan Investment Group, said that the Wuxi (Huishan) Domestic Intelligent Computing Center adopts a "step-by-step" strategy: the first phase first builds a cluster scale capable of supporting general customer business, aiming to improve market acceptance of domestic GPUs, complete adaptation, verification, and conversion of customers, and then seek investors based on actual demand to properly advance the second and third phase cluster expansion, ultimately forming a ten-thousand-card scale.

Domestic GPUs generally face the difficulty of gaining customer trust when entering the market. Customers need to verify first and see results first before they are willing to invest more. Therefore, the two thousand cards in the first phase not only provide computing power services but also become a platform for customers to verify domestic GPUs, debug software, and evaluate subsequent investment.

As one of the very few domestic enterprises with training-grade GPU capabilities, Moore Threads takes full-function GPUs as its core and has built full-stack self-developed capabilities from architecture, chips, and clusters to software stacks. 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 ten-thousand-card intelligent computing center.

According to Moore Threads, its previously built ten-thousand-card intelligent computing cluster based on the MTT S5000 has been deployed and put into service, with multiple key indicators reaching international mainstream levels: floating-point computing power reaches 10 EFLOPS, effective training time can reach 90%, and training linear scaling efficiency reaches 95%.

Addressing the mismatch between the characteristics of prefill (compute-intensive) and decode (memory-bandwidth-intensive) in large model inference, Moore Threads proposed a hybrid inference solution, using the high computing power of the MTT S5000 to handle the prefill stage and using high-memory-bandwidth hardware such as NVIDIA H20 to handle the decode stage.

In relevant use cases, compared with using only H20, its solution improves overall throughput score by about 1.87 times and reduces first-token latency by about 50%, achieving equivalent performance of "3 S5000 + 2 H20 ≈ 9 H20," significantly improving cost-effectiveness.

English text is being prepared.

From R&D verification to larger-scale training, the Wuxi (Huishan) Domestic Intelligent Computing Center is gradually taking on new demand brought by the growth of customer business.

Neolix has completed autonomous driving model training verification based on the MTT S5000 at the intelligent computing center, and will next expand the training scale, with computing power demand approaching the ten-thousand-card level, that is, 5 times the number of cards currently put into operation in the cluster.

According to Zhang Tao, in the future, the expansion of the Wuxi (Huishan) Domestic Intelligent Computing Center will consider a heterogeneous model, combining Moore Threads' PD-separated heterogeneous inference solution to optimize cost-effectiveness and shorten the break-even period.

Expanding to the ten-thousand-card level will bring a new round of tests. Large-scale training requires GPUs to exchange data frequently, and networking, switching, storage, and fault recovery will all affect training output. Fang Zhaoxia said that Moore Threads has accumulated relevant collaboration experience and may in the future also adopt a more intensive supernode approach for delivery.

Conclusion: From infrastructure and R&D to ecosystem innovation, Moore Threads is forming a complete business layout in Wuxi

In the marketization process of domestic GPUs, what needs to be expanded is not only hardware infrastructure, but also customer trust, usable software, and sustainable business. The Wuxi (Huishan) Domestic Intelligent Computing Center is providing a phased answer.

Through continuous investment in talent, technology, and ecosystem, Moore Threads is comprehensively deepening its business layout in Wuxi. The first phase of the Wuxi (Huishan) domestic ten-thousand-card intelligent computing center will echo with the Moore Threads Wuxi R&D Center, the Moore Threads System-Level Product R&D Headquarters, the Moore Threads Wuxi Future Center, the Moore Threads (Wuxi) Industrial Embodied Intelligence Innovation Center, and the AI Education and Training Base in Huishan District, Wuxi, forming a complete layout from technology and product R&D, computing infrastructure, and application demonstration to talent cultivation.

Whether for current large model training and inference or for future explosive scenarios such as physical AI and embodied intelligence, Moore Threads can provide strong computing power support for all scenarios and the full cycle. According to the introduction, Moore Threads' next goal is 100,000 cards, and it has reached a deep cooperation with JD Cloud to plan to build a 100,000-card cluster.