On October 31, Moore Threads signed a strategic cooperation agreement with Wuxi Huishan Investment Promotion (Group) Co., Ltd. and Neolix Technology Co., Ltd. (hereinafter referred to as "Neolix"). Relying on the Wuxi (Huishan) Domestic AI Computing Center and leveraging Moore Threads' full-function GPU computing power, Neolix will conduct larger-scale autonomous driving model training to accelerate the implementation of Physical AI, with computing demand approaching the ten-thousand-card level.
The Phase I project of this AI computing center was officially lit up in June this year; in August, the Phase I cluster of 2,000 cards was fully operational. Yet, this Phase I cluster of only 2,000 cards has already efficiently supported the operation and application of six major AI agents, covering high-value industry tracks such as government services, healthcare, smart education, and chip design.
What is special about this domestic AI computing center project with only 2,000 cards in its Phase I? How can it handle the diverse demands from six core scenarios?
2,000 Cards Fully Utilized in Three Months, Performance Stands Up to Scrutiny
On June 11, the Wuxi (Huishan) Domestic AI Computing Center's Phase I was officially lit up. The project is built on Moore Threads' flagship AI training and inference integrated smart computing card MTT S5000, deploying 2,000 Moore Threads full-function GPUs.
By September, the center achieved full-load operation, efficiently supporting the running and application of six major AI agents across high-value industry tracks including government services, healthcare, smart education, and chip design. Over 20 clients are conducting tests or formal business on the platform, with operations covering model training and inference, where training accounts for more than 62% of the workload.
From "lit up" to "full load," the Wuxi (Huishan) Domestic AI Computing Center took only three months. Moreover, the center pioneered a new market-oriented operational model for purely domestic commercial AI computing centers in China, connecting the entire chain from hardware deployment and model adaptation to commercial delivery.
Clients have given very positive feedback on the cluster's performance. Zhou Ling, Vice President of Langjing Innovation, stated that the technical link for deploying tasks on this cluster has been fully established, and both the stability and performance of the cluster meet the predetermined requirements.
Demand Is Queuing Up, Expansion Will Be Carried Out as Needed
Positive market demand is making expansion increasingly urgent. On October 30, the Wuxi (Huishan) Domestic AI Computing Center signed computing service cooperation agreements with ten enterprises and universities, covering frontier tracks such as model training, embodied intelligence, AI video generation, spatial intelligence, and life sciences. As a leading L4-level autonomous driving unmanned vehicle enterprise in China, Neolix has already conducted autonomous driving model training verification at the center based on the MTT S5000, achieving excellent results. With the deepening of autonomous driving technology R&D, they plan to launch larger-scale training.
Currently, computing supply is tight. Zhang Tao, Deputy General Manager of Wuxi Huishan Investment Promotion (Group) Co., Ltd., introduced that many companies are severely lacking in computing power. During the World Internet of Things Expo, numerous enterprises applied on-site to trial the cluster.
Client needs also reflect the hierarchy of computing demand. Among the client groups interfacing with the cluster, the number of enterprises and universities/research institutes is roughly equal, maintaining a ratio of approximately 1:1. However, their demand structures differ significantly. Research institutes focus more on how to run their self-developed new models, while enterprises prioritize optimizing computing performance utilization and rapidly achieving commercial monetization. Correspondingly, the computing demand from enterprises is about three times that of universities and research institutes.
Zhang Tao stated that the Wuxi (Huishan) Domestic AI Computing Center is intentionally restraining the pace of large-scale expansion. The construction logic follows a "step-by-step" approach: The Phase I cluster of 2,000 cards aims to increase market acceptance of domestic GPUs, first ensuring the cluster can adapt to most vendor needs and run smoothly. Subsequently, there are two development paths: if clients have further usage needs, the center will seek investors to promote Phase II and III expansions, eventually forming a ten-thousand-card cluster scale.
"We currently have some clients in negotiation," said Zhang Tao. "Some clients need 8 units, but the actual remaining capacity is insufficient, so these clients have to wait."
Fang Zhaoxia corroborated this judgment from the industry side: Currently available server rooms are consistently at full load, with many people applying on-site to trial the cluster. What many enterprises lack is computing power—"If you can't even access computing resources, how can you develop AI businesses?" According to Zhang Tao, to cope with the vigorous demand for computing power, Huishan has preliminarily formed an intention to build plans, which will be advanced step by step to ultimately form a ten-thousand-card cluster scale.
Full-Function GPUs Solve Current Problems and Bet on Future Opportunities
Sufficient versatility is the direct reason clients choose Moore Threads. Moore Threads' full-function GPUs can simultaneously support various workloads such as AI computing acceleration, graphics rendering, physical simulation and scientific computing, and ultra-high-definition video encoding/decoding. It is one of the few enterprises in China possessing training-grade GPU capabilities.
Why must it be "full-function"?
In a recent media interview, the case of Langjing Innovation provided the most intuitive explanation.
Beijing Langjing Innovation Technology Co., Ltd., founded in 2020, is a technology transfer company from Peking University's Graphics Intelligence Laboratory and a national-level specialized, refined, unique, and innovative key "Little Giant" enterprise. The company focuses on virtual simulation and spatial intelligence technology R&D and services. It independently developed the domestic spatial intelligence engine ViWo, aiming to create a general-purpose spatial intelligence engine covering multiple application fields to provide tools and platforms for virtual simulation applications. Products independently developed based on the ViWo engine are widely used in fields such as aviation industry and digital twins. Its wholly-owned subsidiary Weishiwei specializes in aircraft simulator visual systems, which won the CCF Technology Invention First Prize and were applied in various types of professional flight simulators, breaking the monopoly of foreign systems in this field. Zhou Ling claimed that currently, 80% of the domestic high-end simulator market is served by this company.
In physics prior models, physical states and visual observations must be strictly aligned; once temporal drift occurs, incorrect results are learned. "In the field of physics prior model AI we engage in, timing and parallel computing are our primary concerns," said Zhou Ling.
To meet such scenario requirements, computing infrastructure cannot merely possess computational functions; it also requires computing chips capable of supporting workloads like 3D rendering and scientific computing. For this reason, Moore Threads' full-function GPUs became the preferred choice for Weishiwei's computing configuration.
Moore Threads' principle for chip design is: not only addressing current needs but also facing the future.
"For clients, investing in building a computing cluster is a long-term calculation: A cluster is an asset used for many years. If it can only serve a specific type of demand, the cluster faces the risk of obsolescence," Fang Zhaoxia, Vice President of Moore Threads and General Manager of Ecosystem Cooperation, explained in a recent interview. "AI technology routes are still changing rapidly—today there might be more training demand, tomorrow more inference, and the day after, perhaps agents. Dedicated architectures can only serve specific applications and struggle to support changes in demand."
For this reason, based on its independently developed MUSA architecture, Moore Threads has created a rare "computing, rendering, simulation" integrated general-purpose computing base in China. Choosing a versatile technology route enables Moore Threads' chips to not only adapt to current needs but also satisfy future requirements.
From 2,000 cards to ten thousand cards, the engineering challenge lies in the overall coordination of networks, switches, and storage. Fang Zhaoxia stated that the Wuxi (Huishan) Domestic AI Computing Center is planned at the ten-thousand-card level and delivered in a modular fashion. The technical difficulties of building a ten-thousand-card cluster have been basically overcome, and subsequent phases may adopt a more intensive super-node approach. However, Moore Threads' goal goes beyond this—the company has reached a deep partnership with JD Cloud to build a hundred-thousand-card full-function GPU cluster.
Author | Ji Xiaoting
Editor | Wu Lilin
Art Editor | Ma Liya
Supervisor | Lian Xiaodong