Moore Threads demonstrated its concept of three AI factories at the WAIC 2026 event: a model training factory, a token generation factory, and an agent production factory. These factories constitute a comprehensive production infrastructure necessary for the era of tokenomics, where the demand for computing power grows exponentially due to the consumption of each other by AI systems.
All three factories are built on the Kuae computing cluster and the company's patented MUSA architecture. This architecture uniquely integrates AI acceleration, 3D graphics rendering, physical simulation, and video codecs into a single chip.
The most significant achievement is the EvoPhys-World model—the world's first 5D model fully trained on Moore Threads' MTT S5000 GPUs. This model ranked first in the Stanford WorldScore global ranking for world generation for 37 consecutive days. A research group from Peking University utilized the Kuae cluster with full support of the MUSA software stack, demonstrating that domestic GPU infrastructure is capable of supporting the training of advanced global models with quality, stability, and scalability comparable to international standards. This confirms the potential of domestic GPUs for complex AI tasks previously considered the prerogative of NVIDIA.
The technical foundation is the MTT C256 supernode, which uses a single-layer Scale-up network architecture. It provides full interconnectivity for 128 GPUs in a standard single-node width chassis, with the possibility of expanding up to 256 GPUs in a combined configuration. The orthogonal structure eliminates issues of bandwidth loss and high latency found in traditional multi-level networks. The model's FLOPs utilization rate reaches 60% for dense models and 40% for MoE models, effective training time exceeds 90%, and linear scaling efficiency reaches 95%, while training accuracy matches major international computing cards. The Kuae cluster successfully demonstrated stable training on 1000 cards for the multimodal model Wan2.1 and completed end-to-end training of the base MoE-236B model from scratch.
Moore Threads also introduced two AI-oriented terminals: the MTT AICUBE home AI center, positioned as a family smart hub, and the MTT AIBOOK laptop for mobile AI tasks. Combined with the MusaCoder code model, the MUSACODE AI programming agent, and the Moore Academy, the company presented a complete stack of solutions—from cloud supercomputing infrastructure to edge deployment and end-user terminals. This allows the company to position itself as the only Chinese GPU manufacturer that simultaneously covers both corporate and consumer markets with a complete scenario coverage strategy.
Beyond the three factories, a broader conclusion is that the Chinese ecosystem of domestic GPUs has reached a level of readiness for industrial application. BAAI used the Kuae cluster with MTT S5000 to complete the full training process of the generalized embodied brain RoboBrain 2.5. The BEVFormer benchmark on Moore Threads infrastructure achieved throughput on one and two machines comparable to leading international GPUs. Since daily token consumption in China exceeded 140 trillion in March 2026, increasing 1000-fold in two years, the ability to train, run, and deploy AI workloads on domestic GPU infrastructure has transformed from a theoretical possibility into an operational reality.
At the WAIC 2026 exhibition, computing clusters featuring domestic AI supernodes are being demonstrated, utilizing Sugon Dawn and Hygon processors and systems scaled up to 100 thousand cards. This indicates a shift in focus within AI computing from competing individual chips to the system level.
The World Artificial Intelligence Conference 2026 signals a fundamental change in the narrative surrounding Chinese AI computing. Instead of comparing individual chips against NVIDIA GPUs, the focus has shifted to systemic computing using domestic supernode clusters capable of competing with the world's best AI infrastructure.
A central exhibit illustrating the transition of Chinese AI computing into a systemic era is Sugon's Dawn 8000 super-AI cluster. This 100,000-card cluster is built on Hygon processors and connected to the National Supercomputer Internet.
Numerous local suppliers are presenting supernode-scale solutions at WAIC. Sugon showcases its super-AI architecture, which integrates computational precision from FP64 to INT8 within a single cluster, eliminating the traditional separation between scientific computing infrastructure and AI training. Hygon displays its latest high-performance processors, specifically designed for large-scale AI tasks, featuring improved memory bandwidth and inter-chip connectivity capabilities.
Other chip manufacturers are also presenting latest-generation products optimized for supernode deployment. The emphasis of all participants is no longer on matching or exceeding the performance of a single NVIDIA H100 or B200 chip, but on demonstrating that domestic chip clusters can achieve competitive training throughput at the system level for models up to one trillion parameters when paired with the correct architecture featuring optimized network, storage, and software stacks.
WAIC also features several next-generation AI computing innovations, such as liquid-cooled supernode racks, disaggregated memory architectures for large model training, and native domestic software stacks ensuring high utilization of heterogeneous accelerators. Some exhibitors are demonstrating training sessions on clusters exceeding 10,000 cards, using exclusively domestic hardware and software, while showing strong scaling efficiency.
This systemic focus reflects a mature understanding that leadership in AI computing is determined by the integrated performance of the entire infrastructure chain, not by the characteristics of individual chips. Network bandwidth, memory architecture, cooling efficiency, power delivery, supported software toolkits, cluster management software, and application optimization contribute more to actual training throughput than the raw performance of the chip itself.
This structural view directs the development of Chinese AI computing toward holistic system optimization rather than component-level competition. Industry observers note that this mirrors the trajectory of global AI computing, where even NVIDIA transformed from a GPU manufacturer into a company creating data center systems with DGX and HGX platforms. The supernode demonstrations at WAIC show that China's domestic AI computing ecosystem has reached a stage where systemic competition is feasible, even if the performance of individual chips still lags behind cutting-edge foreign processors in peak specifications.
Dexmal has introduced a comprehensive suite of products for embodied artificial intelligence aimed at overcoming the final engineering barrier between AI models and real-world performance. Among the announced developments are the universal foundational model DM0.5, the Apex universal robot, the DexOS operating system, and the MaaS platform.
The company's strategy is built on a three-stage model, analogous to a rocket: the first stage focuses on creating a generalized foundational model; the second stage includes developer tools such as MaaS and the OS; and the third stage concentrates on systematic scenario solving capabilities.
The DM0.5 foundational model possesses 4 billion parameters and was trained on 150,000 hours of multi-source data. This represents a fourfold increase in data and a twofold increase in parameters compared to the DM0 model. The model demonstrates a low inference latency of 50 ms, a 25% improvement in inference efficiency, and achieved first place in the RoboChallenge Table30 V2 ranking with an overall score of 60.1, while showing 99.1% complex performance when evaluated by LIBERO.
The key technical breakthrough is the ability to generalize. Compared to DM0, zero-shot navigation success improved by 31%, multi-shot mode success improved by 45%, and fine-tuning success improved by 20%. The cost of fine-tuning decreased by 60%, allowing expert-level tasks to be performed on a single consumer RTX 4090 GPU in just 18 hours.
Three auxiliary platforms help bridge the engineering gap between model capabilities and real-world deployment. DFOL 2.0 utilizes the world model DW0.5 as a high-precision simulation for reinforcement learning in a virtual environment, reducing the need for real robot data by 60% and overall training costs by 40%. DexOS, the industry's first generalized operating system for embodied systems, decouples AI models from robot hardware, allowing developers to control, perceive, and connect robots with just a few lines of Python code, regardless of the robot brand.
The MaaS platform enables one-command zero-shot inference and provides a visual workspace for online training and deployment with pay-as-you-go billing.
The universal Apex robot platform, specifically designed for embodied AI applications, was also presented. It features a modular design with the ability to quickly swap end effectors in under 60 seconds, 360-degree scanning using two lidars, a 3 kg manipulator payload, millimeter precision, and hot-swappable batteries with a brain standby mode for continuous operation. The target MTBF exceeds 1000 hours.
CEO Tang Wenbin noted that embodied AI is at a critical inflection point where scaling laws fully demonstrate their power, but transitioning from models to real productivity requires overcoming the final hurdle of engineering integration. Dexmal is collaborating with robot manufacturers, including Tiangong, Huaqin, and Shihe, as well as chip partners such as Pingtouge, Moore Threads, and Cambricon, to build a domestic embodied AI ecosystem.
