On May 18th, Moore Threads unveiled its latest cloud-edge-end product matrix in Beijing. The event was packed with significant announcements. Products such as the Kuae WanKa-level intelligent computing cluster, the Yangtze SoC, AICUBE, AIBOOK, the E300 edge module, the Wheat Intelligent Agent, the MT Lambda embodied intelligence simulation platform, and progress on the MUSA software ecosystem were all presented in a single launch.
Beyond the product list, the more noteworthy aspect was the strategic organization of these offerings. Moore Threads connected cloud training, software migration, edge devices, terminal agents, and embodied simulation into a cohesive narrative, sending a clear signal that the market's perception of the company is shifting from a GPU chip provider to an AI infrastructure company. In recent years, domestic GPU companies have frequently faced direct questions. Can the cards run? What is the performance relative to competitors? Are mainstream frameworks compatible? Can the CUDA ecosystem be migrated? Can mainstream models be quickly adapted? These questions remain important. However, as AI enters the Agent phase, customer inquiries have become more specific. A large model company is concerned about whether a cluster can sustain continuous training and recover from hardware failures. An autonomous driving company cares about connecting world models and simulation pipelines. A robotics company focuses on whether trained policies can be deployed to edge devices. Enterprise clients also calculate migration and maintenance costs. Single-card performance is the entry point, but system capabilities ultimately influence procurement and repurchase decisions. The system capability is precisely what Moore Threads aimed to emphasize this time.
The cloud is currently its most critical proving ground. Materials indicate that the Kuae WanKa-level intelligent computing cluster is already operational, achieving an MFU of 60% for dense model training, 40% for MoE models, a linear scaling efficiency of 95%, and an effective training uptime of 90%. The complexities of large model training often become apparent only at scale. More GPUs bring greater computing power but also introduce challenges in communication, scheduling, fault tolerance, storage, cooling, and framework adaptation. Longer training cycles make system stability paramount. The Kuae cluster's task is to demonstrate to enterprise clients that Moore Threads possesses the capability for system-level delivery. Building upon its cloud capabilities, Moore Threads is also enhancing its software stack. It has already adapted to major domestic models like DeepSeek, GLM, MiniMax, Kimi, and Qwen, gained official native support in the SGLang mainline code, and open-sourced vLLM-MUSA. The MUSA SDK 5.1.0 aligns with CUDA 12.8 and fully supports all 3,194 PyTorch operators. The strategic value of these developments lies in reducing the friction for developers migrating to Moore Threads GPUs. The difficulties of domestic GPU ecosystems often lie in the long tail. While mainstream frameworks may run, a company's legacy engineering projects might not migrate smoothly. Even after a model is adapted, subsequent versions require ongoing support. With operators being supplemented, custom kernels and dependencies on older versions may still exist in business applications. Developer patience with a hardware platform is often tested by such details.
Moore Threads' emphasis on MUSA compatibility, the open-sourcing of vLLM-MUSA, official SGLang support, along with tools like Automusify for automatic migration and the MUSACODE programming assistant, all address the same core issue: pushing domestic GPUs from being merely usable to genuinely user-friendly. In the edge and terminal segments, a more proactive stance from Moore Threads is evident.
Based on the Yangtze SoC, Moore Threads introduced AICUBE, AIBOOK, and the E300. AICUBE targets the home scenario, integrating the Wheat Intelligent Agent, AI PC, and AI NAS, attempting to consolidate family data, device control, and agent services into a single entry point. AIBOOK targets developers and learners, running MTT AIOS, pre-installed with the Lobster Intelligent Agent OpenClaw, and supports multi-agent collaboration. The E300 is designed for edge scenarios like industrial quality inspection, energy patrol, embodied intelligence, smart vehicles, and the low-altitude economy, offering 50 TOPS of heterogeneous AI computing power, emphasizing local inference, low latency, and stable operation. These products introduce Moore Threads to a more complex market. Home users prioritize high-frequency needs, developers focus on workflow integration, and industry clients evaluate deployment costs, failure rates, and service responsiveness. The value of AICUBE, AIBOOK, and the E300 will be measured by usage frequency, developer retention, and industry project reuse rates. Within this matrix, MT Lambda is a critical variable.
Moore Threads defines MT Lambda as a full-stack embodied intelligence simulation platform. It leverages a full-feature GPU to integrate rendering, physics, and AI computing on a single chip, while the upper layer provides tools for data synthesis, policy training, and simulation verification. This component propels Moore Threads' GPU narrative into the realm of physical AI. Large model training primarily tests cloud computing power. Embodied intelligence broadens the requirements—robots, autonomous vehicles, and industrial equipment need to understand their environment and act within the physical world. They rely on language, vision, motion, physical simulation, graphics rendering, and real-time edge-side response. The cost of trial-and-error in the real world is high. Robots can fall, equipment can break, production lines can be disrupted, and autonomous driving cannot rely on endless real-world experimentation. Hence, simulation training becomes essential infrastructure. This is precisely why Moore Threads emphasizes its full-feature GPU. In embodied intelligence, graphics rendering, physical simulation, and AI computing must be evaluated on a unified computational foundation. The entity that can efficiently generate credible synthetic data and complete policy training and verification in a virtual environment will lower the cost for robotics and autonomous driving to enter real-world scenarios. Increased inter-company collaborations also serve this direction. Moore Threads collaborated with the Beijing Academy of Artificial Intelligence to complete RoboBrain 2.5 training. It is also working with partners like Guanglun Intelligence, Pony.ai, 51VR, and Rays Cloud on adaptation in areas such as simulation data, world models, autonomous driving, and embodied simulation platforms. Viewing the cloud, software, terminals, and simulation together clarifies the company strategy presented at this launch. The Kuae cluster handles large-scale AI computing, the MUSA ecosystem reduces migration costs, the Yangtze SoC and end-side products enter devices and scenarios, and MT Lambda integrates into the embodied intelligence workflow. This is a path extending from chips to systems, platforms, and scenarios. The advantage is that Moore Threads can reduce its reliance on single-point hardware sales and embed itself more deeply into customers' AI system development. When purchasing AI infrastructure, clients simultaneously evaluate computing power pricing, stability, migration costs, service capabilities, and scenario outcomes. The risks are equally apparent. With each deeper layer, the evaluation criteria change. The cloud is judged on scale delivery and stability, software on developer experience, terminals on meeting high-frequency demands, and embodied intelligence on real-world scenario validation. Moore Threads must prove itself across multiple battlefields simultaneously. The launch presented a structurally complete blueprint. What remains to be validated is whether each node can function effectively. Can models trained on the Kuae cluster be deployed to the edge and terminals? Can strategies generated by MT Lambda be integrated into the workflows of robotics and autonomous driving customers? Can the MUSA ecosystem reduce developer migration costs to an acceptable level? From an industry trend perspective, AI competition is spilling over from model capabilities to system capabilities. Over the past two years, industry focus has centered on parameter scale, context length, multimodal performance, inference cost, and Agent capabilities. Models remain core, but the industry is now confronting more concrete problems: how to stably supply computing power, how to generate data, how to train and deploy models, and how robots can learn in simulation and execute in reality. These problems require infrastructure to address them. The significance of Moore Threads' launch lies in placing itself within this new paradigm. The next breakthrough for domestic GPU companies will stem from system delivery, software ecosystems, and scenario closure. The direction Moore Threads outlines involves entering the realms of large-scale intelligent computing clusters, edge terminals, and the embodied intelligence simulation ecosystem. Next, the market will test whether genuine synergy can be formed among these products.
Comments