Core Viewpoint - The news highlights the challenges faced by Muxi Co., particularly its significant financial losses, cash flow issues, and high customer concentration, which raise concerns about its upcoming IPO and overall sustainability in the competitive GPU market [1][2][10]. Financial Performance - Muxi Co. reported cumulative losses of 3.29 billion yuan from 2022 to Q1 2025, with annual losses increasing each year [2]. - The company’s R&D expenses reached 2.466 billion yuan, which is 2.2 times its total revenue of 1.116 billion yuan during the same period [3]. - Operating cash flow has been negative, totaling a net outflow of 4.361 billion yuan across the reporting periods [3]. Debt Structure - By the end of 2024, Muxi Co.'s interest-bearing debt was 2.291 billion yuan, with short-term debt constituting 5.17 billion yuan and long-term debt 1.774 billion yuan [4]. - The company’s cash flow from operations only covered about 24% of its total debt, indicating a precarious financial position [4]. - As of Q1 2025, short-term debt accounted for 92% of total debt, raising liquidity concerns [4][6]. Product and Market Position - The Xiyun C500 series chip, launched in February 2024, is critical for Muxi Co., generating 97.28% of its main business revenue in 2024 [7]. - The chip's performance is competitive, but software compatibility issues with existing ecosystems, particularly with NVIDIA, pose significant challenges [9]. Customer Concentration and Risks - Muxi Co. has a high customer concentration, with the top five clients accounting for 91.58% of revenue in 2023, which increases the risk of revenue loss if any major client withdraws [10]. - The relationship between Muxi Co. and its major clients raises concerns about the sustainability and reliability of its revenue streams [10]. Supply Chain Vulnerabilities - The company relies entirely on Taiwanese foundries for its chip production, making it vulnerable to geopolitical risks and export controls [11]. - The reliance on external suppliers for critical components adds to the financial strain, especially with limited cash reserves [11]. Management and Governance Issues - There are concerns regarding the management's actions, including the establishment of related companies and potential conflicts of interest, which could impact investor confidence [11].

（原标题：【WAIC2025】 AI算力创新竞速，国产化实践走出超节点等新路） 在世博展览馆内，为应用提供底座能力的AI芯片、服务器、智算中心等厂商，展示了在芯片底层架构的自主研发，软件、整机的国产化适配，以 及针对应用场景的解决方案等方面的创新尝试。 AI算力创新 在H1核心技术馆内，阶跃星辰、月之暗面、智谱等大模型厂商的展台人流攒动。与这些模型厂商展台错落有致搭配的是沐曦、无问芯穹、摩尔线 程、燧原科技等算力厂商的展台。 伴随模型的迭代演进，人们对算力的需求也呈指数级增长。 正处于上市辅导备案中的沐曦展示了基于曦云C500系列芯片的服务器以及解决方案。展台人员介绍，服务器从编译、驱动到互联等全链路均已实 现国产化。 沐曦还首次展出了曦云C600通用GPU——一颗基于国产供应链设计、制造，自主可控的训推一体自研芯片。 经济观察报记者在展台未发现C600的性能、参数等内容。展台人员介绍，该芯片配置了业内前沿的显存，能强力支撑大模型训练、推理期间的海 量数据吞吐，主要用于云端AI训练与推理、通用计算、AI for Science等计算任务。 国内端边大模型AI芯片企业后摩智能是第二次参加WAIC，首发并展示了自 ...

Core Viewpoint - The World Artificial Intelligence Conference (WAIC) highlights the importance of embodied intelligence and the underlying computing infrastructure, which includes chips, boards, servers, and computing clusters, as a critical topic for the AI industry [2][5]. Group 1: Computing Infrastructure Transformation - This year's WAIC showcased a shift from a focus on "parameter competition" to practical discussions on "fragmented computing resource coordination," "low power and low cost," and "vertical product integration" [5]. - The trend of "full-chain localization" in computing infrastructure is gaining attention, driven by global supply chain disruptions and the need for self-sufficiency in core technologies [7][8]. - Domestic computing infrastructure manufacturers are expanding localization efforts beyond individual chips to encompass architecture design, software and hardware ecosystems, and industry implementation [9]. Group 2: Notable Companies and Innovations - Muxi, a representative of this localization trend, showcased its latest GPU, the Xiyun C600, which features a self-developed architecture and is designed for cloud AI training and inference [10]. - The Xiyun C600 is equipped with HBM3e memory, significantly enhancing memory bandwidth for large model training and inference [12]. - The previous generation Xiyun C500 series has already achieved full-chain localization in its server solutions, including compilers and interconnect protocols [13]. Group 3: High-Performance Computing Alternatives - Zhonghao Xinying presented its "Shan" series TPU, which utilizes controllable IP cores and a self-developed instruction set, achieving a 30% reduction in energy consumption for the same AI computing tasks [15]. - The "Taize" computing cluster system, based on the "Shan" TPU, can support over 400P (TF32) of floating-point operations, suitable for large AI model computations [17]. Group 4: Industry Applications and Collaborations - Huawei's "384 Super Node" was unveiled, showcasing its capability to interconnect 384 cards for large model adaptation, with over 80 models already developed [20]. - The collaboration between Huawei and partners spans various industries, including finance, healthcare, and transportation, demonstrating the practical applications of their computing solutions [20]. - Moore Threads presented 12 industry-specific demos, including a video super-resolution technology that enhances low-resolution video quality, showcasing the versatility of their solutions [21][22].

Core Insights - The World Artificial Intelligence Conference (WAIC) in Shanghai highlighted the growing importance of embodied intelligence, showcasing robots capable of physical interaction and complex tasks [1] - The focus has shifted from performance parameters to practical applications and solutions in the computing infrastructure sector, indicating a more mature industry approach [3][4] Group 1: Computing Infrastructure Trends - The computing infrastructure exhibition at WAIC displayed a significant change, moving away from the "parameter competition" seen in previous years [3] - Technical specifications are now integrated into industry solutions rather than being prominently displayed, reflecting a more application-oriented mindset [4] Group 2: Domestic Production and Self-Sufficiency - The push for "full-link domestic production" in computing infrastructure is gaining momentum, driven by global supply chain disruptions and technology challenges [5] - Domestic manufacturers are expanding their focus from individual chips to a comprehensive self-sufficient ecosystem, including architecture design and software integration [5] Group 3: Notable Companies and Innovations - Muxi, a representative company, showcased its latest GPU, the Xiyun C600, which features a self-developed architecture aimed at AI training and inference tasks [6] - The Xiyun C600 is equipped with HBM3e memory, enhancing bandwidth for large model training, although it was not yet available in server form at the exhibition [10] - Zhonghao Xinying presented its "Shan" TPU series, which boasts a fully controllable IP core and a design that reduces energy consumption by 30% for AI tasks [12] Group 4: Industry Applications and Collaborations - Huawei's "384 Super Node" technology was unveiled, showcasing its capability to support over 80 large models across various sectors, including finance and healthcare [18][19] - The collaboration between Huawei and partners aims to create solutions across multiple industries, enhancing the adaptability of their computing infrastructure [19] - Moore Threads demonstrated its video super-resolution technology, which significantly improves video quality and can be integrated into various applications, stimulating industry collaboration [20][23]