英伟达在2025年里既有诸多高光时刻，也不乏剧烈震荡，而以下这15件事无疑是最重要的关键节点。 1月6日：CES 2025 英伟达在CES 2025上以一系列产品发布拉开了新一年的序幕，从全新超小型AI芯片到性能更强的游戏 显卡。黄仁勋的主题演讲涵盖了多个话题，包括实体AI(即机器人)、桌面级AI系统，以及一系列AI软件 更新，为公司2025年的发展奠定了基调。 1月27日："DeepSeek时刻" 尽管从几乎所有指标来看英伟达在2025年都取得了惊人的成绩，但这一年却以一场震撼开局。1月， DeepSeek发布了其R-1模型，并声称该模型是在未使用最先进处理器的情况下完成训练的。这一消息引 发了AI交易的剧烈动荡，英伟达市值在单日内蒸发近6000亿美元。 智通财经APP获悉，很难想象有哪家公司在2025年对华尔街和人工智能行业的影响能超过英伟达 (NVDA.US)。投资者密切关注着该公司首席执行官黄仁勋的一举一动，无论是他与美国总统特朗普的 会面，还是与三星和现代的高管晚上一起吃炸鸡喝啤酒。英伟达的营收在2025年飙升，达到1871亿美 元。其市值则继续攀升，一度短暂突破5万亿美元，随后回落并稳定在4万亿美 ...

郑纬民认为，算力自主有三个要求：一是芯片设计能力要自主；二是制造与供应链风险要可控；三是系 统与集群交付能力要强。算法自强则指的是，针对GPU能做大模型适配，能做大模型训练，能让模型优 化。 中新网12月22日电(记者 吴家驹)"真正决定主权AI成败的，在于是否有足够多的开发者，愿意长期在这 套栈上写代码。"近日，摩尔线程首届MUSA开发者大会上，中国工程院院士、清华大学计算机系教授 郑纬民如是表示。 郑纬民表示，过去很长一段时间，做芯片是全球化分工的，有人做架构设计，有人做设备制造，有人做 代工，有人做封装。如今，"主权AI"已成为每个国家必须回答的现实问题，是提升未来国家竞争力的关 键。 在他看来，"主权AI"其核心在于实现"算力自主、算法自强、生态自立"的完整体系。 他表示，在某种程度上，生态自立比算力自主和算法自强更重要。因为生态自立，意味着芯片要从"能 跑"到"愿意用"，而不是能跑一个软件就可以。"真正决定主权AI成败的，在于是否有足够多的开发者， 愿意长期在这套栈上写代码。" 同时，郑纬民表示，当前中国芯片行业面临着内卷与碎片化问题，例如：不同的厂家提供不同的接口， 要做不同的适配。开发者是生态建 ...

Core Viewpoint - The article discusses the challenges and limitations of current AI models, highlighting issues such as "sycophancy" where AI tends to reinforce users' existing beliefs rather than challenge them, leading to potential misinformation and "AI hallucinations" [3][6][12]. AI Model Limitations - A significant flaw in mainstream AI models is their tendency to produce "confident errors," where incorrect information is reinforced rather than corrected, as demonstrated in a case study involving a low-income single mother and vitamin C [6][12]. - The concept of "sycophancy" is introduced, indicating that AI models often cater to users' pre-existing views, which can lead to the propagation of false information [6][7]. Market Dynamics and AI Adoption - Currently, 95% of AI pilot projects in enterprises remain in the experimental phase due to a lack of effective testing mechanisms and clear definitions of what constitutes "good AI," hindering large-scale commercialization [4][12]. - The article notes that the push for "sovereign AI" is leading to the development of localized AI models, which may create a fragmented market rather than a monopolistic one [8][12]. Regulatory Environment - The article critiques the notion that regulation stifles innovation, arguing that clear guidelines are necessary for safe and effective AI development. Companies are calling for well-designed regulatory frameworks to mitigate risks associated with AI [10][11]. - The delay in the implementation of the EU's AI Act reflects the need for updated regulations that address the challenges posed by generative AI, which were not anticipated in earlier drafts [11][12]. Concerns About AI Bubble - There is a growing concern about an "AI bubble," fueled by excessive investment without clear returns, as many companies are hesitant to scale AI solutions due to uncertainties in performance when deployed in real-world scenarios [12][13]. - The article emphasizes that while there is significant potential in AI technology, the exact form this potential will take remains unclear, contributing to the ongoing debate about the sustainability of current investments in the sector [13].

在沙特近期举行的发展融资大会（Momentum）期间，乔杜里接受第一财经的专访时表示，当前的AI模 型显现出一种"谄媚倾向"，即为了留住用户，模型极少挑战用户的既有观点，反而倾向于顺从并强化其 立场。 谈及市场热议的"AI泡沫"，她表示，研究显示，目前企业界高达95%的AI试点项目因缺乏有效压力测试 机制与确定的投资回报，仍停留在实验室阶段，难以实现大规模的商业化推广。而关键瓶颈在于，市场 缺乏"什么是好的AI"的权威指引。 人工智能为求用户留存，正发展出"讨好式"思维模式。 在高效辅助人类生产力跃升的同时，人工智能（AI）模型也常产生看似逻辑自洽、实则背离事实的输 出，包括对用户的盲目顺从。 例如，当流行病学家模拟一位低收入单身母亲询问"我没有钱给小孩买药或去医院，多少剂量的维生素 C能否治疗新冠"时，AI模型不加甄别地认同了"维生素C能治疗新冠"这一假设。进行这一实验的，正是 美国拜登政府时期首任AI科学特使乔杜里博士（Dr. Rumman Chowdhury）的团队。 乔杜里曾为推特（即X）机器学习伦理、透明度与问责总监，现为AI审计与评估公司"人道智 能"（Humane Intelligence）联合 ...

Core Insights - The core focus of the article is the rapid expansion of the domestic GPU leader, Moore Threads, highlighted by the launch of their new GPU architecture "Huagang" at the first MUSA Developer Conference [1][2]. Group 1: Technological Advancements - The "Huagang" architecture represents a significant technological evolution, with a 50% increase in computing density and a 10-fold improvement in energy efficiency compared to the previous generation, set for mass production next year [2]. - The architecture supports full precision from FP4 to FP64 and integrates the first-generation AI generative rendering architecture (AGR) and second-generation ray tracing hardware acceleration engine [2]. - Two core chips based on the "Huagang" architecture were announced: "Huashan," designed for AI training and inference, and "Lushan," focused on high-performance graphics rendering, with AI computing performance improved by 64 times and geometric processing performance enhanced by 16 times [3]. Group 2: Infrastructure and Performance - The "Kua'e" supercomputing cluster was introduced, achieving a floating-point computing capability of 10 Exa-Flops, with a training efficiency of 60% for Dense models and 40% for MOE models [5]. - The MTT S5000 single card achieved a Prefill throughput of over 4000 tokens/s and a Decode throughput of over 1000 tokens/s, indicating substantial breakthroughs in handling large-scale parameter models [6]. Group 3: Software Ecosystem - The MUSA architecture underwent a full-stack software upgrade, with the core computing library muDNN achieving over 98% efficiency in GEMM/FlashAttention and 97% in communication efficiency [7]. - The company plans to open-source key components of its computing acceleration library, communication library, and system management framework to the developer community [7]. Group 4: Market Position and Strategy - Moore Threads officially entered the personal intelligent computing terminal hardware market with the launch of the MTT AIBOOK, priced at 9999 yuan, featuring the self-developed SoC chip "Changjiang" [9]. - The company aims to create a closed-loop for code debugging and application development by bringing its MUSA ecosystem from cloud to desktop [9]. - The stock price of Moore Threads has shown volatility, closing at 664.10 yuan per share on December 19, with a cumulative decline of 29.4% from its peak on December 11, yet maintaining a market capitalization of 312.146 billion yuan [10].

Core Insights - The focus on "Mole Thread," the first domestic GPU stock, is shifting from its high valuation to its technological advancements, product iterations, and operational performance following its debut on the Sci-Tech Innovation Board [1] Group 1: Technological Developments - Mole Thread held its first MUSA Developer Conference on December 20, showcasing its full-function GPU technology roadmap and announcing a series of technological and product advancements, including the new GPU architecture "Huagang" [1] - The new architecture boasts a 50% increase in density and a 10-fold improvement in efficiency, supporting intelligent computing clusters of over 100,000 cards [1] - Future products based on this architecture will include the high-performance AI training and inference chip "Huashan" and the graphics rendering-focused chip "Lushan" [1] - The company also introduced the AI computing power notebook "Changjiang," equipped with an intelligent SoC chip, serving as a core entry point for developers into the MUSA ecosystem [1] Group 2: Industry Context - The development of "sovereign AI" is deemed crucial for enhancing national competitiveness, focusing on achieving a complete system of "autonomous computing power, self-reliant algorithms, and independent ecosystems [2] - The performance gap between domestic graphics cards and foreign mainstream products is narrowing, although building ultra-large-scale intelligent computing systems remains a significant challenge [2] - The current Chinese GPU industry is in the early stages of constructing a core technology stack and a complete ecosystem, facing challenges such as high R&D difficulty and the construction of computing ecological barriers [2] Group 3: Market Performance - Mole Thread's stock has seen recent adjustments, with a 5.9% drop on December 19, closing at 664.10 yuan per share, marking a cumulative decline of 29.4% from its peak of 941.08 yuan on December 11 [2] - Despite the recent decline, the stock remains over 480% higher than its issue price, with a total market capitalization exceeding 300 billion yuan [2]

Core Insights - The core focus of the article is the rapid expansion of the domestic GPU leader, Moore Threads, highlighted by the launch of their new GPU architecture "Huagang" at the first MUSA Developer Conference [1][2]. Group 1: Product Development - Moore Threads introduced the "Huagang" architecture, which boasts a 50% increase in computing density and a 10-fold improvement in energy efficiency compared to the previous generation, set for mass production next year [1]. - The "Huagang" architecture supports full precision from FP4 to FP64 and integrates the first-generation AI generative rendering architecture (AGR) and second-generation ray tracing hardware acceleration engine [1]. - Two core chips based on the "Huagang" architecture were announced: "Huashan," designed for AI training and inference, and "Lushan," focused on high-performance graphics rendering, with AI computing performance improved by 64 times and geometric processing performance increased by 16 times [2]. Group 2: Infrastructure and Performance - The "Kua'e" supercomputing cluster was launched, achieving a floating-point computing capability of 10 Exa-Flops, with a training efficiency of 60% on Dense models and 40% on MOE models [4]. - The MTT S5000 single card achieved a Prefill throughput of over 4000 tokens/s and a Decode throughput of over 1000 tokens/s on the DeepSeek R1 671B model, indicating significant breakthroughs in system-level engineering optimization for large-scale parameter models [5]. Group 3: Software Ecosystem - The MUSA architecture received a full-stack software upgrade, with the core computing library muDNN achieving over 98% efficiency in GEMM/FlashAttention and 97% in communication efficiency [6]. - The company plans to open-source key components of its computing acceleration library, communication library, and system management framework to the developer community [6]. - A new intermediate language, MTX, compatible with cross-generation GPU instruction architectures, and a programming language, muLang, aimed at rendering and AI integration, will be introduced to lower adaptation barriers for developers [6]. Group 4: Market Position and Strategy - Moore Threads officially entered the personal intelligent computing terminal hardware market with the launch of the MTT AIBOOK, priced at 9999 yuan, expected to be available on January 10, 2026 [7][8]. - The MTT AIBOOK features the self-developed intelligent SoC chip "Changjiang," integrating a high-performance CPU and full-function GPU, with heterogeneous AI computing power reaching 50 TOPS [8]. - The company aims to transition from a single hardware supplier to a platform-level computing infrastructure provider, as reflected in the showcased "Huagang" architecture and the "chip-edge-end-cloud" full-stack system [9]. Group 5: Financial Performance - The company's stock price closed at 664.10 yuan per share on December 19, down 5.9%, with a cumulative decline of 29.4% from the peak on December 11, although it remains up over 481% from the issue price, maintaining a market capitalization of 312.146 billion yuan [9].

12月20日，摩尔线程首届MUSA开发者大会（MDC 2025）在北京中关村国际创新中心开幕。 在主论坛环节，中国工程院院士、清华大学计算机系教授郑纬民提出，在芯片产业全球化分工遭遇技术封锁的背景下，构建中国"主权AI"计算引擎成为紧迫 任务。要实现"主权AI"，需从算力自主、算法自强、生态自立三方面入手。 从"主权AI"基建的角度出发，发展国产万卡/十万卡系统是不得不走的一步，但仍需解决互联网络与拓扑、可靠性与运维、能耗与供电散热等方面的问题。 至于国产芯片厂商都要面对的终极问题——生态建设，在郑纬民看来，真正决定"主权AI"生态成败的，是有没有足够多的开发者愿意长期在这套栈上写代 码。未来国产平台要提高用户的开发体验，还需解决迁移成本高、工具链不成熟、文档/社区与支持不足等问题。 郑纬民教授现场演讲，图片来源：每经记者杨卉摄 "主权AI"三大支柱：算力自主、算法自强、生态自立 过去很长一段时间，芯片产业一直处于全球化分工的状态，架构设计、制造装备、代工、封测等环节均涉及不同领域。然而，近年来高端AI芯片面临出口 管制、技术封锁等困境，算力从一般生产要素上升为战略资源，"主权AI"也从学术讨论逐步变为每个国 ...

除主论坛外，大会还设置了20余场技术分论坛与超过1000平方米的"MUSA嘉年华"沉浸式展区，全面呈 现了国产GPU在AI大模型、科学智能、数字孪生、工业仿真、数字文娱、智慧医疗等前沿与产业场景 中的应用潜力。 海淀区委书记、中关村科学城党工委书记张革说，摩尔线程自2020年成立以来扎根海淀，专注于全功能 GPU自主研发，以"一年一芯片"的迭代速度推出四代产品，在AI智算、数字孪生等领域填补了国内技术 空白，是海淀区硬科技企业创新发展的典型代表。本次大会不仅是企业技术成果的集中展示，也是海淀 区人工智能与集成电路产业生态协同共进、向优发展的生动体现。 12月20日，首届摩尔线程MUSA开发者大会在中关村国际创新中心举办。大会上，不久前刚成功登陆资 本市场的海淀全功能国产GPU企业摩尔线程集中发布了一系列技术与产品新成果，如计算提升50%、效 能提升10倍的全功能GPU架构"花港"，在多项关键精度指标上达到国际主流水平的夸娥万卡智算集群 等，推动国产GPU技术与生态实现进一步突破。 中国工程院院士、清华大学计算机系教授郑纬民发表主题演讲时指出，发展"主权AI"是提升未来国家竞 争力的关键，其核心在于实现"算力自 ...

Core Insights - The first MUSA Developer Conference (MDC 2025) was held by Moore Threads in Beijing, where the company unveiled its new GPU architecture "Huagang" and a series of technological advancements [2] - Moore Threads has established a complete technology stack based on its self-developed unified architecture, covering "chip-edge-end-cloud" integration, and plans to increase R&D investment [2] Group 1: New Architecture and Chip Roadmap - The MUSA (Meta-computing Unified System Architecture) has been upgraded to version 5.0, achieving key breakthroughs in full-stack unification, performance, and ecological openness [3] - The "Huagang" architecture supports full precision calculations from FP4 to FP64, with a 50% increase in computing density and a 10-fold improvement in energy efficiency, capable of supporting over 100,000 card-scale intelligent computing clusters [3] - Two upcoming chip technologies based on the "Huagang" architecture were announced: "Huashan," focusing on AI training and ultra-large-scale intelligent computing, and "Lushan," specializing in high-performance graphics rendering [3][5] Group 2: AI Training and Computing Clusters - The newly launched "Kua'e" intelligent computing cluster achieves full precision and general computing capabilities, with a floating-point computing capacity of 10 Exa-Flops and training efficiency rates of 60% for Dense models and 40% for MOE models [7] - The MTT S5000 single card has achieved breakthroughs in inference performance, with a throughput of over 4000 tokens/s for Prefill and 1000 tokens/s for Decode [7] - Future architecture planning for the MTT C256 super node aims to enhance training efficiency and inference capabilities for large-scale intelligent computing centers [7] Group 3: Graphics Computing and AI Technologies - Moore Threads' products support major graphics and computing APIs, including DirectX 12 and Vulkan 1.3, and have achieved compatibility with mainstream domestic CPUs and operating systems [8] - Key breakthroughs in rendering technology include hardware-level ray tracing acceleration and self-developed AI generative rendering technology, enabling realistic lighting effects on domestic GPUs [8] - The MT Lambda embodiment intelligence simulation training platform integrates physics, rendering, and AI engines for efficient development and training environments [8] Group 4: Ecosystem Development and Education - The concept of "ecosystem" was emphasized, with the company focusing on building a self-reliant domestic computing industry ecosystem through collaboration and innovation [11] - The company has established a developer growth system through the Moore Academy, gathering nearly 200,000 developers and learners, and engaging over 100,000 students in over 200 universities [11] - The company plans to open-source key simulation acceleration components to enhance research and development efficiency in the robotics industry [9]