过去一年，具身智能持续处于资本与产业关注的交汇点。一方面，融资规模不断扩大，技术演示频繁刷 新行业想象；另一方面，稳定落地、规模复制以及成本控制仍是行业绕不开的现实挑战。 2月10日晚间，星海图合伙人、CFO（首席财务官）罗天奇在接受包括《每日经济新闻》在内的媒体记 者采访时表示，具身智能最终依然是一个由Scaling Law（规模定律）驱动的AI（人工智能）行业，胜 负手不在于短期的融资额，而在于每一元钱能换回多少智能。 2月11日，星海图完成10亿元B轮融资。截至本轮，星海图累计融资额近30亿元，估值百亿元，成为继 宇树、智元、银河通用之后具身智能行业又一只百亿元"独角兽"。 行业正经历结构性转折 当前具身智能行业正经历一次重要的结构性转折，资本逻辑正在从"广撒网"转向"押头部"，行业也从早 期技术探索阶段逐步迈向资源密集型竞争阶段。 硬件成本不是比拼关键 关于具身智能的"ChatGPT时刻"何时到来，业内争议颇多。罗天奇认为，"ChatGPT时刻"不一定是一两 年内很快到来的，但这并不妨碍商业化的率先开启。 罗天奇将具身智能的商业化拆分为技术驱动与商务驱动，后者包括机器人表演等场景。 罗天奇将当下具身智 ...

Core Insights - The article discusses the imminent release of new AI models in the Chinese market, particularly focusing on the GLM-5 model from Zhipu AI, which is expected to leverage advanced technologies and compete effectively in the AI landscape [1][16]. Group 1: Model Development and Features - The GLM-5 model has been linked to multiple technical platforms, indicating a strong collaborative effort in its development [2][4]. - GLM-5 incorporates a 78-layer Transformer decoder with a total parameter count of approximately 745 billion, which includes a mixture of dense and sparse architectures [6][8]. - The model utilizes a hybrid expert (MoE) architecture, activating only a small fraction of its parameters during inference, which enhances efficiency while maintaining performance [9][10]. Group 2: Technological Innovations - The integration of DeepSeek's Sparse Attention (DSA) mechanism allows GLM-5 to handle long sequences more efficiently, reducing computational costs significantly [12][13]. - Multi-Token Prediction (MTP) technology is employed to accelerate token generation, allowing the model to predict multiple tokens simultaneously, which is particularly beneficial for structured text generation tasks [15][16]. - The model's architecture reflects a shift towards efficiency over sheer parameter count, indicating a trend in the AI industry towards optimizing performance rather than simply increasing size [9][17]. Group 3: Market Position and Challenges - GLM-5 is expected to excel in code generation and logical reasoning tasks, positioning it competitively in software development and algorithm design [16]. - However, the model currently lacks multi-modal capabilities, which may limit its applicability in creative AI-generated content (AIGC) scenarios, especially as competitors advance in this area [16]. - The article highlights a broader industry trend where companies are moving towards open-source technology integration, emphasizing efficiency and practicality in AI model development [16][17].

一凡 发自 凹非寺 量子位 | 公众号 QbitAI 准备回家过年了吗？ 有没有感觉今年回家比去年还堵？据说今年春运流量再创新高，官方预计 40天内人员流动量将达95亿人次 ，其中多数人仍然选择自驾出行， 占比达到了8成， 人次超过70亿 。 如果你也是自驾回家的一员，可能会发现今年春运还真有点不一样，因为AI含量更高了，现在， AI不仅在加持你的出行，甚至在关键时刻真 的能救命 。 有的AI在算命，有的AI在救命 今年是AI味儿最浓的一年，我们习惯了用AI学知识、写代码、辅导孩子作业，甚至算命……但AI啥时候能救命了？ 有人第一次产生"AI能救命"的认知，是因为一起事故。 2024年5月1日，广东梅大高速茶阳路段发生塌方。事故发生时，有车主提前刹车，逃 出生天。 据其事后讲述，他会提前刹车，正是因为在到达事发路段前，车上的高德导航突然发出语音警告 "前方有车辆急刹" ，所以他才赶紧刹车，躲 过一劫。 类似的用户当时还有很多，据了解，梅大高速茶阳路段发生塌方灾害时，高德首次急刹预警发出的时间是5月1日凌晨2：00左右，提醒后方车 辆注意前车急刹，安全驾驶。"当晚半小时内，我们共检测到多个速度骤降的急刹，并通过 ...

Core Insights - Generalist is a leading company in the robotics field with significant long-term competitive potential, focusing on data scale, team capability, and a clear scaling path [2][4]. Data Collection and Quality - High-quality real-world data is recognized as a core scarce resource in the robotics industry, with Generalist claiming to have accumulated 270,000 hours of training data, positioning it as the first robotics team to reach a data scale comparable to GPT-1 [4][6]. - The current mainstream methods for data collection include real machine data, human-operated data, pure video data, and synthetic data, with a consensus that real machine data is essential for training usable robotic models [5][6]. - Generalist's data collection strategy involves deploying thousands of data collection devices globally, utilizing egocentric data, and collaborating with data foundries to ensure diverse data sources [40][44]. Team and Technical Expertise - The core team of Generalist consists of members from prestigious institutions like OpenAI, Boston Dynamics, and Google DeepMind, contributing to significant projects such as PaLM-E and RT-2, showcasing strong technical capabilities [2][53]. - The team has demonstrated a clear research path and model dexterity through various demos, indicating a focus on achieving high levels of agility in robotic tasks [3][30]. Model Development and Performance - Generalist's GEN-0 model exhibits remarkable dexterity and the ability to perform complex tasks autonomously, showcasing its potential in physical interaction challenges [30][37]. - The model architecture employs Harmonic Reasoning, integrating perception and action tokens in a single Transformer flow, allowing for continuous and intelligent action generation [52]. Competitive Landscape - Generalist operates in a competitive environment with other companies like Physical Intelligence and Google, each with distinct strategies and strengths. Generalist's primary advantages lie in its extensive real machine data and strong team expertise, while facing challenges from competitors with more comprehensive team structures and funding [62][63]. - The company is positioned in the second quadrant of the robotics industry landscape, focusing on developing a general robotic brain, while competitors like Sunday are advancing faster in practical applications [61][62].

Core Insights - OpenAI CEO Sam Altman announced an online seminar to gather public feedback before developing the next generation of AI tools [1] - The seminar focused on the future of AI, model evolution, and the transition from static software to real-time generated applications [3][5] - Altman highlighted the asymmetry in GPT-5's performance, acknowledging a trade-off in writing ability for enhanced reasoning and programming skills [4][7] Group 1: AI Model Development - Altman revealed that by the end of 2027, the intelligence cost of GPT-5.2 is expected to decrease by at least 100 times [4][8] - The emphasis is shifting from cost to speed, with developers prioritizing rapid output for complex tasks [5][8] - OpenAI aims to provide two pathways: extremely low-cost intelligence and high-speed feedback systems [5] Group 2: Software Evolution - The future of software is envisioned as dynamic and personalized, moving away from static applications [5][9] - Users will expect computers to generate immediate solutions tailored to specific problems, leading to a restructured operating system [5][9] - This shift will create a unique, evolving productivity system for each user, fundamentally changing software interaction [5] Group 3: Economic and Social Impact - Altman believes AI will empower individuals, reduce barriers, and enable low-cost innovation and entrepreneurship [6] - However, there are concerns about potential wealth concentration and the need for policy focus to prevent this [6] - The integration of AI is expected to enhance human collaboration rather than diminish it, fostering new forms of teamwork [20] Group 4: Challenges and Future Directions - Altman acknowledged the challenges in achieving long-term autonomous operation for AI agents, emphasizing the need for task simplification [11] - The importance of resilience in AI safety, particularly in biological and cybersecurity contexts, was highlighted [16][17] - OpenAI is exploring how to effectively integrate AI into educational settings, particularly for younger children [24]

外面的VC仍旧挥舞着支票，只是这次不再为"流量买单"，而是要为"AI革命"买单。VC们希望买断的， 是他们脑子里在大厂内部无法孵化的"非共识"。 2021年，原商汤科技副总裁闫俊杰创立了MiniMax；2023年，曾于Google和Meta 参与核心模型研发的 杨植麟创立了月之暗面；原微软亚洲研究院负责人姜大昕博士创立了阶跃星辰；原百度小度科技CEO 景鲲创立了 Genspark。这些名字背后，是一群对 Transformer 架构有深度理解、对 Scaling Law 有信 仰、用数亿算力"喂"出工程直觉的人。 整个2025年，创始人背景带有科技大厂（包括但不限于百度、阿里、腾讯、字节、美团、滴滴、华为、 网易、大疆等企业）的创业融资事件数量超70起。其中，大疆、字节、腾讯、华为、阿里5家的高管创 业事件数量遥遥领先。相应地，智能硬件、AI应用、具身智能成为最集中的创业赛道。 十年，足够让一代技术从实验室走向千家万户，也足够让一种权力结构从巅峰滑向黄昏。 曾几何时，大厂是抵御风险的避风港。但在AI彻底重构生产力的今天，那个严丝合缝的层级组织，正 在逐渐异化为一个巨大的认知囚笼。当"确定性管理"遇上"探索性创 ...

Group 1 - The core discussion revolves around the technological landscape of AI and autonomous driving, focusing on the anticipated developments in 2026 and the implications for investment opportunities [1][2][3] - The transition from theoretical discussions about AI, such as Scaling Law, to practical applications is highlighted, with industry leaders emphasizing the need for localized and practical AI solutions [2][5] - The concept of "DeepSeek Moment" signifies a shift away from the dominance of major tech companies in AI model development, suggesting that innovation may increasingly occur outside these established firms [3][4] Group 2 - The debate on whether Meta should focus on model development or application capabilities reflects broader strategic challenges faced by tech giants in the evolving AI landscape [6][7][8] - The performance of Google's Gemini and its integration with TPU showcases the importance of efficient computing solutions in the AI sector, indicating a potential shift in market dynamics [29][30] - The discussion on the operational costs of autonomous driving technologies, particularly comparing Tesla and Waymo, underscores the significance of long-term operational efficiency and maintenance in evaluating investment potential [24][25][26] Group 3 - The potential for AI applications to emerge as "killer apps" in 2026 is debated, with emphasis on the need for applications that integrate seamlessly into workflows rather than merely enhancing existing functionalities [10][11] - The financial landscape for AI investments is characterized by a belief in the ongoing growth of AI capabilities, with concerns about potential market corrections if expectations are not met [32][34] - The macroeconomic risks, including geopolitical factors and monetary policy changes, are identified as critical elements that could impact the tech sector's performance in 2026 [34][35]

Core Insights - The article discusses the recent IPOs of AI companies, highlighting the significant oversubscription rates and initial stock price surges, indicating strong market interest in AI ventures [3][5] - It emphasizes the challenges faced by AI startups in a landscape dominated by major tech firms like Tencent, ByteDance, and Alibaba, suggesting that these giants create a difficult environment for smaller companies to thrive [7][15] Group 1: Market Dynamics - The IPO of Zhihua Huazhang on January 8, 2026, had an issue price of HKD 116.2 per share, with a subscription rate of approximately 1,159 times, and a first-day price increase of 13.17%, leading to a market cap of nearly HKD 90 billion [3] - MiniMax, established only four years prior, went public on January 9, 2026, at HKD 165 per share, with an oversubscription of over 1,800 times and a first-day price increase of 109.1%, resulting in a market cap exceeding HKD 100 billion shortly thereafter [5] Group 2: Technological Paradigms - The article argues that the current AI landscape is shaped by the "Scaling Law," which suggests that increasing model size, data, and computational power leads to predictable improvements in performance [9][10] - It notes that the success of OpenAI is seen as a unique historical occurrence that may not be replicable, as the current environment is characterized by concentrated computational resources and homogenized model capabilities [12][13] Group 3: Competitive Landscape - The emergence of DeepSeek has altered industry perceptions by significantly reducing training and inference costs, challenging the narrative that only large investments can yield viable models [19][22] - Major companies are now treating models as foundational infrastructure rather than profit centers, which complicates the ability of startups to justify their value propositions to clients [22][23] Group 4: Strategies for Startups - Startups like MiniMax and Zhihua Huazhang are finding sustainable paths by avoiding direct competition with large firms, focusing instead on niche markets or specific applications [26][30] - MiniMax is targeting overseas markets with products centered on companionship and interaction, while Zhihua focuses on complex enterprise applications that larger firms may overlook [28][31] - The article suggests that successful startups must carve out unique positions within existing paradigms rather than attempting to replicate the success of giants like OpenAI [42]

Core Viewpoint - The company has actively revised its logic regarding "DAU/traffic = barrier," recognizing that the only moat in the AI era is the intellectual advantage of models. By reducing inefficient ToB sales teams and C-end user acquisition costs, resources are intensely focused on high-intensity model research and technological breakthroughs. This "anti-consensus" contraction is essentially a strategic move to gain an edge in the second half of the Scaling Law, focusing on reasoning and architectural innovation. The management team possesses top-tier research and ToB commercialization and delivery experience [1]. Industry Transformation - AI is defining a new generation of productivity, with the Total Addressable Market (TAM) shifting from software budgets to labor budgets. The industry is undergoing a qualitative change from discriminative AI to generative AI. The Scaling Law is driving exponential improvements in model intelligence while reasoning costs are decreasing exponentially, making AI not just a traditional SaaS "tool" but a "digital employee" with reasoning and planning capabilities. For instance, the traditional software market targets about $300 billion of IT budgets, while AI, as a production factor, is expected to penetrate a global labor cost market of approximately $13 trillion [1]. Technical Evolution - The technical path is evolving, with increasing engineering barriers and a gradual consolidation of model frameworks. The extensive pre-training Scaling (computational power/data) is facing diminishing marginal returns, leading the industry into a new phase of "architectural innovation & reasoning-side Scaling." Companies like DeepSeek (MLA/architecture compression) and Google (multimodal association) represent different technical breakthrough directions, with technical barriers returning to a combination of "engineering capability + architectural innovation" [2]. Company Advantages - The MiniMax team benefits from a founder with both research capabilities and ToB delivery experience. The founder, Yan Junjie, previously served as Vice President of SenseTime and CTO of the Smart City Business Group, demonstrating exceptional technical and management skills. Under his leadership, a team of over 700 achieved industry-leading facial recognition algorithms, with the smart city business generating over 2 billion RMB in revenue in 2021, targeting government and large enterprises with a focus on engineering delivery. Unlike the mobile internet era, the management team understands that "user scale ≠ model intelligence." Therefore, the company's strategic focus is shifting from "revenue generation/user acquisition" to "technological iteration" by 2025 [3]. Financial Projections - The company expects to achieve revenues of $80 million, $185 million, and $351 million for FY25-27, representing year-on-year growth of 162%, 131%, and 90%. AI-native product revenues (To C) are projected to be $58 million, $139 million, and $263 million, with growth exceeding 140%. Open platform revenues (ToB) are expected to be $22 million, $46 million, and $88 million. As reasoning costs optimize and high-margin ToB business stabilizes, Non-GAAP gross profits are projected to be $20 million, $74 million, and $193 million, corresponding to Non-GAAP gross margins of 25.0%, 40.0%, and 55.0%. Despite ongoing competition in computational power, Non-GAAP net losses are expected to narrow, recording losses of -$240 million, -$180 million, and -$80 million for FY25-27, with a significant trend of decreasing loss rates [4]. Investment Logic - Within the AI sector, the investment logic between the "infrastructure layer (e.g., DeepSeek)" and the "native application layer" is diverging. Compared to DeepSeek, which establishes barriers on the cost side through architectural innovation, MiniMax's deep focus on multimodal (voice/video) interaction experience provides a stronger moat in user stickiness and commercialization. The integration of "high-sensory interaction" and "productivity tools" is key, with the former (Talkie/Xingye) providing vast RLHF data and the latter (Hailuo/open platform) generating high-margin cash flow. Greater revaluation potential lies in the technological unlocking in the second half of the Scaling Law. As the company's strategic focus returns to technological research, new multimodal models (e.g., Video-01, end-to-end voice) are expected to significantly contribute to incremental ARR starting in FY26-27 [5].

Investment Rating - The report maintains a "Buy" rating for key companies in the domestic computing power industry, including Cambrian (688256) and SMIC (0981.HK) [7][12]. Core Insights - The report emphasizes the certainty of investment opportunities in the domestic computing power sector, driven by the upcoming release of the new generation of computing power chips, represented by H Company’s 950 series, which will enter mass production in the first quarter [5][6]. - The domestic computing power industry is expected to see significant growth, with ByteDance projected to invest 150 billion in global computing power procurement in 2026, of which 60-65 billion is expected to be allocated domestically, with over 40 billion for domestic computing power [6]. - The report suggests that the impact of the H200 release on the domestic computing power industry will be limited, as the primary application scenarios differ from those of domestic solutions [6]. Summary by Sections Supply Side - The domestic computing power sector has faced challenges due to U.S. restrictions on advanced semiconductor processes and key materials. However, breakthroughs are anticipated starting in the second half of 2025, with improved collaboration between chip design companies and foundries expected to enhance production yields by 2026 [5][6]. - The report highlights the optimization of the supply chain and the collaboration between hardware and software, which has significantly improved the usability of domestic computing power in AI inference scenarios [5]. Demand Side - The demand for computing power in 2026 is becoming clearer, with major internet companies like Alibaba and Tencent also planning significant investments in domestic computing power [6]. - The report notes that the three major telecom operators are expected to increase their procurement of domestic computing power to meet the growing demand from AI applications [6]. Key Companies to Watch - The report recommends focusing on Cambrian (688256) as a representative of domestic computing power card suppliers and SMIC (0981.HK) as a leading foundry. Additionally, attention is drawn to Huahong Semiconductor (1347.HK) for its advancements in advanced processes [7]. - The report also suggests monitoring companies related to domestic IC substrates due to supply bottlenecks caused by shortages of upstream materials, recommending companies like Shenzhen South Circuit (002916) and Pengding Holdings (002938) [7]. Overseas Computing Power Industry - The report observes a shift in the driving force of AI computing power from training large models to deploying inference applications, with companies like Google leading advancements in model capabilities [8][9]. - The report anticipates continued high growth in AI application-driven computing power demand, with major companies expected to double their computing power every six months over the next few years [10].