高盛认为，AI视频生成领域非"赢家通吃"，快手可灵与字节即梦均将受益于市场扩张。预计全球AI视频生成市场将从 2025年30亿美元增至2030年290亿美元，增长10倍。AI技术将重塑娱乐业价值链，使其向上游IP设计和分发平台转移， 这些环节将获得更高附加值。 硬·AI 作者 | 董 静 编辑 | 硬 AI 尽管字节跳动旗下即梦近期推出的Seedance 2.0引发了市场高度关注，但高盛认为，AI视频生成领域并 非"赢家通吃"的零和游戏，快手可灵与字节跳动即梦2.0等头部模型都将从快速扩张的市场中受益，且AI技 术的进步将重塑整个娱乐行业的价值分布。 在字节即梦2.0于2月12日正式向公众开放后，投资者对两大平台的竞争格局高度关注。高盛分析师Lincoln Kong、Ronald Keung及Luqing Zhou在最新研报中表示，可灵3.0在2月5日升级，较即梦2.0提前数日推 出，两者均在音视频一致性、视频时长（15秒）及叙事控制方面实现重大突破。 尽管部分测试者认为即梦2.0在流畅度和多场景连贯性上表现更优，但高盛强调，可灵3.0在影视级细节和 定价优势上仍保持竞争力，且在第三方基准测试中持续位居全球顶 ...

作者 | 杨 宸 编辑 | 硬 AI 美国市场对AI可能带来行业"被颠覆"的担忧 ， 正在把全球资金从华尔街的潜在受害者，推向亚洲的AI基础 设施赢家，芯片制造股成为最直接的受益者。 据彭博， MSCI亚太指数2026年以来上涨逾12%，而标普500年内下跌0.2%，纳斯达克100下跌约2%。 过去10个交易日，纳斯达克100下挫4.6%，市值蒸发约1.5万亿美元，软件股等被认为易受新AI工具冲击 的板块领跌。 资金流向也在快速反映这一分化。周四，三星出现"最大规模"的海外净买入，股价上涨6.4%，周五继续上 行；全球投资者对台湾股票录得第三大单周净买入。日本存储芯片厂商铠侠控股周五股价大涨15%。 AI对各行业的颠覆恐慌引发全球资金大迁徙！华尔街资金正加速撤离美股相关标的，涌向亚洲AI硬件生产商。MSCI亚太 指数年内大涨逾12%超越标普500（-0.2%）与纳斯达克100（-2%）。三星、台积电等拥有定价权的芯片巨头遭外资爆 买，亚洲AI硬件供应链正成为这轮AI浪潮的大赢家。 硬·AI 02 外资加码韩国、台湾与日本，指数权重放大芯片股的带动效应 在增量资金进入的背景下，亚洲主要芯片股在本地指数中的高权重 ...

Core Viewpoint - MiniMax has launched its latest M2.5 model series, achieving a significant breakthrough in both performance and cost, aiming to address the economic feasibility of complex agent applications while claiming to have reached or refreshed the industry SOTA (state-of-the-art) levels in programming, tool invocation, and office scenarios [3][4]. Cost Efficiency - The M2.5 model demonstrates a substantial price advantage, costing only 1/10 to 1/20 of mainstream models like Claude Opus, Gemini 3 Pro, and GPT-5 when outputting 50 tokens per second [3][4]. - In a high-speed environment of 100 tokens per second, the cost for continuous operation for one hour is just $1, and it can drop to $0.3 at 50 tokens per second, allowing a budget of $10,000 to support four agents working continuously for a year [3][4]. Performance Metrics - M2.5 has shown strong performance in core programming tests, winning first place in the Multi-SWE-Bench multi-language task, with overall performance comparable to the Claude Opus series [4]. - The model has improved task completion speed by 37% compared to the previous generation M2.1, with an end-to-end runtime reduced to 22.8 minutes, matching Claude Opus 4.6 [4]. Internal Validation - Internally, MiniMax has validated the M2.5 model's capabilities, with 30% of overall tasks autonomously completed by M2.5, covering core functions such as R&D, product, and sales [4]. - In programming scenarios, M2.5-generated code accounts for 80% of newly submitted code, indicating high penetration and usability in real production environments [4]. Task Efficiency - M2.5 aims to eliminate cost constraints for running complex agents by optimizing inference speed and token efficiency, achieving a processing speed of 100 TPS (transactions per second), approximately double that of current mainstream models [7]. - The model has reduced the total token consumption per task to an average of 3.52 million tokens in SWE-Bench Verified evaluations, down from 3.72 million in M2.1, allowing for nearly unlimited agent construction and operation economically [9]. Programming Capability - M2.5 emphasizes not only code generation but also system design capabilities, evolving a native specification behavior that allows it to decompose functions, structures, and UI designs from an architect's perspective before coding [11]. - The model has been trained in over 10 programming languages, including GO, C++, Rust, and Python, across tens of thousands of real environments [12]. Testing and Validation - M2.5 has been tested on programming scaffolds like Droid and OpenCode, achieving pass rates of 79.7% and 76.1%, respectively, outperforming previous models and Claude Opus 4.6 [14]. Advanced Task Handling - In search and tool invocation, M2.5 exhibits higher decision maturity, seeking more streamlined solutions rather than merely achieving correctness, saving approximately 20% in rounds consumed compared to previous generations [16]. - For office scenarios, M2.5 integrates industry-specific knowledge through collaboration with professionals in finance and law, achieving an average win rate of 59.0% in comparisons with mainstream models, capable of producing industry-standard reports, presentations, and complex financial models [18]. Technical Foundation - The performance enhancement of M2.5 is driven by large-scale reinforcement learning (RL) through a native Agent RL framework named Forge, which decouples the underlying training engine from the agent, supporting integration with any scaffold [23]. - The engineering team has optimized asynchronous scheduling and tree-structured sample merging strategies, achieving approximately 40 times training acceleration, validating a near-linear improvement in model capabilities with increased computational power and task numbers [23]. Deployment - M2.5 is fully deployed in MiniMax Agent, API, and Coding Plan, with model weights to be open-sourced on HuggingFace, supporting local deployment [25].

Core Viewpoint - ByteDance's video model Seedance 2.0 has gained significant popularity overseas, with Elon Musk commenting on its rapid development, indicating a growing market interest in video generation capabilities [2][3][10]. Group 1: Product Launch and Features - Seedance 2.0 has been officially released and is fully integrated with Doubao and Jimeng products, along with the launch of the Huoshan Ark experience center for user trials [7][12]. - The model emphasizes capabilities such as original audio-visual synchronization, multi-camera long narrative, and multi-modal controllable generation, targeting a broader range of creators and commercial content scenarios [7][15]. - Key features include: 1. Multi-modal input supporting text, images, audio, and video, allowing for mixed input of composition, actions, camera movements, effects, and sounds [16]. 2. Original audio-visual synchronization with multi-track output, supporting background music, sound effects, or character narration, aligned with visual rhythm [17]. 3. Multi-camera long narrative capabilities that automatically parse narrative logic, generating shot sequences while maintaining character, lighting, style, and atmosphere consistency [17]. 4. Enhanced video editing and extension capabilities, reinforcing "director-level control" workflow attributes [18]. Group 2: Limitations and Future Developments - Despite its leading industry performance, ByteDance acknowledges that Seedance 2.0 is "far from perfect," with areas for improvement including detail stability, multi-character matching, multi-subject consistency, text restoration accuracy, and complex editing effects [20]. - Compliance and usage boundaries have become clearer, with restrictions on using real human images or videos as reference subjects unless verified or authorized, impacting certain commercial material production and deployment [23]. - The upcoming release of Doubao model upgrades on February 14, 2026, will include significant enhancements to the foundational model capabilities and enterprise-level agent capabilities [25].

Core Viewpoint - The article discusses the evolution of AI video generation from entertainment to industrial production, emphasizing the importance of controllability and reduced waste rates in video production processes [1][2][5]. Group 1: AI Model Updates - The recent updates in domestic multimodal models, such as Keling 3.0 and Seedance 2.0, significantly enhance controllability, marking a shift in AI video from entertainment to industrial production [2][5]. - The upgrades focus on improving controllability over video generation, including consistency across scenes, adherence to complex instructions, and the ability to edit generated content, thereby reducing waste rates [5][10]. Group 2: Cost Structure and Commercialization - The report indicates that the marginal cost of video production is increasingly aligning with computational costs, driven by advancements in AI models that lower waste rates [2][20]. - The commercialization threshold has shifted from "can it be done" to "can it be delivered consistently," highlighting the importance of stability in output quality [8][20]. Group 3: Competitive Landscape - ByteDance focuses on efficiency and cost reduction in content production, Kuaishou emphasizes professional storytelling, and Alibaba targets vertical e-commerce applications, indicating distinct strategic paths among these companies [19]. - The competition is not solely based on performance rankings but rather on the strategic differentiation of each company's offerings [19]. Group 4: Supply-Side Revolution - The report predicts a supply-side revolution where the marginal cost of content production will increasingly resemble computational costs, leading to improved efficiency for marketing and e-commerce service providers [20]. - As content production becomes easier, the scarcity and pricing of intellectual property (IP) will become more pronounced, with top-tier IP and derivatives gaining higher value [20].

Core Viewpoint - The article discusses the competitive landscape of AI model releases during the Chinese New Year period, highlighting the shift from model performance to efficiency and practical applications, with DeepSeek's V4 model being a focal point for potential industry transformation [2][4][15]. Group 1: Industry Dynamics - The 2026 Chinese New Year is expected to see a surge in flagship model releases, with multiple companies, including ByteDance, Alibaba, and DeepSeek, preparing significant updates [4][5][8]. - ByteDance has already launched three models, signaling a strong market entry with its Seedance 2.0 model [4]. - Alibaba plans to release Qwen 3.5 in mid-February, supported by a substantial customer acquisition incentive [5]. - DeepSeek's V4 model is anticipated to enhance coding and long prompt processing capabilities, with a reported support for up to 1 million tokens [8][15]. Group 2: Competitive Implications - The simultaneous release of multiple models may lead to a "winner-takes-all" scenario, where underperforming models face significant disadvantages [9][10]. - The Chinese New Year period is characterized by scarce attention, making it crucial for companies to present credible flagship updates to remain relevant [12][13]. Group 3: DeepSeek's Strategic Focus - DeepSeek's potential release aims to improve efficiency through a novel "conditional memory" approach, which could shift expensive computations to more cost-effective retrieval operations [15][16]. - If successful, this could enable AI to transition from being an expensive "toy" to a cost-effective "tool," facilitating its integration into high-frequency consumer products [17]. Group 4: Beneficiaries of the Model War - Tencent is positioned as a major beneficiary of the upcoming model releases, leveraging its high-frequency communication platforms like WeChat and QQ to enhance user experience with improved AI capabilities [19][20]. - In contrast, Alibaba and Baidu may face a dual challenge: while stronger models can enhance user experience, a price war initiated by DeepSeek could pressure the entire API service market [21][22]. Group 5: Market Sentiment and Future Outlook - Despite the excitement in the capital markets, there is skepticism regarding the actual performance of consumer-facing AI models, with large-scale user testing during the Chinese New Year serving as a critical evaluation point [24][25]. - The true signal for adoption will be whether major players integrate AI as a default feature in high-frequency interfaces, which would drive sustained demand for AI capabilities [25]. Group 6: Valuation Perspectives - Morgan Stanley maintains a bullish outlook on model developers like Zhiyu and MiniMax, projecting significant long-term growth based on their advancements and market positioning [27][28]. - The valuation logic is shifting towards long-term profitability, with target prices set at 400 HKD for Zhiyu and 700 HKD for MiniMax, based on expected earnings by 2030 [29][30].

Core Viewpoint - The article discusses the global debut of Z-Angle Memory (ZAM), a next-generation AI memory technology developed by Intel and SoftBank, which promises lower power consumption, higher capacity, and wider bandwidth compared to traditional DRAM [3][6]. Group 1: ZAM Technology Overview - ZAM utilizes a vertical stacking architecture that addresses long-standing thermal management issues associated with planar stacking designs [3][7]. - Early data indicates that ZAM can reduce power consumption by 40% to 50% and achieve a single-chip capacity of up to 512GB [3][7]. - The prototype of ZAM is expected to be launched in 2027, with full commercialization anticipated by 2030 [3][6]. Group 2: Market Positioning and Strategic Goals - ZAM aims to fill the gap between high bandwidth memory (HBM) and traditional DDR DRAM, providing significantly higher energy efficiency without sacrificing capacity [10][11]. - The technology is designed to overcome the limitations of current high bandwidth memory, which often compromises capacity for higher bandwidth [10][11]. - The collaboration between SoftBank and Intel seeks to address the energy consumption bottlenecks and supply chain challenges faced by current AI applications [11].

Core Viewpoint - Investors are shifting from seeking AI winners to rapidly selling stocks of companies that may be disrupted by AI, leading to a panic selling mentality across various sectors, including software, financial services, wealth management, insurance brokerage, and legal services [2][3]. Group 1: Market Reaction to AI Disruption - The latest wave of selling was triggered by the launch of a tax strategy tool, Hazel, by Altruist Corp., which caused significant stock price drops of over 7% for wealth management firms like Charles Schwab, Raymond James Financial Inc., and LPL Financial Holdings Inc., marking the largest decline since the market crash in April [3][5]. - The panic began when Anthropic introduced a new tool that led to a deep correction in software, financial services, asset management, and legal services sectors, indicating a turning point in market sentiment [6][8]. - The insurance brokerage sector was also heavily impacted after Insurify launched a new application using ChatGPT to compare auto insurance rates, resulting in substantial stock losses for U.S. insurance brokers [6][8]. Group 2: Concerns Over AI's Impact - The introduction of AI tools like Hazel highlights deep-seated anxieties about AI disrupting traditional financial services, as these tools can perform tasks that typically require entire teams, with costs as low as $100 per month [5][6]. - Market participants are increasingly concerned that any intermediary services that could be replaced by AI face existential threats, leading to widespread selling [6][8]. Group 3: Diverging Market Opinions - Despite the prevailing panic, some market analysts express skepticism about the speed and extent of AI disruption, suggesting that technological upheaval often takes longer to materialize than anticipated [8]. - Historical context indicates that industries like banking have faced challenges from emerging technologies, such as cryptocurrencies and electronic services, but these have not significantly undermined their dominance [8]. Group 4: Market Sensitivity and Valuation Concerns - The current sell-off reflects broader anxieties regarding elevated stock valuations, which have been pushed up by a surge in AI spending and unexpected economic resilience in the U.S., making investors highly sensitive to negative signals [10]. - In a tense market environment, even minor product launches from small startups can lead to significant volatility in large public companies, as investors prefer to err on the side of caution regarding potential AI disruptions [10].

Core Viewpoint - Nomura Securities believes that DeepSeek's upcoming next-generation model V4 may further reduce training and inference costs through innovative architectures mHC and Engram technology, accelerating the innovation cycle of China's AI value chain [2][4][5]. Group 1: Innovation in Technology Architecture - The report indicates that computing chips and memory have been bottlenecks for China's large models, and V4 is expected to introduce two key technologies—mHC and Engram—to optimize these constraints from both algorithmic and engineering perspectives [7]. - mHC, or "Manifold Constraint Hyperconnection," aims to address the bottleneck of information flow and training instability in deep Transformer models, enhancing the communication between neural network layers [8]. - Engram is a "conditional memory" module designed to decouple "memory" from "computation," allowing static knowledge to be stored in a sparse memory table, which can be quickly accessed during inference, thus freeing up expensive GPU memory for dynamic calculations [11]. Group 2: Impact on AI Development - The combination of these two technologies is significant for China's AI development, as mHC provides a more stable training process to compensate for potential shortcomings in domestic chips, while Engram smartly manages memory to bypass HBM capacity and bandwidth limitations [13]. - Nomura emphasizes that the most direct commercial impact of V4 will be a further reduction in the training and inference costs of large models, stimulating demand and benefiting Chinese AI hardware companies through an accelerated investment cycle [13][14]. Group 3: Market Dynamics and Competition - Nomura believes that major global cloud service providers are still in a race for general artificial intelligence, and the capital expenditure competition is far from over, suggesting that V4 is unlikely to create the same level of shockwaves in the global AI infrastructure market as last year [15]. - However, global large model and application developers are facing increasing capital expenditure burdens, and if V4 can significantly lower training and inference costs while maintaining high performance, it will serve as a strong boost for these players [15][16]. - The report reviews the market landscape one year after the release of DeepSeek's V3 and R1 models, noting that these models accelerated the development of Chinese LLMs and applications, altering the competitive landscape and increasing attention on open-source models [16]. Group 4: Software Evolution - On the application side, the more powerful and efficient V4 is expected to give rise to more capable AI agents, transitioning from "dialogue tools" to "AI assistants" that can handle complex tasks [20][21]. - This shift will require more frequent interactions with underlying large models, increasing token consumption and thereby raising computing demand [21]. - Consequently, the enhancement of model efficiency is not expected to "kill software," but rather create value for leading software companies that can leverage the capabilities of the new generation of large models to develop disruptive AI-native applications or agents [22].

Core Viewpoint - Cloudflare reported a strong fourth-quarter performance with revenue of $614.5 million, a year-over-year increase of 33.6%, driven by surging demand for AI services, leading to the fastest growth since 2021 [2][3][4]. Group 1: Financial Performance - The company signed its largest annual contract ever, with an average annual value of $42.5 million, and the total value of new annual contracts increased by nearly 50% year-over-year, marking the fastest growth since 2021 [2][3]. - Cloudflare's earnings per share for the fourth quarter were $0.28, exceeding analyst expectations, while the net loss narrowed from $12.8 million in the previous year to $12.1 million [4]. - For 2026, the company expects total revenue to reach between $2.79 billion and $2.8 billion, surpassing market expectations of $2.74 billion [3][11]. Group 2: AI Integration and Market Position - The transition towards AI and intelligent agents is fundamentally restructuring internet platforms, which is driving demand for Cloudflare's services [4][7]. - Cloudflare is seen as a direct beneficiary of the rapid integration of AI across industries, with approximately 80% of leading AI companies utilizing its solutions [12]. - The company is positioned as the platform for new internet users represented by intelligent agents, as stated by CEO Matthew Prince [8]. Group 3: Market Outlook - Cloudflare's first-quarter revenue guidance for 2026 is projected to be between $620 million and $621 million, above market expectations of $613.9 million [11]. - Adjusted earnings per share for 2026 are expected to be between $1.11 and $1.12, slightly below market expectations of $1.19 [11]. - Despite the positive outlook, the company has not been immune to the broader sell-off in the software sector, with software ETFs down 21% year-to-date [13].