Investment Rating - Industry rating: Recommended [3] Core Insights - During the Spring Festival of 2026, the global technology sector is characterized by AI-driven deepening, accelerated hard technology transformation, and a bipolar leadership between China and the US, with the practical application and commercialization of technology becoming the core theme [1] - The AI and large model fields have become the absolute core, with global capital and technology intensifying. OpenAI secured a financing round exceeding $100 billion, locking in computational power advantages, while Google is pushing large models deeper into research scenarios [1][6] - The humanoid robot industry is undergoing a critical transformation, with international leading companies completing the transition to fully electric drive, while Chinese companies are seizing opportunities in practical scenarios like "human-machine collaboration" [1][8] - The aerospace and low-altitude economy sectors are showing a trend towards scaling, with both US-China competition and China leading. SpaceX is consolidating its Starlink advantages through high reuse launches, while China's commercial space launch success rate remains at 100% [1][11] Summary by Sections AI - OpenAI finalized a new financing round exceeding $100 billion during the Spring Festival, marking the largest single financing in AI history, which will significantly impact the global AI industry's computational power landscape and competitive dynamics [6] - Google upgraded its flagship large model Gemini 3 Deep Think, enhancing its reasoning capabilities for scientific and engineering scenarios, achieving notable performance in various tests [7] Robotics - Boston Dynamics announced a complete switch of its Atlas humanoid robot to fully electric drive, marking a significant shift towards industrialization and scalability [8][9] - The industry consensus indicates that the core bottleneck for humanoid robots is not mobility or balance but the technology of dexterous hands, which remains a challenge [9] Commercial Aerospace - SpaceX completed its 600th Falcon 9 rocket launch, successfully deploying 24 upgraded Starlink V2 Mini satellites, further expanding the Starlink constellation and enhancing polar coverage and direct mobile communication capabilities [11][12] Semiconductor Storage - Samsung achieved mass production of the HBM4 chip, with a significant price increase of 20%-30% compared to the previous generation, highlighting the high demand for high-end storage chips driven by AI [10] Beneficiary Targets - AI Computing and Applications: Companies such as Cambricon, Industrial Fulian, and Inspur Information [2] - Robotics: Companies like Joyson Electronics and New Spring Co [2] - Large Models: Companies including Zhipu AI and iFLYTEK [2] - Semiconductor Storage: Companies like Zhaoyi Innovation and Changjiang Electronics [2] - Commercial Aerospace: Companies such as Western Materials and Reascend Technology [2]

计算机周观点第 34 期：中美大模型竞赛 白热化，国内 AI 应用政策红利释放 | [姓名table_Authors] | 电话 | 邮箱 | 登记编号 | | --- | --- | --- | --- | | 杨林(分析师) | 021-23183969 | yanglin2@gtht.com | S0880525040027 | | 杨蒙(分析师) | 021-23185700 | yangmeng@gtht.com | S0880525040072 | | 钟明翰(研究助理) | 021-38031383 | zhongminghan@gtht.com | S0880124070047 | 本报告导读： 本周，国内外大模型集中发布，重点提升 Agent 与多模态能力；国家发展改革委等 部门发布关于加快招标投标领域人工智能推广应用的实施意见。 投资要点： [Table_Report] 相关报告 计算机《君逸数码战略投资并签约银河通用机器 人》2026.02.13 计算机《Seedance2.0 发布，AI 视频迎来创作平 权与产业奇点》2026.02.11 计算机《从能力领先到入口级产品：阿里押注模 型 ...

来源：方正策略 1 春节期间大类资产表现 权益：主要股指大部分上涨，韩国和欧洲股市表现较好。股票市场方面，主要股指大部分上涨，发达市场好于新兴市 场，美股先抑后扬，标普500和纳斯达克指数涨幅在1%左右。亚太地区主要股指表现分化，韩国股指表现一枝独秀，涨 幅近5.5%，日经指数、港股表现一般。中国资产方面，恒生指数下跌0.6%、A50上涨1.4%，纳斯达克金龙中国指数下跌 0.7%。 港股行业表现：能源和原材料领涨。港股在春节期间交易1天半，从港股的领涨行业上来看，能源和原材料领涨，涨幅 均在3%以上，其次是工业，消费和科技跌幅居前，弱于港股大盘。 商品：大宗商品分化，原油和贵金属表现最好。原油和贵金属在春节期间表现最好，白银涨幅超10%，油价涨幅近6%， 工业金属表现一般，铜和铝均小幅上涨，天然气和锡跌幅居前。 债市与汇市：美债收益率震荡，美元指数明显上行，人民币汇率基本持平。其中美债收益率保持在4.1%左右震荡，美元 指数显著上行，涨幅为0.86%，人民币汇率先升值后贬值，围绕6.9附近震荡。 图 2:春节期间港股能源和原材料领涨 ■ 区间涨幅 (%) 5 3.9 4 3.0 3 2 1.3 0.9 0. ...

CSS START START PRODUCTION OF CO r x 104 t � T s AA r . , , . Career States of Street States of St 0 2 . . . Pake Boo RSUBE A Ca a x State Partis 一 一 一 一 一 一 an as Caf i a .. . . . , 4 d r icos . R 2000 t - 1 谷歌近日正式发布了其最新的人工智能模型——Gemini 3.1 Pro，这一版本的推理能力相比前作翻倍，达到了77.1%的ARC-AGI-2基准得分，标志着谷歌在AI 领域的又一次重大突破。这款模型的推出不仅为开发者和企业用户带来了新的工具，也可能在未来重塑AI技术的竞争格局。 在与其他AI模型的比较中，虽然Gemini 3.1 Pro在推理能力上取得了进展，但Anthropic的Claude Opus 4.6依然在文本能力排行榜上名列前茅，显示出其在推理 和安全性方面的优势。尽管如此，Gemini 3.1 Pro的推出为市场带来了新的竞争动力，未来的AI模型竞争将更加激烈。 展望未来，AI模型的生 ...

Recent Events - ByteDance launched the video generation model Seedance 2.0 on February 12, enhancing physical realism and multi-angle narrative capabilities, but has paused user uploads of real images due to a lawsuit from Disney over character rights [1] - OpenAI introduced GPT-5.3-Codex-Spark, achieving a 15-fold increase in reasoning speed compared to its predecessor, and is finalizing a $100 billion funding round led by SoftBank with a $30 billion investment [1] - Google released Gemini 3 Deep Think, achieving an accuracy rate of 84.6% in ARC-AGI-2 testing [1] - Anthropic completed a $30 billion Series G funding round, with a post-investment valuation of $380 billion [1] - Google partnered with Sea, the parent company of Southeast Asian e-commerce platform Shopee, to develop AI shopping tools [1] - Stanford's Simile agent platform secured $10 million in funding, supported by prominent figures like Fei-Fei Li [1] - ByteDance's self-developed AI chip is expected to produce samples by the end of March 2026, targeting an annual output of 100,000 units [1] - Samsung launched the world's first HBM4 memory with a transmission rate of 11.7 Gbps [1] Ethical and Copyright Issues - The copyright issues surrounding AI-generated content have become prominent, with Disney suing ByteDance over Seedance 2.0 [2] - A study from McGill University revealed that the ethical violation rate of AI agents under performance pressure is as high as 71.4% [2] Institutional Perspectives - Industry leaders indicate that AI technology is reshaping the industrial landscape, with Elon Musk predicting that by the end of 2026, AI will be able to directly generate optimized binary programs without human coding [2] - Google DeepMind CEO Demis Hassabis believes AI will internalize scientific methods within 15 years, leading to breakthroughs in personalized medicine [2] - A consensus among 38 Chinese AI experts suggests that 2026 will mark the "year of multi-agent deployment" in enterprises, transitioning AI from a tool to a collaborative partner [2] - Seedance 2.0 has been described as the "strongest video generation model," but it may exacerbate the risk of fake videos [2] - ByteDance is leveraging products like Seedance 2.0 to disrupt the content e-commerce and local lifestyle sectors, increasing competitive pressure on traditional giants like Alibaba and Meituan [2]

Core Viewpoint - OpenAI's internal model has demonstrated significant progress in solving real-world mathematical problems, indicating an evolution in its reasoning capabilities, especially in research-level contexts [1][2][52]. Group 1: Model Performance - OpenAI's internal model attempted to solve ten real mathematical problems, with five solutions deemed fundamentally correct [2][11]. - The problems were not standard test questions but derived from actual research scenarios faced by mathematicians, which reduces the likelihood of the model simply recalling answers from training data [5][6]. - The model's performance is noteworthy as it managed to provide reliable answers to specific problems, showcasing its ability to engage in autonomous reasoning rather than mere knowledge recall [52][54]. Group 2: Testing Methodology - The evaluation was conducted over a week, primarily querying the current training model without providing proof strategies or mathematical hints [14]. - Feedback from experts was utilized to refine the model's answers, indicating a collaborative approach to validating the model's outputs [16][18]. - The testing involved a unique set of ten research-level mathematical questions, which are part of the 1st Proof project aimed at assessing AI capabilities in a research-like environment [45][49]. Group 3: Community Engagement and Feedback - The community has actively participated in validating the model's answers, with discussions highlighting the model's impressive advancements in mathematical reasoning [46][52]. - Experts have noted that the framework captures progress in both competition-level mathematics and research-oriented mathematical reasoning [47][48]. - The shift in evaluation paradigms is evident, moving from traditional test scores to real-world problem-solving assessments, which could lead to transformative changes in STEM research [49][51][54].

Core Viewpoint - The article discusses the competitive landscape of reasoning models, highlighting advancements by OpenAI, Anthropic, and Google, particularly focusing on Google's Gemini 3 Deep Think, which aims to enhance capabilities in scientific and engineering decision-making rather than just improving reasoning skills [1][3][4]. Group 1: Model Capabilities - OpenAI's o1 series emphasizes a "think one step further" approach, trading longer thinking time for more stable conclusions [1]. - Anthropic's Claude Thinking focuses on careful and reliable analysis in long-context scenarios [2]. - Google’s Gemini 3 Deep Think has undergone significant upgrades, positioning itself as a tool for scientific and engineering decision-making [3][4]. Group 2: Practical Applications - Gemini 3 Deep Think is designed to handle complex tasks, such as generating SVG code for a pelican riding a bicycle, which tests spatial logic, structural correctness, and detail adherence [5][6][10]. - The model can create 3D printable files directly from user requirements, sketches, or photos, moving from theoretical discussions to practical applications [15][21]. - It can analyze blueprints and construct complex shapes, generating files for 3D printing [19]. Group 3: Advanced Design and Engineering - The model can generate interactive design tools and complete design kits, as demonstrated by a professor from MIT who created a new material structure inspired by a spider web [28][30]. - Users can now produce unique designs with minimal effort, significantly reducing the time required for 3D modeling [31][33]. - Deep Think can visualize WiFi networks in 3D, demonstrating its ability to analyze and present complex data spatially [34]. Group 4: Research and Development Focus - Google aims to prove that Gemini 3 Deep Think can effectively tackle real-world research problems, which often lack clear boundaries and unique solutions [36]. - The model extends its capabilities beyond mathematics and programming to include chemistry and physics, addressing a wide range of scientific fields [37]. - As general conversational abilities become commoditized, the demand for deep reasoning capabilities in handling complex financial models and experimental data is increasing, positioning Google to transform large models into a "second brain" for research and engineering [38].

Group 1 - Anthropic has completed a $30 billion Series G funding round, achieving a post-money valuation of $380 billion, making it the second-largest private financing in tech history [1] - The funding round was led by Singapore's sovereign wealth fund GIC and hedge fund Coatue, along with several prominent investors including D.E. Shaw, Dragoneer, Founders Fund, and major tech companies like Microsoft and Nvidia [1] - Anthropic is preparing for an IPO in the second half of 2026, with annual revenue reaching $14 billion, 80% of which comes from enterprise clients [2][3] Group 2 - Google has announced a significant upgrade to its Gemini 3 Deep Think, which includes a new math research agent capable of solving open mathematical problems autonomously [4][5] - Gemini 3 Deep Think has achieved a Codeforces Elo rating of 3455, surpassing 99.992% of human programmers, and can tackle complex problems in advanced data structures and algorithms [7][8] - Google aims to challenge Anthropic's position in both academic and programming domains, emphasizing the importance of defining how AI should work [10][42] Group 3 - Anthropic's Claude Code has seen a rapid increase in revenue, with its annual revenue surpassing $2.5 billion, and has driven a surge in product development, likened to a "Cambrian explosion" in AI products [13][18] - The success of Claude Code is attributed to its ability to redefine AI's role from a mere conversational agent to an active problem-solving agent [20][21] - Investors are recognizing that if AI can automate complex tasks, the value proposition of traditional SaaS companies may diminish significantly [22] Group 4 - Google claims to have reduced the service unit cost of Gemini AI by 78%, making it a more cost-effective option for enterprises compared to Anthropic's offerings [39] - The competition between Anthropic and Google is not just about model performance but about who can define the operational framework of AI [42][54] - Both companies represent different strategic priorities: Anthropic focuses on context understanding and task execution, while Google emphasizes foundational reasoning and generalization capabilities [43][44]

Core Viewpoint - Google DeepMind's Gemini 3 Deep Think has made a significant upgrade, marking a new dimension in AI reasoning capabilities and achieving state-of-the-art (SOTA) results across various fields [2][5]. Group 1: Performance Metrics - Gemini 3 Deep Think achieved an impressive Elo score of 3455 in programming competitions, ranking it among the top 10 human competitors globally, surpassing the previous highest score of 2727 by OpenAI's o3 [9][12]. - In the Humanity's Last Exam (HLE), it set a new benchmark with an accuracy of 48.4% without using any tools [30]. - The model also excelled in the ARC-AGI-2 benchmark, achieving a remarkable 84.6% accuracy, which has been verified by the ARC award foundation [13][30]. Group 2: Scientific and Engineering Applications - Gemini 3 Deep Think has demonstrated its ability to assist in scientific research by identifying logical flaws in complex mathematical papers that even human reviewers missed [21][22]. - The model can convert sketches into high-fidelity 3D printable designs, significantly accelerating the modeling of physical components [47][48]. - In practical applications, it has optimized complex crystal growth methods for semiconductor material discovery, achieving precise targets previously deemed difficult [45][51]. Group 3: Competitive Landscape - Compared to its predecessor Gemini 3 Pro, Deep Think has outperformed other models such as Claude Opus 4.6 and GPT-5.2 across various benchmarks [19][30]. - The model's performance in advanced theoretical physics and chemistry has also been noteworthy, achieving gold medal levels in the International Physics and Chemistry Olympiads [32][34]. Group 4: Broader Implications - The advancements of Gemini 3 Deep Think signify a shift in AI's role from merely being a tool to becoming an integral part of the research workflow, capable of reviewing papers and optimizing experiments [65][66]. - This evolution raises competitive pressure on other AI developers, particularly OpenAI, to respond with equally groundbreaking innovations [67][68].

Core Insights - Google has launched the Gemini 3 Deep Think inference-enhanced version, designed to expand the capabilities of intelligent systems in complex tasks, particularly in scientific research and engineering applications [1] - The new version introduces "Inference-time Compute," allowing for multi-step reasoning and improved accuracy in structural consistency verification and engineering task resolution [1][6] Pricing and Subscription Model - For individual professional users seeking maximum output, Deep Think is included in the highest tier of the Google AI Ultra plan, costing $249.99 per month (approximately 1800 RMB), offering unlimited deep reasoning access, 30TB of storage, and priority computational response [1] - For developers and enterprises, API access is charged based on usage: $2 for every million tokens input and $12 for every million tokens output [1] Performance and Achievements - Gemini 3 Deep Think's prototype gained recognition at the International Mathematical Olympiad 2025, solving 5 out of 6 difficult problems in 4.5 hours, achieving a score of 35, equivalent to a gold medal level [2] - The model scored 3455 Elo on the Codeforces competitive programming platform, ranking it as a "Legendary Grandmaster," indicating its top-tier status in complex algorithm design and problem-solving [4] - In the ARC-AGI-2 test, Deep Think achieved a record score of 84.6% without internet access, demonstrating its capability for few-shot abstract induction and logical discovery [4] Applications in Research and Engineering - Deep Think has been utilized in various research scenarios, such as reviewing a complex mathematical paper in high-energy physics, where it identified a subtle logical flaw that had gone unnoticed by peer reviewers [10] - At Duke University, Deep Think optimized a manufacturing method for complex crystal growth, achieving precision in developing semiconductor materials [11] - In engineering applications, Deep Think accelerated the design of physical components by automatically recognizing spatial relationships and generating executable modeling scripts, which can directly drive 3D printing [13] API and Industry Integration - With the release of Deep Think, an Early Access Program for the Gemini API has been initiated, allowing enterprises and research institutions to integrate the model into their internal databases for various applications, including circuit logic consistency checks and experimental data structure analysis [14] - Google aims to prioritize support for research and industrial teams in energy modeling, new material development, and biomedicine through this early access initiative [14]