中信证券WAIC2026观察：建议沿四条具备明确催化的方向布局
Related
Agent of "1011 Insider Whale": AI computing power trading is undergoing a shift, with funds flowing from storage chips to hyperscale cloud providers.
According to Odaily Odaily, Garrett Jin, an agent of "1011 Insider Whale," pointed out that the market structure has changed significantly this week, with funds being redistributed within the AI industry chain. Change 1: Signs of a temporary peak in memory chips are emerging. He said that Micron's stock price encountered resistance and fell back around $1,250. Although the financial report was stronger than expected, the stock price was still falling with increasing volume, showing the typical top characteristics of "weakening after the good news is realized". Meanwhile, funds flowed out of the memory sector rapidly. DRAM-related ETFs saw a significant drop in trading volume, and SK Hynix and Samsung Electronics in the South Korean market also weakened. Data shows that foreign capital has withdrawn more than 100 trillion won (approximately US$65 billion) from the South Korean stock market in the past two months. Change 2: Funding shifts to AI hyperscale cloud providers. He pointed out that the real destination for funds is not small and mid-cap AI concept stocks, but core cloud computing giants such as Google, Microsoft, and Amazon. When the chip sector came under pressure last Friday, GOOG and MSFT had already stabilized with increased trading volume, and this week META further strengthened this trend with increased trading volume. Garrett Jin believes that the logic behind this round of capital migration is the "token optimization trend": as more and more simple tasks are handled by low-cost models, value will gradually be concentrated in token-billed cloud services and orchestration layers, rather than the basic model layer. This also constitutes the core moat of hyperscale cloud vendors, and the current strategy should focus on the catch-up opportunities of hyperscale cloud vendors.
The Ministry of Industry and Information Technology and six other departments have issued a directive to guide platform companies to strengthen their innovation in artificial intelligence fields such as general-purpose large-scale models, industry-specific large-scale models, and intelligent agents.
According to Mars Finance, seven departments, including the Ministry of Industry and Information Technology, jointly issued the "Action Plan for Promoting the Collaborative Development of Large, Medium, and Small Enterprises in the Platform Economy (2026-2028)". The plan guides platform enterprises to strengthen their innovation layout in artificial intelligence fields such as general-purpose large-scale models, industry-specific large-scale models, and intelligent agents; accelerate breakthroughs in key cutting-edge technologies and products such as high-end chips, next-generation operating systems, and next-generation intelligent terminals; and promote the verification, application, and dissemination of new technologies and products. It supports platform enterprises and SMEs in carrying out collaborative innovation, jointly applying for and undertaking national major science and technology projects and national key R&D programs with universities and research institutions, and conducting joint research on key technologies. It supports platform enterprises in issuing challenges and SMEs in accepting them, guiding SMEs to focus on their specialized technical fields. It also promotes the establishment of an innovation investment growth mechanism for platform enterprises, benchmarking against leading international companies, and enhancing their ability to guarantee investment in basic, original, and disruptive technologies. (Cailian Press)
Ministry of Human Resources and Social Security: By 2030, the large-scale model of the human resources and social security industry will be mature and complete, and the artificial intelligence application system will be basically sound.
According to Mars Finance, the Ministry of Human Resources and Social Security, the National Development and Reform Commission, the Ministry of Industry and Information Technology, and the National Data Administration recently jointly issued the "Implementation Opinions on Accelerating the Application and Development of 'Artificial Intelligence + Human Resources and Social Security'". A relevant official from the Ministry of Human Resources and Social Security answered reporters' questions regarding the "Implementation Opinions". In terms of promoting the work, a three-step work goal will be achieved within five years. First, building the foundation. Taking this year (2026) as a benchmark, the initial formation of the artificial intelligence application system, standard system, and guarantee system in the human resources and social security sector will be promoted. The infrastructure for "Artificial Intelligence + Human Resources and Social Security" applications will be deployed, a number of high-performance human resources and social security industry large-scale models and intelligent agent applications will be cultivated, and about 20 application scenarios based on human resources and social security industry large-scale models and corresponding high-quality datasets will be created, forming a collaborative development ecosystem of computing power, models, data, and scenario applications. Second, popularization and promotion. By 2027, a number of human resources and social security industry large-scale models and intelligent agents will be widely applied, and about 50 high-value application scenario empowerment paths will be explored, achieving significant results in intelligent development. Third, widespread application. By 2030, high-quality datasets will be effectively supplied, large-scale models for the human resources and social security industry will be mature and complete, and the artificial intelligence application system will be basically sound, forming an innovative landscape where artificial intelligence is widely applied in the human resources and social security sector. (Cailian Press)
B.AI Officially launched on Binance Wallet's Featured Recommendations, opening a new entry point for AI Agent.
According to ChainCatcher, the AI Agent underlying infrastructure project B.AI has been featured on Binance Wallet's Hot Picks section. Global users can easily find and access B.AI by opening their Binance Wallet and going to the "Discover - DApps - Tools" section to experience its cutting-edge AI capabilities. B.AI is dedicated to providing developers and ecosystem users with an efficient, convenient, and barrier-free platform for building and applying AI intelligent agents. This launch on Binance Wallet's Hot Picks section will allow B.AI to reach a wider Web3 community, further promoting the integration and large-scale application of AI and decentralized technologies.
Meituan open-sourced its trillion-parameter large-scale model LongCat-2.0, and simultaneously released the inference code for domestically developed Chinese card processors.
According to Beating's monitoring, Meituan has officially open-sourced its trillion-parameter large-scale model, LongCat-2.0, with a total of 1.6T parameters and an average activation of approximately 48B, designed specifically for real-world agentic coding tasks. Architecturally, it innovatively introduces LongCat sparse attention and N-gram embedding. The former reduces fragmented memory access through flow-aware indexing and hierarchical indexing, accelerating training and inference with millions of contexts; the latter, while achieving nearly 97% sparsity in MoE, invests 135B parameters into the embedding layer, balancing parameter gains and structural stability. Post-training employs multi-teacher online distillation, categorizing experts into Agent, Inference, and Interaction types, seamlessly integrating them on a domestic computing power cluster through the MOPD architecture. As the industry's first trillion-parameter model to complete inference on a 50,000-card domestic computing power cluster, LongCat-2.0 validates the mature capability of domestic chips to handle complex large-scale model tasks. To address the multiple limitations of domestically produced Chinese chips in terms of memory, bandwidth, and interconnects, Meituan has made breakthroughs in three areas: model, chip adaptation, and deployment. At the model level, ScMoE leverages the core control capabilities of domestically produced chips to achieve physical core-level parallelism for Dense and MoE branches, combined with KV-cache partitioning to alleviate the pressure on ultra-long context memory. At the chip adaptation level, Super Kernel reduces operator startup overhead, and Weight Prefetch hides I/O latency, maximizing hardware utilization under constrained conditions. At the deployment level, PD separation is adopted to balance TTFT and TPOT, along with asynchronous Expert-Parallel load balancing to solve load unevenness under high EP (efficiency level). This open-source release simultaneously provides multiple precision versions, including BF16, FP8, and INT8, and fully opens up inference results optimized for domestic computing power, aiming to enable existing domestically produced cards and even older cards to smoothly deploy trillion-model inference services. --------------------------------- Click the original link below to join the Beating · Lark AI news channel and monitor global AI hot topics and news 24/7.
Report: Fixed computing power masks the true capabilities of AI; the evolutionary speed of cutting-edge intelligent agents is underestimated by 60%.
According to Beating's monitoring, the UK AI Security Institute points out that current mainstream AI agent testing has significant blind spots. Evaluation methods with fixed computing power limits severely underestimate the true capabilities and iteration speed of models. The research team tested the performance of several cutting-edge large-scale models in benchmarks such as cybersecurity, software engineering, and mathematics. The test results show that the performance of an agent is not a fixed score, but rather a curve that continuously increases with test-time compute. In network attack and defense tests, when the computing power budget increased from 2.5 million tokens to 50 million tokens, the upper limit of the complexity of tasks that the most advanced agents could overcome (equivalent to human time) skyrocketed from 2 hours to 14 hours. Many attempts that failed with low computing power could eventually complete the task if given sufficient computing power to allow the agent to explore and correct errors. The new model's utilization efficiency of test-time compute power is significantly higher than that of the old model. Under a sufficient budget, the measured trend of cutting-edge capability evolution (the slope of the fitted curve) is approximately 60% steeper than in low-computing-power tests, demonstrating that traditional assessments severely underestimate the true iteration speed of AI. However, this computing power advantage has its limits; in fields such as healthcare where immediate feedback is lacking, increasing computing power does not necessarily improve the performance of intelligent agents. As inference costs decrease, low-budget assessments may lead decision-makers to underestimate the risks of AI agents in practical applications.