Core Insights - A new reasoning framework for agents called DualPath has been introduced, which addresses I/O bottlenecks in long-text reasoning scenarios by optimizing the speed of loading KV-Cache from external storage [1][3]. Group 1: DualPath Framework - DualPath changes the traditional Storage-to-Prefill loading mode by introducing a second path, Storage-to-Decode, allowing for more efficient data handling [3][6]. - The framework utilizes idle storage network interface card (SNIC) bandwidth from the decoding engine (DE) to read caches and employs high-speed computing networks (RDMA) to transfer data to the prefill engine (PE), achieving global pooling of storage bandwidth and dynamic load balancing [3][13]. Group 2: Performance Improvements - In tests with a production-level model of 660 billion parameters, DualPath demonstrated a remarkable increase in offline inference throughput by 1.87 times and an average increase in online service throughput by 1.96 times [3][14]. - The framework significantly optimizes first token latency (TTFT) under high load while maintaining stable token generation speed (TPOT) [5][14]. Group 3: Technical Innovations - DualPath allows KV-Cache to be loaded into the decoding engine first, which is then transmitted to the prefill engine, alleviating bandwidth pressure on the prefill side [7][9]. - The architecture includes a central scheduler that dynamically allocates tasks based on I/O pressure and computational load, preventing congestion on any single network interface or computational resource [14][18]. Group 4: Research and Development - The first author of the paper, Wu Yongtong, is a PhD student at Peking University, focusing on system software and large model infrastructure, particularly in optimizing inference systems for large-scale deployment [15][16].

Summary of Key Points from the Conference Call Industry and Company Involved - The discussion centers around the AI industry, specifically focusing on the upcoming release of DeepSeek V4 "Sea Lion Lite" and its implications for the domestic AI industry chain [1][2]. Core Insights and Arguments - DeepSeek V4 "Sea Lion Lite" is set to be released with a remarkable 1M ultra-long context window, which is expected to impress the industry [1]. - In a significant deviation from standard practice, DeepSeek has granted exclusive early access to domestic suppliers like Huawei, rather than to Nvidia or AMD, highlighting a strategic shift towards domestic collaboration [1]. - This decision emphasizes the synergy between domestic large models and computing power, marking a pivotal moment for the domestic AI industry [1]. - The naming of "Sea Lion Lite" suggests agility and efficiency, aligning with the strategic considerations behind choosing Huawei for technical compatibility and security [1]. Additional Important Content - DeepSeek has successfully completed model migration on the Ascend platform, achieving over a 35-fold increase in inference speed through the KernelCAT tool, which is crucial for closing the loop in the domestic AI industry chain [2]. - Huawei's recent launch of the Atlas 950 super node, capable of supporting 8192 Ascend 950DT chips with an FP8 computing power of 8 EFLOPS, represents a significant leap in computing scale and challenges the notion of "single-chip gap" [2]. - The release schedule of DeepSeek V4 is strategically aligned with the launch of Huawei's Atlas 950 in Q4 2026, indicating a coordinated effort in the market [2]. - Key beneficiaries identified include: - **High-tech Development and Tuo Wei Information** as core partners in the Ascend ecosystem, expected to benefit directly from the shipment of super node servers. - **Hua Feng Technology** as a core supplier of high-speed backplane connectors, likely to see increased demand due to the super node's high interconnect bandwidth of 16.3PB/s. - **Tai Jia Co., Xing Sen Technology, Yi Hua Co., and Hua Zheng New Materials** as foundational infrastructure for computing hardware. - **Chuan Run Co.** providing full liquid cooling solutions, and **Heng Wei Technology** collaborating deeply with Ascend on heterogeneous intelligent computing operations [2].

Market Overview - Major global stock indices showed mixed performance, with the Dow Jones Industrial Average up by 17.05 points (0.03%) to 49,499.2, while the Nasdaq fell by 273.7 points (-1.18%) to 22,878.38 [1] - The S&P 500 decreased by 37.27 points (-0.54%) to 6,908.86, while the European Stoxx 50 dropped by 11.76 points (-0.19%) to 6,161.56 [1] - In Asia, the Hang Seng Index fell by 220.46 points (-2.44%) to 8,814.29, while the Nikkei 225 rose by 170.27 points (0.29%) to 58,753.39 [1] Geopolitical Developments - Tensions escalated between Pakistan and Afghanistan following intense border clashes, with Pakistan conducting airstrikes in response to casualties and territorial losses [2] - Significant progress was reported in US-Iran nuclear negotiations, with plans for further discussions in Vienna, aimed at de-escalating regional tensions [3] - The US Treasury proposed new regulations to cut off Swiss MBaer Bank from the US financial system due to alleged support for illegal activities related to Russia and Iran [3] Technology and Innovation - DeepSeek, in collaboration with Tsinghua and Peking University, released a paper on the DualPath model inference system, enhancing offline inference by up to 1.87 times and online service by 1.96 times [4] - ASML confirmed that its next-generation EUV lithography machines are ready for mass production, which will significantly impact advanced semiconductor manufacturing [5] - Broadcom announced the shipment of the industry's first 3.5D face-to-face computing SoC, utilizing 2nm technology to support next-generation AI computing needs [6] Corporate Actions - Netflix rejected a bid to increase its offer for Warner Bros. Discovery, citing financial concerns, and announced a stock buyback plan, leading to a 10% increase in its after-hours stock price [5]

Market Overview - On February 26, 2026, the A-share market saw all three major indices close higher, with the Shanghai Composite Index at 3888.60 points, up 0.34%, and the Shenzhen Component Index at 12984.08 points, up 0.85% [1] - The Hong Kong stock market experienced a decline, with the Hang Seng Index falling 1.44% to 26381.02 points, and the Hang Seng Technology Index down 2.87% to 5109.33 points [1] - In the U.S. stock market, the S&P 500 Index decreased by 0.54% to 6908.86 points, while the Dow Jones Industrial Average saw a slight increase of 0.03% to 49499.2 points [1] Banking Sector - The People's Bank of China announced support for domestic banks to conduct RMB cross-border interbank financing under lawful, compliant, and risk-controlled principles [2] - Domestic banks are required to manage these financing activities centrally and establish robust risk management and internal control mechanisms [2] - A warning mechanism must be established for when the net RMB financing balance approaches 80% of the upper limit [2] AI Industry - For the first time, China's AI model API usage surpassed that of the U.S., with 4.12 trillion tokens called in the week of February 9-15, compared to 2.94 trillion tokens in the U.S. [3] - The following week, China's usage further increased to 5.16 trillion tokens, marking a 127% rise over three weeks, while U.S. usage fell to 2.7 trillion tokens [3] - Four out of the top five AI models by usage are from Chinese companies, contributing 85.7% of the total calls [3] Baidu Financial Performance - Baidu Group reported fourth-quarter revenue of 32.74 billion RMB, a 5% increase from the previous quarter, slightly above market expectations [3] - AI business revenue accounted for 43% of Baidu's general business revenue, with smart cloud infrastructure revenue at 5.8 billion RMB and AI high-performance computing subscription revenue up 143% year-over-year [3] - The adjusted EBITDA for the fourth quarter was 4.7 billion RMB, with an EBITDA margin of 14% [3]

Group 1: DeepSeek and AI Models - DeepSeek's new model "sealion-lite" is in active testing, supporting a 1M context window and native multimodal reasoning, surpassing the V3.2 thinking mode [1] - DeepSeek has provided early access to V4 for domestic chip manufacturers like Huawei to optimize processor software, while Nvidia and AMD have not received access [1] - Initial SVG examples indicate that V4 Lite has a simpler and higher quality code, with speculations around 285 billion parameters, preparing the market for another "DeepSeek moment" [1] Group 2: Grok 4.20 Update - Grok 4.20 features a "4 Agents" architecture, including a coordinator and three specialists, which collaborate automatically for complex queries [2] - It ranked first in Search Arena, surpassing GPT-5.2 and Gemini 3.0 Pro, and also topped the Alpha Arena real stock trading benchmark [2] - The model employs a rapid learning mechanism, iterating weekly through real user interactions, significantly reducing hallucinations by about 65% and improving reliability in multi-step reasoning [2] Group 3: Perplexity and Anthropic Developments - Perplexity launched a Computer product that orchestrates up to 19 AI models for end-to-end research, design, coding, and deployment, capable of running autonomously for hours or days [3] - The founder claims "AI is the computer," enabling the creation of a real-time financial terminal comparable to Bloomberg [3] - Anthropic acquired AI startup Vercept, with its core capabilities to be integrated into Claude, which has improved its performance in OSWorld benchmark tests from under 15% to 72.5%, nearing human levels [3] Group 4: Samsung Galaxy S26 Series - Samsung's Galaxy S26 series features a customized Snapdragon 8 Gen 2 chip, enabling AI to autonomously perform tasks like ride-hailing and shopping [4] - The S26 Ultra introduces an embedded anti-peep display and supports professional video standards, significantly enhancing night photography and video stabilization [4] - The starting price for the standard version is 6,999 yuan, an increase of 1,000 yuan from the previous generation, while the S26 Ultra starts at 9,999 yuan, up by 300 yuan, with a target of over 400 million AI-supported Galaxy devices by the end of 2025 [4] Group 5: Talent Movement in AI - A prominent Chinese talent, Pang Ruoming, left Meta after seven months for OpenAI, despite Meta offering over $200 million in a multi-year compensation package [5][6] - Pang previously expanded a small team at Apple into a large-scale model team and led the development of key AI features [6] - His departure coincided with a critical period for Meta's AI lab, which had just delivered its first core AI models [6] Group 6: AI Programming Transformation - Karpathy asserts that a significant transformation in AI programming began in December 2022, predicting that coding agents will be ineffective until December 2025 [7] - Programming is being restructured to involve AI agents managing multiple parallel code instances rather than traditional coding methods [7] - The author of Ruby on Rails describes this as the fastest and most significant change in 40 years of computing, emphasizing that skilled programmers will enhance their capabilities rather than be replaced [7] Group 7: AI Agent Audit Findings - A joint report from MIT and other institutions audited 30 top AI agents across 45 dimensions, revealing that 23 are completely closed-source, with a high concentration of underlying models among GPT, Claude, and Gemini [8] - The actual autonomy of browser-type agents is rated at L4-L5, while companies often misrepresent them as L1-L2, with only four agents disclosing dedicated security documentation [8] - Programming accounts for nearly half of agent usage, but only 0.04% of the global population has tried AI programming, highlighting a significant gap in governance frameworks [8]

Group 1: C3.ai Performance and Strategy - C3.ai (AI) shares fell 23% in premarket trading after reporting a quarterly earnings miss and projecting revenue below expectations [5] - The company announced a major restructuring plan, cutting $135 million in expenses, which includes a 26% reduction in workforce [5] - For Q4, C3.ai expects total revenue of $48.0 – $52.0 million, significantly below the consensus estimate of $77.72 million [5] Group 2: C3.ai Bookings and Market Position - C3.ai highlighted strong federal, defense, and aerospace bookings, with federal bookings increasing by 134% year-over-year, accounting for 55% of total bookings [5] - Key customer wins include the U.S. Department of Agriculture, U.S. Department of Energy, NATO, Royal Navy, GSK, Thales, ExxonMobil, and U.S. Steel [5] Group 3: Nvidia Financial Results - Nvidia (NVDA) shares rose 1.3% after reporting fiscal fourth-quarter results, with adjusted earnings of $1.62 per share and revenue of $68.13 billion, a 73% year-over-year increase [5] - For the fiscal first quarter, Nvidia expects revenue to be around $78 billion, exceeding analysts' expectations of $72.78 billion [5] Group 4: DeepSeek's Strategic Move - Chinese AI firm DeepSeek has withheld its upcoming AI model from U.S. chipmakers, which deviates from standard industry practices [5] - This move is perceived as part of a broader strategy by the Chinese government to disadvantage U.S. hardware and models in China [5]

Core Insights - In February, China's AI model API call volume surged, surpassing that of the United States for the first time, with 41.2 trillion tokens compared to the U.S.'s 29.4 trillion tokens during the week of February 9-15 [1][7] - The following week, China's model call volume increased to 51.6 trillion tokens, marking a 127% growth over three weeks, while the U.S. model's call volume dropped to 27 trillion tokens [1][7] - Four out of the top five models in global API call volume are from Chinese manufacturers, indicating a collective rise rather than reliance on a single product [1][10] Token Call Volume Growth - The global token call volume for major models has seen explosive growth, increasing from 12.4 trillion tokens in early March 2025 to 139.5 trillion tokens by mid-February 2026, a tenfold increase in less than a year [6] - In early 2026, U.S. models showed signs of fatigue in growth, while Chinese models began to accelerate rapidly, with a call volume of 22.7 trillion tokens in the first week of February [6][7] Leading Models and Their Performance - The top five models by call volume during the week of February 16-22 included four from China: MiniMax's M2.5, Kimi K2.5, GLM-5, and DeepSeek's V3.2, contributing 85.7% of the total call volume [10] - MiniMax's M2.5 model achieved 14.4 trillion tokens in its first week, while Kimi K2.5's innovative architecture significantly boosted its call volume and revenue [10][13] Cost Competitiveness - Chinese models are significantly cheaper than their U.S. counterparts, with input costs at $0.3 per million tokens compared to $5 for U.S. models, and output costs at $1.1 and $2.55 versus $25 for U.S. models [15][16] - The cost advantage is attributed to innovative algorithm architectures, such as the Mixture-of-Experts (MoE) model, which reduces computational costs and increases efficiency [18] Market Dynamics and Future Trends - The demand for AI tokens is expected to grow exponentially, with a projected compound annual growth rate of 330% from 2025 to 2030, leading to a 370-fold increase in consumption [19] - The shift in AI usage from simple Q&A to complex task execution is driving this growth, as users increasingly rely on AI for productivity [20] - Future pricing models for AI services are anticipated to become highly customized and flexible, reflecting the complexity of tasks and consumption patterns [22]

Core Insights - In February, China's AI model API call volume surged, surpassing that of the United States for the first time, with 41.2 trillion tokens compared to the U.S.'s 29.4 trillion tokens during the week of February 9-15 [1][7] - The following week, China's model call volume increased to 51.6 trillion tokens, marking a 127% growth over three weeks, while U.S. model calls dropped to 27 trillion tokens [1][7] - Four out of the top five models in global API call volume are from Chinese companies, indicating a collective rise of Chinese AI manufacturers rather than reliance on a single product [1][10] Token Call Volume Growth - The global model token call volume has experienced explosive growth, increasing from 12.4 trillion tokens in the week of March 3-9, 2025, to 139.5 trillion tokens by mid-February 2026, a tenfold increase in less than a year [6] - In early February 2026, China's model call volume reached 22.7 trillion tokens, signaling a strong competitive push against U.S. models [6][7] Leading Models and Their Performance - The top five models by call volume during the week of February 16-22, 2026, included four from Chinese manufacturers, contributing 85.7% of the total call volume [10] - MiniMax's M2.5 model, launched on February 13, 2026, quickly became the top model, contributing 14.4 trillion tokens to the total call volume of 32.1 trillion tokens during the week of February 9-15 [10] Cost Competitiveness - Chinese models are significantly cheaper than their U.S. counterparts, with MiniMax's M2.5 and Zhiyu's GLM-5 priced at $0.3 per million tokens for input, compared to $5 for Claude Opus 4.6, making Chinese models approximately 16.7 times cheaper [15][16] - For output, MiniMax's M2.5 costs $1.1 per million tokens, while Claude Opus 4.6 costs $25, representing a cost difference of about 22.7 times [16][17] Technological Innovations - The "Mixture-of-Experts" (MoE) architecture is a key factor in reducing inference costs for Chinese models, allowing for significant reductions in memory usage and increases in throughput [18] - Chinese AI companies are also pursuing vertical integration to optimize costs further, combining model algorithms, cloud infrastructure, and AI chips for better efficiency [19] Market Trends and Future Projections - The demand for tokens is expected to grow exponentially, with a projected compound annual growth rate of 330% from 2025 to 2030 in China [19] - The concept of "Token inflation" reflects a structural increase in token consumption per user, driven by deeper engagement with AI tools for complex tasks [20] - Future AI service pricing is anticipated to become highly customized and flexible, influenced by task complexity and resource consumption [22]

Core Insights - In February, China's AI model API call volume surged, surpassing that of the United States for the first time, with 41.2 trillion tokens compared to the U.S.'s 29.4 trillion tokens during the week of February 9-15 [2][9] - The following week, China's model call volume increased to 51.6 trillion tokens, marking a 127% growth over three weeks, while U.S. model calls dropped to 27 trillion tokens [2][9] - Four out of the top five models in global API call volume are from Chinese manufacturers, indicating a collective rise rather than reliance on a single product [2][12] Token Call Volume Growth - The OpenRouter platform, which aggregates AI models, reported a dramatic increase in global model token call volume, rising from 12.4 trillion tokens in early March 2025 to 139.5 trillion tokens by mid-February 2026, a growth of over tenfold in less than a year [8] - In early February 2026, Chinese models accounted for a significant increase in call volume, signaling a shift in market dynamics [8][9] Competitive Landscape - The top five models by call volume during the week of February 16-22, 2026, included four from Chinese companies, contributing 85.7% of the total call volume [12] - MiniMax's M2.5 model quickly became the top model within a week of its launch, contributing 14.4 trillion tokens to the total call volume [12][15] Cost Advantages - Chinese models, such as MiniMax's M2.5 and Zhiyu's GLM-5, offer significant cost advantages, with input costs at $0.30 per million tokens compared to $5 for U.S. counterparts like Claude Opus 4.6, making them approximately 16.7 times cheaper [18][19] - The output costs for Chinese models are also significantly lower, with MiniMax's M2.5 at $1.10 per million tokens versus $25 for Claude Opus 4.6, highlighting a cost disparity that influences developer choices [18][19] Technological Innovations - The "Mixture-of-Experts" (MoE) architecture is a key factor in reducing inference costs for Chinese models, allowing for efficient resource utilization by activating only relevant parts of the model for specific tasks [20] - This architecture can reduce memory usage by 60% and increase throughput by up to 19 times, contributing to the overall cost advantage [20] Market Trends - The demand for AI tokens is expected to grow exponentially, with a projected compound annual growth rate of 330% from 2025 to 2030 in China, indicating a significant market opportunity [21] - The evolution of AI from a simple Q&A tool to a productivity tool is driving increased token consumption, as users engage in more complex tasks [22][23] Future Pricing Models - The pricing of AI services is anticipated to shift towards a more customized and flexible model, influenced by task complexity and resource consumption, moving away from a one-size-fits-all approach [24]

Core Insights - The capital landscape in the large model sector is rapidly evolving, with companies like Jieyue Xingchen preparing for an IPO to raise approximately $500 million, following a record-breaking financing round of over 5 billion RMB [3] - The industry is experiencing significant differentiation, with companies like Zhipu AI and MiniMax successfully listing on the Hong Kong Stock Exchange, while others like Lingyi Wanshu and Baichuan Intelligence are shifting focus to more practical business models [5][6] - The competitive dynamics have shifted from a focus on model capabilities to a more brutal "ecological positioning battle," as companies adapt to the harsh realities of high operational costs and market competition [5][9] Company Strategies - Zhipu AI has chosen to focus on enterprise services, targeting large clients such as banks and government entities, thereby avoiding direct competition with tech giants in the consumer market [11] - MiniMax is betting on the overseas consumer market, aiming to differentiate itself with lightweight models and engaging user experiences, while facing challenges in maintaining user retention and profitability [12][14] - Yuezhianmian is concentrating on technological advancements in the domestic market, developing competitive models like K2 to carve out a niche despite the competitive pressures from larger players [16] - Jieyue Xingchen is pursuing a hybrid model that integrates AI capabilities into physical devices, aiming to become a core component in smart hardware, thus linking its success to the growth of the hardware industry [16][18] Industry Outlook - The future landscape of the large model industry is expected to resemble a "pyramid structure," with a few dominant players at the top, specialized vertical players in the middle, and a vibrant ecosystem of applications at the base [22] - The complex competitive relationships may replace simple confrontations, with top players potentially launching their own vertical applications, leading to alliances among different vertical players to counteract giants [24] - The emergence of new technological paradigms could disrupt existing ecological positions, leaving the future of the industry uncertain [25]