Recently, three group standards proposed and coordinated by the China Advertising Association—"Requirements for Generative Engine Optimization Services", "Requirements for Monitoring and Verifying the Optimization Effects of Generative Engines", and "Requirements for Classification, Grading, and Selection of Content Sources for Generative Engine Optimization"—were officially released and will be implemented starting from October 23, 2026. Currently, generative artificial intelligence is deeply integrated with brand communication, corporate marketing, and consumer decision-making. The application scenario of generative engine optimization as an emerging service form continues to expand. At the same time, the overall development in this field is still immature, and market behaviors need to be regulated. To guide market entities to fulfill compliance obligations, strictly abide by laws and regulations such as the "Anti-Unfair Competition Law", "Advertising Law", and "Consumer Rights Protection Law", effectively safeguard consumer rights, purify the industry environment, and meet the industry's demand for high-quality standardized development, under the guidance and support of government departments such as market supervision and standardization authorities, and industry organizations, the China Advertising Association has gathered the wisdom of various fields, including universities, research institutions, mainstream media, large model platforms, brand enterprises, technical service institutions, and professional legal service institutions, to compile these three group standards for generative engine optimization in the AI marketing field. These three standards provide strong support and standardized guidelines for rectifying chaos and strengthening industry self-discipline. Among them, "Requirements for Generative Engine Optimization Services" focuses on standardizing the entire service process, clarifying entity qualifications, implementation norms, behavioral boundaries, and compliance requirements, providing self-disciplinary guidance for service operations; "Requirements for Monitoring and Verifying the Optimization Effects of Generative Engines" focuses on a closed-loop measurement of effects, unifying monitoring dimensions, sampling rules, and verification methods, solving common problems such as inconsistent evaluation criteria and difficulty in verifying effects that have long existed in the industry. "Requirements for Classification, Grading, and Selection of Content Sources for Generative Engine Optimization" focuses on governing the content source, establishing a classification and grading model for content sources and selection rules, and laying a foundation for credibility standards from the information entry point. This standard was drafted led by Frost & Sullivan (hereinafter referred to as "Frost & Sullivan"). These three standards cover the entire chain of service supply, effect verification, and content norms, forming an organic self-disciplinary standard system that supports each other. Zhang Guohua, President of the China Advertising Association stated that the wave of generative artificial intelligence technology is reconstructing the underlying logic of information dissemination, and also brings new development challenges and governance issues for brand marketing. While embracing the technological benefits, the industry must maintain the bottom line of authenticity and compliance, protect consumer rights and industry credibility. Promoting standardization in the field of generative engine optimization is an inevitable choice for the industry to actively adapt to technological changes and improve the self-disciplinary standard system. It is also an important measure for the association to fulfill its self-disciplinary responsibilities and serve the high-quality development of the industry. The introduction of these standards will define a clear development coordinate system for the industry and help brand marketing progress healthily in the AI era. Huo Yan, Secretary General of the China Advertising Association said that the compilation of these three standards emphasizes both problem orientation and systematic thinking. They provide a systematic self-disciplinary reference for service operation, effect evaluation, and content norms in the field of generative engine optimization, further enriching and improving the standard system construction in the AI marketing field that the China Advertising Association is systematically promoting. At the same time, the release of the standards is not the end, but the starting point of the industry's standardized self-discipline norms. The association will take these three standards as the core, continuously collaborate with government departments, market entities, and industry organizations, promote the implementation and iteration of standards, and effectively transform institutional norms into the internal driving force for industry development. Wang Chenhui, Managing Partner and President of Frost & Sullivan Greater China, and the main drafter of "Requirements for Classification, Grading, and Selection of Content Sources for Generative Engine Optimization" said that the content source is the "first gate" for information distribution in the era of generative AI. Without order at the source, any optimization cannot be carried out. Frost & Sullivan is honored to lead the drafting of the requirements for classification, grading, and selection of content sources, transforming global consulting research experience and methodologies into an industry public infrastructure. The next phase of GEO competition is essentially a competition of content source credibility. As a responsible content source, Frost & Sullivan will take the lead in following the standards, integrating standard requirements into its own research and consulting processes; taking the lead in promoting and conveying, helping brands and ecosystem partners understand and make good use of the rules; and continuously providing feedback on practice, working with the association to transform "paper standards" into "industry practices", making credible content the new foundation for brand growth in the AI era. Next, the China Advertising Association will fully promote the implementation of the standards, conduct industry compliance assessments based on the standards, promote the construction of industry public services such as the AI marketing review platform, and continuously track industry practice feedback and technological iteration trends, timely revising and improving standard content to meet the needs of high-quality industry development. The full text of the standards can be accessed on the official website of the China Advertising Association ( https://www.chinacaa.org/cnaa/newsdetail/1331) or through the National Group Standard Information Platform (https://www.ttbz.org.cn/) Please check it out. Event Preview On October 23 (Friday) from 14:00 to 17:45, during the 33rd China International Advertising Festival, the "2026 New Consumption Brand Transformation and Growth Exchange Meeting", co-hosted by the China Advertising Association and Frost & Sullivan, and co-organized by the New Consumption Professional Committee of the China Advertising Association, the Internet Advertising Work Committee of the China Advertising Association, LeadLeo Research Institute, and IAB China, will be held at Beijing Haidian Zhongguancun International Innovation Center. With the theme "Leading the New Consumption Wave, Opening a New Era for Global Branding", this exchange meeting focuses on five core topics: global consumption trend insights, AI-empowered brand marketing, innovation in new consumer categories, practical strategies for global branding, and GEO and brand communication in the new era. Research reports such as SULLIVAN's "Global Product Innovation Index" and "China Leading Index" will be shared on site. Two thematic roundtables and an interpretation session of the "2026 China GEO Marketing Practice and Compliance Development Report" will be held. A GEO map and typical GEO cases will be released, and a ceremony for issuing certificates to participants of the three group standards for GEO will be held. Industry leaders, platform managers, leading brand representatives, investment institutions, and senior experts are expected to share their insights on the growth logic and global expansion path of new consumption brands. Industry partners are welcome to register and attend! Please scan the code to fill in Exclusive QR code for "2026 New Consumption Brand Transformation and Growth Exchange Meeting": Contact Person: Lu Jiabin 18310819186 (same as WeChat number) Email: lvjiabin@china-caa.org

Frost & Sullivan China Global Industry Cooperation Conference On October 15, 2026, the Frost & Sullivan China Global Industry Cooperation Conference, with the theme "Integrating the World, Achieving Joint Success," was held in Nanjing. Wang Chenhui, Managing Partner and President of Frost & Sullivan China, attended the conference. During the main forum in the afternoon, titled "The Governance Capability of Chinese Companies in the AI Era," he delivered a keynote speech titled "Value Perception and Information Governance of Listed Companies in the AI Era: New Challenges and New Paradigms," focusing on the positioning and co-construction path of the IP-KNIT industry public trustworthy knowledge network. Wang Chenhui, Managing Partner and President of Frost & Sullivan China, delivered a keynote speech 01 AI-related risks pose new requirements for information governance Wang Chenhui pointed out that AI's extraction, compression, and reorganization of public information are changing how investors understand listed companies. Key omissions, risk amplification, and outdated information may cause the corporate profile to deviate from reality. Even if the data is accurate, the lack of comparative benchmarks and business context can lead to structural distortions. Companies should integrate investor relations (IR) and public relations (PR) into a unified workflow, and focus on the accuracy, traceability, and updateability of their information within the AI ecosystem. In September 2026, under the guidance of the China Listed Companies Association, the first batch of 5 listed companies underwent testing. Using 29 standardized questions covering 6 AI platforms, 57 distortion issues were identified, including only 5 fabricated cases, 23 structural distortions, accounting for approximately 40%. 29 errors were caused by data conversion, cross-period mismatches, and outdated reporting periods, accounting for about 51%. Relevant standards and service systems are still in the discussion stage. Wang Chenhui, Managing Partner and President of Frost & Sullivan China, delivered a keynote speech at the 2026 China Global Industry Cooperation Conference 02 IP-KNIT connects knowledge entries with industry research Wang Chenhui explained that under Frost & Sullivan's "Knowledge 3.0" research framework, knowledge governance further focuses on whether information can be reliably used after leaving the original text and who is responsible for it. In 2026, at the COMMERCIAL AVIATION, Frost & Sullivan first systematically released the IP-KNIT (Industry Public Knowledge Network of Integrity & Trust) co-construction plan, connecting knowledge, evidence, versions, and responsibilities through industry collaboration. The knowledge in IP-KNIT remains in each participant's original system, establishing connections through common data structures, verification mechanisms, identity and rights rules, and machine invocation interfaces. The public trustworthy knowledge layer is based on a shared ontology, organizing regulatory status, test results, and first-party facts of enterprises into verifiable knowledge entries. The industry research layer includes entries marked with sources, boundaries, and versions, combined with the industrial chain, market benchmarks, and estimation models to form research outcomes supporting decision-making. The two layers are connected by "bidirectional accessibility": each high-impact conclusion in the report can return to the knowledge entry and original evidence, and the entry can also return to the complete report, restoring the industrial context. Thus, the research conclusions retain the source of evidence—"bidirectional accessibility" is the key feature of IP-KNIT compared to ordinary content libraries. The shared ontology defines concepts and their relationships, expressing business meanings in machine-readable standardized language, and is jointly maintained by multiple industry parties. Taking the magnesium thione case mentioned in the speech as an example, application acceptance, subsequent status, and evidence boundaries should be recorded separately. Acceptance and postponement documents cannot support already approved conclusions, and issues in a single batch cannot be directly extrapolated to the entire industry. 03 Verifying the governance loop and public trust through pilot projects IP-KNIT will start with leading industries and collaborate with industry partners. The first 90-day pilot plan will produce 30–50 minimum knowledge entries, 12–15 real frequent issues, a knowledge map, a prototype of industry research, and complete operation records, verifying the closed loop of knowledge from provision, organization, verification, publication to maintenance and correction. When conditions are met, the quality of AI responses will be tested to see if it improves, and its effectiveness in real decisions will ultimately be evaluated. Wang Chenhui emphasized that co-construction must clarify roles and power boundaries. Companies should contribute first-party facts and real issues, universities and independent experts should conduct scientific verification and judgment of evidence boundaries, industry research institutions should handle overall industry aspects, independent evaluators should test response quality when conditions permit, and the secretariat should maintain coordination. Sponsors should not purchase facts or rankings, producers should not independently verify their key conclusions, and operators should not be the only fact adjudicators. Listed companies can start with a knowledge entry, a frequent issue, or a status clarification, organizing product specifications, technical capabilities, certification information, and operational data into verifiable, updatable, comparable, and correctable facts. Based on AI self-checks and optimizing text structures, companies are encouraged to incorporate AI information governance into the ESG framework, participate in the construction of the industry public trustworthy knowledge network, ensuring that key knowledge retains evidence, boundaries, versions, and responsibility even after leaving the original text.

As generative AI rapidly reshapes information retrieval and consumer decision-making processes, the channels through which users obtain information about brands and products are gradually expanding from traditional search engines to generative AI platforms. In this context, Generative Engine Optimization (known as "GEO") is evolving from a standalone content optimization tool into a systematic ecosystem that connects brand awareness monitoring, user intent analysis, content creation and optimization, and omnichannel information dissemination. To systematically organize the GEO ecosystem in China, LeadLeo Research Institute has gathered high-quality technologies, products, services, professional content, data resources, media channels, and practical achievements in the GEO field. This initiative aims to continuously improve the industry classification system and supply-demand matching mechanism. The China Advertising Association has organized the "China GEO Marketing Ecosystem Map 2026" collection project, and Frost & Sullivan provides intellectual support. This collection will primarily base on the core business, resource capabilities, and functional positioning of enterprises or institutions within the GEO ecosystem. It will cover three main categories: large model and AI answer platforms, GEO service and technology providers, and content and information sources. The "China GEO Marketing Ecosystem Map 2026" will be compiled to have authority, professionalism, practicality, and industry guidance value. It will help brands accurately identify the capabilities and applicable scenarios of different ecosystem participants, reduce cooperation screening and trial costs, and promote the standardization, professionalism, and high-quality development of the GEO marketing field in China. 1. Participation Requirements (1) Basic Requirements Participants must be legally registered enterprises/institutions with no negative business records; Their main business should match the specific track of the map, and they should have at least one year of service/operational experience in that track; They must provide real and complete supporting materials related to their main business (ensuring traceable information for map verification); Each enterprise must apply for the corresponding track based on its most core main business. (2) Submission Materials Electronic materials must be submitted, including but not limited to: · Business license/institutional registration proof; · Enterprise/product logo; · Company introduction and core team overview; · Explanation of main business and corresponding niche tracks; · Annual revenue scale, major partners, and main operating areas; · Typical cases, etc. (Note: The above content is required for the selection process; the actual presentation shall follow the map display) 2. Participation Value Authoritative Exposure: The names of enterprises/institutions, brand logos, and core businesses may be included in the "China GEO Marketing Ecosystem Map 2026" according to the corresponding ecosystem type, and promoted through relevant channels. Resource Connection: Selected enterprises or institutions may participate in brand demand matching, special salons, and industry exchange activities, connecting brands and related ecosystem partners. Case Promotion: Excellent products, research results, and typical cases may be promoted as outstanding industry practices. 3. Participation Method (1) Collection Phase September 30 - October 15 (2) Application via QR code: Long-press to recognize the following QR code and fill out online (3) Application Guidelines Each enterprise or institution can choose up to one ecosystem type for application. It is recommended to select the type that best represents its core business or resource capabilities; Submitted materials must be true and valid. Incomplete information or false applications will affect the selection results or disqualify applicants; Submitting an application means the applicant authorizes the China Advertising Association to use the submitted information in the map and related content for industry analysis, content publication, and display; Submission does not guarantee automatic selection. The final classification and inclusion results of the map shall be based on the officially released version. Please submit your materials through the application portal as soon as possible to jointly create the "China GEO Marketing Ecosystem Map 2026", seize new opportunities for brand awareness and growth in the generative AI era, and build a standardized and trustworthy GEO industry ecosystem.

Frost & Sullivan Insight "一小时一个价。"近期，MLCC（多层陶瓷电容器）市场流传着这样的说法。部分高容量MLCC现货价格快速上涨，极端情况下，较此前价格最高涨至数倍甚至十倍；与此同时，普通消费级MLCC却开始降价。这种分化，正在成为这一轮MLCC行情最鲜明的特征。MLCC被称为"电子工业大米"，几乎存在于所有电子产品中，包括手机、家电、新能源汽车、AI服务器等。过去一年，MLCC经历多轮涨价。2025年下半年至2026年上半年，部分产品价格持续上调；进入7月，村田、国巨、三星电机等海外厂商再次调价，部分产品涨幅达到30%甚至更高。 在此背景下，《时代周报》记者就MLCC的价格分化采访弗若斯特沙利文（Frost & Sullivan，以下简称"沙利文"）沙利文中国业务主管合伙人兼董事总经理陆景。 "Time Weekly" Q: How do you expect MLCC prices to be in the coming year? How long will this price increase last, and what key factors determine it? Lu Jing Frost & Sullivan China Business Director Partner and Managing Director (1) Second half of 2026: The pricing logic remains valid, but it has shifted from "full price increases" to "price increases for high-end specifications". From the second half of 2025 to February 2026: Frost & Sullivan reports that global passive components (MLCC, resistors, inductors, tantalum capacitors, magnetic beads, etc.) are entering a period of price increases across all categories, multiple rounds, and regional levels. The increase ranges generally from 5% to 30%. The prices of high-end categories have seen even greater increases. Starting from April 2026: Global Passive Components Frost & Sullivan Frost & Sullivan China LeadLeo Research Institute LeadLeo YUAN CAPITAL SULLIVAN TELE-TREND CLOUD TECHNOLOGY TradeGo MagnaTEC Automotive & Mobility Environmental Protection & Energy Saving Technology Logistics & Supply Chain MATERNAL AND INFANT Education & Training Real Estate & Property Catering & New Retailing Advanced Materials Healthcare & Life Sciences Semiconductor & Chip COMMERCIAL AVIATION Technology, Media and Telecom LANDSCAPING Big Data & AI Infrastructure Construction & Utilities Culture & Entertainment AGRICULTURE, FORESTRY ANIMAL HUSBANDRY AND FISHERY Food & Beverage Fintech SHIPPING AND PORTS Dual Carbon & New Energy Mining & Metals Public Sector Cross-Border E-commerce Trade Building Technology, Construction & Decoration Beauty & Fashion Smart Homes Digital Infrastructure Enterprise Services Consumer Electronics Frost & Sullivan prioritizes ensuring supply for AI and automotive-grade major clients, resulting in "price changes daily" for channel partners. Driven by factors such as channel inventory replenishment, customers locking in orders in advance, and manufacturers tightening low-price orders, price increases in the first half of 2026 have gradually spread to some high-specification and automotive-grade materials, as well as some ordinary consumer-grade products, creating a broad price increase trend. Taiyo Yuden increased the prices of some consumer and automotive-grade MLCCs by approximately 6%–13% in April; after June, the prices of some mainstream X5R consumer MLCCs in the Chinese market rose by about 15%–25%. Since July 2026, passive components have seen a third round of price increases. Starting in July 2026, Murata has increased prices by 10% to 40% for AI and automotive-grade products; Guojue raised its entire range of prices by approximately 50% on July 1, 2026; Samsung Motor sent a notice to customers in July, stating that the shipping price for MLCC products would be uniformly increased by 30% starting from August 1; Sunsei sent a notice about MLCC price adjustments to customers on July 23, announcing price changes for certain products starting from September 1. At the same time, since the second half of the year, the market has gradually shown two distinct trends: On one hand, ordinary consumer-grade MLCCs for traditional terminals such as mobile phones, PCs, and home appliances, after experiencing rapid increases in the early stage, will gradually return to a more rational price level due to channel inventory replenishment, normal customer procurement rhythms, and renewed supply competition. The prices will mainly stabilize at high levels or experience minor fluctuations; on the other hand, High-capacity, small-size, low-voltage, high-temperature performance, and highly reliable MLCCs for AI servers, automotive applications, and industrial uses are still expected to experience shortages and price increases. Order and delivery data have reflected this differentiation: as of the end of June 2026, the order-to-delivery ratios of Murata, Samsung Electro-Mechanics, and Taiyo Yuden were approximately 1.30, 1.31, and 1.25 respectively, significantly higher than the industry average of about 1.04; The delivery times for some high-end X6S products have also increased from about 8 weeks to a maximum of about 20 weeks. Currently, some domestic manufacturers' orders for high-end AI specifications have been scheduled for the second quarter of 2027. Consequently, downstream procurement strategies have changed. In the past, orders were placed 1-2 months in advance, but now, 3-6 months in advance is often used to lock in production capacity. This indicates that the previous price increases for ordinary consumer-grade products mainly reflect price recovery and inventory replenishment, while the rises in high-end MLCCs are driven by actual supply-demand gaps. AI servers continue to increase power density and power management complexity, leading to rapid growth in demand for high capacitance, high stack layers, and similar product types. At the same time, due to material systems, stacking processes, yield improvement, and customer certification, effective supply cannot expand rapidly in the short term, and Japanese and Korean manufacturers will prioritize allocating limited production capacity to high-value-added customers such as servers, automobiles, and industrial applications. Therefore, the key characteristics of the MLCC market in the second half of 2026 will be: ordinary consumer-grade products will return to a rational price level after the rapid increase in the first half of the year, while high-end AI servers and automotive high-end MLCCs will continue to experience shortages and price increases due to tight supply-demand balance. (2) 2027: It is more likely to be "stabilization at high levels + structural price increases," rather than continuous significant general increases In the first half of 2027, the prices of high-end MLCCs are expected to remain resilient, and some scarce product types may continue to rise. On the demand side, AI server platform updates and continuous deployment by cloud service providers will continue to increase the unit value of high-specification MLCCs. For example, the GB200 board requires approximately 6,500 MLCCs, and the Rubin platform is expected to increase this to approximately 12,000 MLCCs, with high-capacity, small-size, low-voltage, and high-temperature product types still benefiting. On the supply side, although Japanese and Korean manufacturers have started production expansion, the new capacity still needs to go through equipment installation, yield improvement, and customer certification, so it will not fully convert into effective supply in the short term, Therefore, the tight supply-demand balance for high-end products is expected to continue. After the second half of 2027, the market is more likely to shift from continuous price increases driven by shortages to structural differentiation under stable prices. On one hand, if actual AI server shipments and the value per MLCC per unit continue to increase, high-end high-capacity MLCCs will still maintain stronger price resilience compared to ordinary consumer-grade products; on the other hand, as the previous production expansion by Japanese and Korean leaders is gradually released and Taiwanese and Chinese manufacturers continue to penetrate mid-to-high-end product types, the supply shortage for some high-end specifications is expected to ease marginally, and the price increase rate will slow down. Meanwhile, ordinary consumer-grade MLCCs, after the rapid increase in the first half of 2026, will still face inventory digestion and production capacity competition if traditional terminal demand does not recover as expected, and prices will depend more on terminal demand and channel inventory changes. Overall, the baseline scenario for 2027 is not a continuous general increase across the industry, but rather high-end high-capacity MLCCs will maintain relatively high prices and strong bargaining power, while ordinary consumer-grade MLCC prices will stabilize or return to their fundamental levels, and the differentiation in industry product structure, customer structure, and manufacturer technical capabilities will further increase. (3) Several core variables determining the duration of price increases MLCC usage per AI server board: determines the slope of high-end demand growth The demand for MLCCs from AI servers is not just due to increased server shipments; more importantly, the number of high-specification MLCCs used per device and their value increase simultaneously. According to TrendForce data, the NVIDIA GB200 board uses approximately 6,500 MLCCs, and the next-generation Rubin platform will use about 12,000 MLCCs, an increase of approximately 85%; during the verification process for AMD MI450, the number of MLCCs per board for some 47μF, 2.5V, X6S, 0402 specifications increased from 1,440 to 10,544 pieces, a increase of 632%. This change indicates that AI server upgrades involve not only adding general electronic components but also, due to increased power density of GPUs, ASICs, HBM, and high-speed network modules, and greater current transient changes, the power distribution network requires more high-capacity, low-ESR, low-voltage, and high-temperature performance MLCCs to achieve more stable decoupling, filtering, and transient response. Since the new demand is concentrated in a few high-specification product types, it actually results in a sharp increase in demand for specific product combinations, rather than an average increase for all MLCCs. As the single-board MLCC usage on the new AI platform continues to increase, even if server shipment growth slows, the unit value of high-end MLCCs can still increase, continuously occupying the original manufacturers' high-end production capacity; conversely, if the platform design becomes more mature, the growth in usage per unit slows, or server deployment pace is lower than expected, the supply-demand tension will ease quickly. Japanese and Korean manufacturers' speed of releasing high-end capacity and the progress of domestic substitution: determine whether the supply gap can be filled The supply constraints for high-end MLCCs do not necessarily mean insufficient total industry capacity, but rather that the effective production capacity to steadily produce high-capacity, small-size, high-stack layers, and high-reliability products is limited. Currently, Japanese and Korean manufacturers such as Murata (Japan's Murata), Samsung Electro-Mechanics (South Korea's Samsung Electro-Mechanics), and Taiyo Yuden (Japan's Taiyo Yuden) remain the main suppliers of high-end MLCCs. Their new capacity still needs to go through equipment installation, process improvement, yield enhancement, and customer certification, so it will not quickly convert into effective supply for AI servers and automotive customers in the short term. At the same time, domestic substitution will become an important factor affecting the duration of this price increase. Manufacturers in Mainland China already have strong supply capabilities in general consumer-grade and some mid-range MLCC areas, enabling them to quickly alleviate the shortage of general specifications. Therefore, general consumer-grade products are more likely to return to a reasonable level after the initial price increase. However, in terms of high capacitance, small size, ultra-thin stacked structures, high-temperature characteristics, and high-reliability materials, there are still gaps between domestic manufacturers and leading Japanese and Korean companies in product performance, yield stability, and customer certification timelines. In particular, AI servers and automotive customers usually do not easily switch suppliers due to short-term shortages. Even if domestic manufacturers have the product capabilities, they still need to go through a longer verification and integration process. Thus, in the short term, domestic substitution is more likely to intensify price competition for general and mid-range MLCCs rather than immediately eliminating the supply-demand gap for high-end MLCCs. However, if domestic manufacturers make breakthroughs in yield, reliability, and customer certification for high-specification materials, it will significantly accelerate the release of high-end supply and become a key factor in narrowing the price increase for high-end MLCCs after 2027. Speed of AI server platform iteration Demand for AI servers does not grow uniformly but is highly dependent on generational changes in GPU, ASIC, and server architecture. Each platform iteration, such as from GB200 to Rubin, may lead to a stepwise increase in MLCC usage per unit and the proportion of high-specification materials due to increased computing density, complex power supply architectures, and higher board integration levels. In addition to NVIDIA, AMD and cloud service providers such as Google, AWS, and Meta's self-developed ASICs will also create new demand sources. If multiple platforms increase production simultaneously during the same period, original manufacturers will face greater pressure to allocate high-end capacity. Therefore, the speed of platform iteration determines not only the scale of MLCC demand but also whether demand growth occurs in a concentrated manner. If Rubin and subsequent platforms increase production as expected, and CSP's self-developed ASICs continue to be deployed, demand for high-end MLCCs will continue to exceed the original supply plan, supporting shortages and continued price increases. Conversely, if the introduction of new platforms is delayed, AI capital expenditure decreases, or other critical components such as chips, HBM, CPU, and power modules become bottlenecks in server delivery, MLCC demand will delay, and the price increase cycle will shorten. Demand for Consumer Electronics and Automotive Electronics Phones, PCs, tablets, and home appliances remain the largest application base for MLCCs. Demand for consumer electronics determines the overall industry utilization rate and the price floor for general specifications. In the first half of 2026, general consumer-grade MLCCs experienced rapid increases driven by original manufacturers' price adjustments, channel inventory replenishment, and market sentiment. However, in the second half of the year, if terminal demand such as phones and PCs remains weak, after channels complete inventory replenishment, general specifications will face inventory digestion and capacity competition again, and prices will tend to be reasonable. Conversely, if consumer electronics demand shows a significant recovery in 2027, it will directly absorb general capacity, improve the utilization rate of original manufacturers, and reduce the motivation for original manufacturers to offer discounts to low-end customers, thus providing stronger price support for the entire MLCC market. Automotive electronics has a different role from consumer electronics: its demand growth may not be a direct catalyst for rapid short-term price increases. However, automotive MLCC certification cycles are long, lifecycles are extensive, and supply chain stability requirements are high, continuously occupying mid-to-high-end capacity and raising the industry price level. If new energy vehicles, advanced driver assistance systems, and intelligent cabins continue to increase the electronic complexity per vehicle, demand for high-reliability automotive MLCCs will compete with AI server demand for the high-end capacity of Japanese and Korean manufacturers, further strengthening supply constraints for high-end products. Q: AI servers are increasing MLCC usage. What is the impact of this on overall demand? After high-end MLCC capacity is directed towards AI servers, how much pressure will it put on traditional markets such as consumer electronics and automotive electronics? Lu Jing Partner and Managing Director, Frost & Sullivan Frost & Sullivan China (1) AI servers are currently not the largest downstream market for MLCCs, but they have become one of the most important structural growth areas, and their impact on high-end MLCC demand is significantly higher than that of traditional servers The impact of AI servers on MLCC demand mainly comes from two aspects: increased usage per unit and product specification upgrades. Compared with traditional general servers, AI servers are equipped with more GPUs, ASICs, and high-power power management modules, resulting in significantly increased overall power consumption and transient current, thereby increasing the MLCC requirements for decoupling, filtering, voltage stabilization, and energy storage. AI server MLCC usage is approximately 8–13 times that of traditional general servers (For example, according to Samsung Electro-Mechanics' official data, a general server uses about 2,200 MLCCs per unit, while an AI server uses approximately 28,000 MLCCs per unit). As GPUs and ASICs continue to evolve toward higher computing power and higher power consumption, MLCC demand also shows a trend of upgrading from quantity to high capacity, miniaturization, and high reliability. Leading manufacturers are continuously expanding their high-capacity MLCC products for AI servers: In September 2026, the global leader Murtaka announced the suspension of production of some consumer-grade and automotive conventional models, prioritizing capacity allocation to high-value-added AI high-end products. Additionally, the company announced a planned additional investment of 80 billion yen for MLCC expansion Frost & Sullivan, Frost & Sullivan China, LeadLeo, LeadLeo Research Institute, LeadLeo, YUAN CAPITAL, SULLIVAN TELE-TREND CLOUD TECHNOLOGY, TradeGo, MagnaTEC, Automotive & Mobility, Environmental Protection & Energy Saving Technology, Logistics & Supply Chain, MATERNAL AND INFANT, Education & Training, Real Estate & Property, Catering & New Retailing, Advanced Materials, Healthcare & Life Sciences, Semiconductor & Chip, COMMERCIAL AVIATION, Technology, Media and Telecom, LANDSCAPING, Big Data & AI, Infrastructure Construction & Utilities, Culture & Entertainment, AGRICULTURE, FORESTRY ANIMAL HUSBANDRY AND FISHERY, Food & Beverage, Fintech, SHIPPING AND PORTS, Dual Carbon & New Energy, Mining & Metals, Public Sector, Cross-Border E-commerce Trade, Building Technology, Construction & Decoration, Beauty & Fashion, Smart Homes, Digital Infrastructure, Enterprise Services, Consumer Electronics, focusing on the AI server sector, aiming to increase production capacity by approximately 20% by fiscal year 2027, and the first phase of new production capacity is expected to begin operation in the fourth quarter of 2027. Samsung Electro-Mechanics signed a long-term supply contract for AI server MLCC worth trillions of yen in September. In terms of demand growth, AI servers have become one of the fastest-growing downstream applications for MLCCs. According to public information, in 2026, the demand for AI server MLCC is approximately 72.6 billion pieces, representing a 87% increase compared to previous years. By 2027, the demand is expected to reach around 136.7 billion pieces, showing an additional 88% growth. Although AI servers have not yet replaced traditional markets such as mobile phones, PCs, and automobiles in terms of overall MLCC shipments, they have become an important growth direction for high-end MLCC demand and new production capacity allocation. (2) The impact on consumer electronics and automotive electronics is not entirely the same Consumer Electronics: The pressure is relatively evident Traditional consumer electronics such as mobile phones and PCs have not seen comparable demand growth for AI servers. Therefore, the recent shortage of consumer-grade MLCC supply is mainly due to reallocation of production capacity from the supply side, rather than a significant expansion in end-demand. Since some mid-to-high capacity consumer X5R products compete with high-end products in terms of production equipment and manufacturing resources, after Korean and Japanese manufacturers increased their production capacity for AI-related products, the effective supply flexibility of consumer-grade MLCC decreased, and some orders were shifted to manufacturers in Taiwan and mainland China. Price changes further indicate that capacity allocation adjustments have been transmitted from high-end products to consumer markets. In August 2026, TrendForce predicted that Samsung Electro-Mechanics' consumer X5R products for OEM and ODM customers in the fourth quarter would see an average price increase of about 25%–30%, while the high-end X6S products used in AI servers were expected to see a price increase of about 10%–20%. One of the main reasons for raising prices for consumer specifications is to control related order demands and release capacity for expanding high-end product lines. Thus, the impact on the consumer electronics market mainly manifests as a reduction in supply for some mid-to-high capacity products, lower inventory levels, and higher prices, rather than a general shortage of all consumer-grade MLCCs. Automotive Electronics: The short-term impact is relatively limited, but high-end automotive MLCCs also face competition in capacity and supply resources Automotive MLCCs have high technical and certification requirements. They need to operate stably for long periods in complex environments such as high temperature, low temperature, high humidity, and vibration, and must pass AEC-Q200 and customer verification. Samsung Electro-Mechanics revealed that the development cycle for automotive MLCCs is about three times longer than that of ordinary products, and their prices are more than three times higher. Therefore, automotive products that have passed customer certification and established stable supply relationships have relatively rigid capacity allocations, and will not shift quickly and大规模 due to increased AI demand like ordinary consumer products. However, both AI servers and automotive electronics are important growth markets for high-end MLCCs. There is still some competition in high-capacity, high-reliability products, as well as advanced equipment, materials, and manufacturing resources. According to TrendForce data, given the market's expectation of tight supply in the second half of 2026, some automotive ODMs have delayed the planned MLCC procurement scheduled for July to May, indicating that the automotive industry chain has also started increasing inventory levels to cope with potential supply pressures. At the same time, the continuous improvement of new energy vehicles, ADAS, and automotive electronics levels is also increasing the demand for high-performance MLCCs. Therefore, the impact on the automotive market from AI servers is relatively weaker than that on consumer electronics. However, since the demand for AI servers and automotive electronics is growing simultaneously, it will jointly increase the capacity utilization rate of high-end MLCCs and strengthen the supply-demand tension for high-spec products. Q: What is the competitiveness of domestic manufacturers such as Fenghua High-tech and Sanhuan Group in high-end MLCCs and MLCCs for servers? If we refer to Changxin Storage's path of domestic substitution, does the MLCC industry have similar industrial conditions? Where is the real gap? Lu Jing Partner and Managing Director, Frost & Sullivan Frost & Sullivan China (1) Domestic manufacturers have moved from "low-end substitution" to "mid-to-high-end breakthrough," but there is still a significant gap compared to leading Japanese and Korean companies AI computing infrastructure is being built at an accelerated pace, simultaneously increasing the usage and performance requirements for MLCCs. AI servers require multiple high-power GPUs, with a single module using hundreds of MLCCs. For example, the MLCC usage in Nvidia's 8-GPU Hopper HGX server is about 7–8 times that of traditional servers, and the usage increase for the GB200 NVL72 rack solution can be more than 100 times. At the same time, system power consumption and integration continue to increase, putting higher demands on the capacity density, high-temperature resistance, and high reliability of MLCCs. This drives products to upgrade towards higher capacity, smaller size, and higher reliability, creating new opportunities for domestic manufacturers with relevant technologies and mass production capabilities to break through to mid-to-high-end markets from general products. Sanhuan Group: Sanhuan's advantages come more from its long-term accumulated material technology, ceramic forming and sintering processes, as well as large-scale manufacturing capabilities. The company's MLCCs have achieved 01005 to 2220 (Imperial) Full range of packaging mass production, covering miniature, high-capacity, high-reliability, high-voltage, and high-frequency series . In 2023, Sanhuan's stacking breakthrough 1,000 layers Daguan has entered the world's top tier; By 2025, the dielectric layer will be further reduced from approximately 5μm to less than 1μm (1 micron, equivalent to four tens of the diameter of a hair strand), and mass production will be achieved successfully. This indicates that its ability to produce thin layers and high stacking capabilities is continuously improving. For high-frequency applications such as optical communication and servers, the company's high-frequency low-resistance MLCC products have reached industry-leading performance levels, and it has deeply integrated into the supply chain of AI servers and high-speed optical modules. The company's 2025 annual report clearly states that it has introduced multiple size options of high-capacity MLCCs, as well as high-capacity MLCCs for 48V power systems. The supply capacity of data center products continues to increase; vehicle-grade MLCCs now cover sizes from 0201 to 2220, and from 0.1pF to 47μF, and are gradually being applied in new energy vehicle power supplies and electric drives, having passed certifications from several benchmark clients. In terms of production capacity, the two major bases established by Sanhuan in Nanchong and Deyang are continuously expanding the production capacity of high-capacity MLCCs. Its production base in Thailand is being developed steadily. After listing on the Hong Kong Stock Exchange in July this year, Sanhuan's raised funds will be used for overseas production capacity expansion, automation upgrades, and material innovation, further strengthening its global supply capacity. At the same time, the business that complements MLCC is optical communication component business . With the increased demand for 800G and 1.6T optical modules, the demand for core components such as ceramic inserts and sleeves is also growing simultaneously. Sanhuan Group is "the world's leading supplier of ceramic inserts and sleeves (with a market share of approximately 70% in 2025)" and possesses significant technical and scale advantages, directly benefiting from the surge in optical interconnection needs brought about by AI computing infrastructure development. Furthermore, as one of the few domestic manufacturers that has completed the entire chain from upstream ceramic powder to downstream components , the company has mastered the core technologies of ceramic forming and sintering, eliminating the constraints of imported materials. Fenghua High-tech: Fenghua's advantages lie in its industrial accumulation, large-scale production, and complete product system, and it is currently upgrading to high-end MLCC. In the first half of 2026, the company's high-end MLCC revenue, including high capacitance, automotive-grade, and medium-high voltage products, accounted for approximately 40% of the total MLCC revenue. The company also stated that related AI server products are being gradually certified by customers. In terms of products and processes, Fenghua has mastered technologies such as homogenization and dispersion of barium titanate powder below 100nm, superthin film casting, and precision layered manufacturing of more than 1,000 layers. It has also developed a sintering process system for extreme high-capacity MLCCs, and has launched 0805 size, 100μF, 2.5V, X6S products, indicating that its high-end development has progressed from technical reserves to product implementation and commercialization stages. Therefore, Fenghua is no longer just a mid-to-low-end domestic replacement enterprise. However, the customer certification and scale expansion of high-end MLCCs for AI servers are still in the ongoing progress stage. (2) MLCCs have certain conditions similar to those of "Changxin Storage" for domestic replacement, but the industrial logic is not entirely the same. Advantages China has the world's largest downstream demand market. China is a global important manufacturing base for Consumer Electronics, new energy vehicles, communication equipment, and servers. It has a large number of complete machine manufacturers, automotive companies, server ODMs, power supply manufacturers, and electronic manufacturing service providers, providing rich product validation and customer introduction scenarios for domestic MLCCs. As the importance of supply chain security increases, downstream customers are actively introducing domestic suppliers, creating conditions for domestic MLCCs to penetrate into mid-to-high-end applications such as automotive-grade and servers. The upstream industry chain is gradually improving. MLCC manufacturing involves multiple steps, including dielectric powder, internal electrode slurry, release film, casting, printing, layered manufacturing, sintering, and sealing. In the past, high-end materials and equipment mainly relied on Japanese and Korean companies. Currently, the domestic industry chain has begun to achieve collaborative breakthroughs: Guoci Material's MLCC dielectric powder customers include Samsung Motor, Guju, and Fenghua High-tech, and it is expanding the production capacity of automotive-grade and AI server MLCC powders; Boqian New Material has also established the ability to supply nano- and sub-micron nickel powder. Its 2025 annual report shows that AI servers and automotive electronics are driving the upgrade of MLCC internal electrode materials to smaller particle sizes, higher sphericity, and better dispersion. This is somewhat similar to Changxin's approach of completing product iteration and large-scale production through collaboration with domestic equipment, materials, and downstream customers. Changxin has gradually expanded from DDR4 and LPDDR4X to DDR5 and LPDDR5/5X, and continues to promote the development and verification of the domestic supply chain, indicating that the追赶 path of "domestic demand - customer validation - large-scale production - industry chain collaboration" is feasible. Domestic manufacturers already have a foundation for large-scale manufacturing. Domestic manufacturers such as Sanhuan, Fenghua, and Weirong did not start from scratch; they have already achieved mass production capabilities for general MLCCs and some high-capacity and automotive-grade products. These companies have a complete iteration foundation from material and process optimization, product development, to customer validation, providing conditions for further breaking through high-end specifications for AI servers. The logic of domestic replacement for MLCCs is not entirely the same as that of DRAM. The technical routes and interface standards for DRAM products are relatively concentrated. Leading manufacturers can transform technological progress into market share quickly through process platform upgrades, capacity expansion, and increased production of standard products. In contrast, MLCC features a large number of part numbers, complex specification combinations, and diverse application scenarios, with different customers having varying requirements for size, capacity, rated voltage, temperature characteristics, DC bias voltage, and reliability. Therefore, MLCC is more likely to be gradually replaced along a path of "single part number breakthrough - customer certification - small-scale supply - expansion of adjacent specifications," making the overall process more fragmented compared to storage chips with higher standardization levels. (3) Core Gap The biggest gap is not simply the presence or absence of equipment, but the comprehensive capabilities of materials, processes, and yield rates. The main challenges of MLCC are ultra-thin dielectric layers, multi-layer stacking, high-precision printing, sintering control, and long-term reliability. Especially as MLCC develops towards smaller sizes and higher capacitance values, the dielectric layer becomes thinner, and the number of layers increases, requiring higher demands for material purity, particle size control, slurry quality, printing accuracy, sintering process, and yield rates. Therefore, mid-to-low-end MLCC production lines cannot directly upgrade to high-end AI server production lines by simply adding equipment. Domestic manufacturers have already made breakthroughs in reducing the dielectric layer thickness and increasing the number of layers, but further progress toward high yield rates, low failure rates, and stable mass production across batches still requires continuous process development. Customer certification for high-end products is the second significant barrier MLCC used in AI servers and automotive electronics typically needs to operate for long periods under high-temperature, high-load, and frequent current fluctuations conditions, requiring high effective capacity at actual operating voltage, temperature stability, failure rate, and batch consistency. After product development is completed, it must also go through component selection, board-level testing, system validation, and long-term reliability testing, before gradually entering the procurement systems of server ODMs, power supply manufacturers, or end customers. Thus, the fact that domestic manufacturers have launched high-capacity, automotive-grade, and server products does not mean they have entered the core supply chain and achieved substantial revenue. There is still a significant time gap between product development, certification, small-scale supply, and large-scale production. The top-tier products still have a technological gap compared to international leaders Japanese and Korean leaders can already produce 0402 (small size, using sub-micron dielectric layer *This interview was published in The Times Weekly. The author is Zhu Chengcheng. The original title is: MLCC "Price Changes Every Hour": High-end Product Capacity Under Competition, Do Domestic Manufacturers Seize the "Changxin Moment"?

导语 当AI Agent逐渐深入操作系统、走向跨应用操作，商业生态、数据安全与产业治理随之面临结构性风险，治理规则需要同步跟进。针对这一趋势，弗若斯特沙利文在《2026年侵入式Agent产业治理白皮书》中系统分析了跨应用执行带来的生态影响，并提出以双重授权和全链路可审计为治理基础，逐步形成以API协同为主、GUI模拟为辅的跨应用协作机制。提升执行效率的同时，更应保障用户知情、平台可控、责任可追溯。智能体产业能否形成长期价值，取决于效率与信任能否保持平衡。 AI Agents are evolving from information interaction toward task execution, with cross-application operations emerging as an important direction of product development. However, as Agents gain deeper access to operating systems and applications, risks associated with insufficient authorization and unclear accountability are also increasing, requiring governance rules to evolve in parallel.   2026年上半年以来，Agent产业密集演进，在应用生态内、应用与应用、应用与终端、终端与终端均涌现多款基于接口或协议的交互方案。这表明，白皮书所揭示的产业损害逻辑——流量价值迁移、开发成本攀升、安全风险集中——正被越来越多的行业参与者所验证和认知。Agent权限边界的合规压力也在推动产业回归可治理、可追责的轨道上。 Against this backdrop, Frost & Sullivan’s White Paper on the Governance of Invasive Agents 2026 systematically examines the ecosystem implications of cross-application execution. It proposes a governance framework built on dual authorization and end-to-end auditability, with API-based collaboration as the primary approach and GUI simulation as a supplementary mechanism. The objective is to improve execution efficiency while ensuring that users remain informed, platforms retain control, and accountability remains traceable. The Agent industry’s ability to create long-term value will depend on maintaining a sustainable balance between efficiency and trust.   但与此同时，应用授权不充分、责任边界模糊、高危权限滥用等侵害产业的潜在问题依然存在。白皮书指出，随着全球AI竞争加速，治理能力和信任基础正成为企业进入国际市场的重要竞争条件。若以牺牲信任换取短期扩张，不仅会压缩中国AI企业的国际合作空间，也可能损害中国AI产业的长期国际信誉。Agent技术的演进方向，不应以侵入替代协同，而应在可信治理的框架内，实现从“能做什么”到“该做什么”的产业共识收敛。 As global competition in AI accelerates, governance capabilities and a strong foundation of trust are becoming increasingly important for companies seeking to enter international markets. Pursuing short-term expansion at the expense of trust could not only constrain international cooperation opportunities for Chinese AI companies, but also cause lasting damage to the global reputation of China’s AI industry.   根据沙利文《2026年侵入式Agent产业治理白皮书》，AI Agent加速进入系统层，跨应用执行正在重塑移动互联网入口与生态治理规则 Frost & Sullivan’s 2026 White Paper on the Governance of Invasive Agents: AI Agents Move Deeper into the System Layer, Reshaping Mobile Entry Points and Ecosystem Governance   随着多模态大模型从概念验证步入商业化落地，AI智能体（AI Agent）正逐渐深度介入设备底层操作，并展现出独立执行跨应用复杂任务的能力。在此背景下，部分绕过标准应用程序接口（API）、直接利用系统底层特权干预应用运行的“侵入式Agent”迅速蔓延，给现有的互联网信任机制与生态协同带来了空前挑战。 As multimodal large language models move from proof-of-concept to commercial deployment, AI Agents are increasingly becoming deeply involved in low-level device operations and are demonstrating the ability to independently execute complex tasks across applications. Against this backdrop, invasive Agents that bypass standard application programming interfaces (APIs) and directly use low-level system privileges to interfere with application operations are spreading rapidly, creating unprecedented challenges for existing internet trust mechanisms and ecosystem collaboration.   基于对全球及中国AI Agent市场的系统调研，弗若斯特沙利文（Frost & Sullivan，以下简称“沙利文”）于2026年3月正式发布《2026年侵入式Agent产业治理白皮书》。本报告聚焦侵入式Agent机制对产业流量分配、商业生态运作及底层数据安全造成的冲击，并对行业未来的规范化落地与治理路径进行了前瞻性研判。 In March 2026, based on systematic research into the global and Chinese AI Agent markets, Frost & Sullivan has officially released the White Paper on the Governance of Invasive Agents 2026. The white paper focuses on the impact of invasive Agent mechanisms on industrial traffic allocation, commercial ecosystem operations, and underlying data security, while providing forward-looking analysis of the industry’s future standardization and governance pathways.   01 侵入式Agent依托系统层权限突破应用边界，为产业治理带来新风险 Invasive agents breach application boundaries through system-level privileges, introducing a new layer of governance risk 侵入式Agent的关键特征，在于其不依靠标准接口或协议，借助系统签名级权限直接读取界面信息、识别页面内容，并模拟用户完成点击、输入、跳转等操作。借助这一路径，Agent可以在未获得应用方业务授权的情况下，跨越多个软件连续执行任务。相较于基于标准接口开展协同的路径，这种方式虽然降低了跨应用联动对生态合作的依赖，却也突破了原有应用边界与权限边界，将风险前置到系统层。 The defining feature of invasive agents is that they do not depend on standard interfaces or protocols exposed by third-party applications. Instead, they rely on high-level system privileges, including signature-level permissions, to directly read interface content, interpret on-screen information, and simulate user actions such as clicking, inputting text, and navigating between pages. Through this approach, agents are able to execute continuous tasks across multiple applications without obtaining business authorization from the applications involved. Compared with models built on standardized interface-based collaboration, this pathway reduces dependence on ecosystem cooperation for cross-application orchestration, but at the same time pushes risk upward into the system layer by breaking through established application and permission boundaries.   过去，用户需要先进入具体应用，再逐步完成信息获取与操作决策；如今，系统层Agent开始承接用户意图理解、任务拆解和动作调度，应用则更多退居为任务执行的承载界面。随着移动端原有以应用为中心的入口逻辑被改写，系统层Agent在提升任务执行便利性的同时，也开始对既有生态秩序、权限规则与治理边界形成新的冲击。 Historically, users entered individual applications and completed information discovery and decision-making step by step within each app environment. Today, system-level agents are increasingly taking over intent understanding, task decomposition, and action scheduling, while applications are being relegated to execution interfaces. As this shift rewrites the app-centric entry logic of the mobile internet, system-level agents are improving task convenience while simultaneously placing new pressure on existing ecosystem order, permission rules, and governance boundaries. 侵入式Agent的定义 Defining Features of Invasive AI Agents   02 侵入式Agent前置用户决策环节，第三方应用商业价值承压 By moving user decision-making upstream, invasive agents place pressure on the commercial value of third-party applications 当侵入式Agent逐步成为用户发起任务的主要入口，传统移动应用在生态中的角色也随之发生变化。原本由应用自身承接的搜索、浏览、比较、点击和下单等行为，被越来越多地前置到Agent侧完成，应用留给用户的直接交互时长和触达机会显著减少。 As invasive agents increasingly become the primary entry point for users to initiate tasks, the role of conventional mobile applications within the ecosystem is also changing. Activities that were previously carried out within apps, including search, browsing, comparison, clicking, and purchasing, are increasingly being completed on the agent side before users ever meaningfully interact with the application itself. As a result, applications face a sharp reduction in direct engagement time and user touchpoints.   这一变化将直接传导至应用开发者的商业模式。工具类应用首当其冲，交易类与社交类的生态价值也将被冲击。无论是依赖停留时长和曝光分发的信息流广告，依赖深度使用形成粘性的会员订阅，还是依赖交易链路闭环获得收益的佣金模式，都会受到不同程度影响。 This shift is likely to have direct consequences for application developers’ monetization models. Utility applications are expected to come under the greatest pressure first, but transactional and social/content platforms will also face material disruption. Whether monetization depends on information-flow advertising tied to user time spent and exposure allocation, subscription models built on deep engagement, or commission-based models supported by closed-loop transaction paths, each of these commercial mechanisms may be weakened to varying degrees.   白皮书测算显示，若未来侵入式Agent在用户侧渗透率达到25%，工具类应用商业价值预计下降39%，内容与社交类应用预计下降19.5%，交易类应用预计下降15.4%。这意味着，侵入式Agent带来的并非单点效率优化，而是应用生态商业价值的重新分配。 Frost & Sullivan estimates that if invasive agents reach a 25% user-side penetration rate in the future, the commercial value of utility applications could decline by 39%, while content and social applications could see a 19.5% decline and transactional applications a 15.4% decline. This suggests that the rise of invasive agents is not merely about marginal efficiency enhancement. It represents a broader redistribution of commercial value across the application ecosystem. 用户触达流量入口转移 Shift in User Traffic Entry Points   03 流量迁移未带来显著增量，反而导致产业内卷与高治理成本 Traffic migration does not generate meaningful incremental value, but instead intensifies internal competition and drives up coordination and governance costs across the industry 白皮书进一步指出，侵入式Agent带来的影响，并不主要体现为产业新增量，而更多体现为对现有流量分配关系的重新切分。当系统层Agent试图绕过既有合作机制直接承接用户需求时，平台方、应用开发者与Agent提供方之间的关系会由协同合作转向零和博弈。 The white paper further notes that the impact of invasive agents is not primarily reflected in the creation of substantial new market value. Rather, it is more accurately understood as a redivision of existing traffic allocation relationships. When system-level agents attempt to bypass established cooperation mechanisms and directly intermediate user demand, the relationship among platform operators, application developers, and agent providers may shift from coordinated collaboration to a zero-sum game. 侵入式Agent内卷式工具特征 Invasive AI Agents and Zero-Sum Competition 在这一过程中，应用开发者往往需要通过界面调整、反自动化识别、权限收紧和策略更新等方式进行防御，以应对未经许可的数据读取和界面操作。这使得软件迭代频率、兼容维护难度以及安全防护投入同步抬升。 In this process, application developers are often compelled to respond through interface adjustments, anti-automation detection, tighter permission controls, and ongoing strategy updates in order to defend against unauthorized data extraction and interface manipulation. This raises software iteration frequency, increases compatibility-maintenance complexity, and drives additional spending on security protection.   研究显示，在侵入式Agent渗透率达到25%的情况下，移动应用综合开发成本预计上升16%，包括合作协调、合规审查与安全防御在内的产业链综合治理成本预计上升34.4%。这一趋势说明，若缺乏清晰规则约束，侵入式发展路径很可能将产业资源更多消耗在对抗之中，而非创新本身。 Research indicates that at a 25% penetration rate for invasive agents, overall mobile-application development costs could increase by 16%, while total ecosystem governance costs, including coordination, compliance review, and security defense, could rise by 34.4%. The implication is clear. Without an explicit governance framework, an invasive development path is likely to consume more industry resources in defensive confrontation than in genuine innovation.   04 高权限集中叠加指令诱导，显著放大数据与资产安全风险 The concentration of high-level permissions, combined with instruction-based manipulation, significantly amplifies data and asset-security risks 为实现跨应用、跨场景的连续任务执行，侵入式Agent通常需要长期持有较高等级的系统权限，并在多个业务场景中维持账户登录与操作能力。这意味着，原本分散在不同应用中的数据边界和权限边界，开始向单一Agent集中。一旦出现权限滥用、模型误判或安全缺口，影响范围将不再局限于单个应用，而可能扩散至整个终端环境。 To enable continuous execution across applications and use cases, invasive agents typically require long-term access to elevated system permissions and the ability to maintain logged-in operational states across multiple business scenarios. This means that data boundaries and permission boundaries that were previously distributed across different applications begin to converge into a single agent. Once permission abuse, model misjudgment, or security vulnerabilities occur, the consequences are no longer confined to a single application, but may spread across the entire terminal environment.   与此同时，由于侵入式Agent具备读取屏幕内容并触发后续动作的能力，其还可能受到外部信息的诱导。攻击者可以通过网页、邮件、文档等载体嵌入特定指令，诱导Agent在用户未充分察觉的情况下执行删除、转发、修改、支付等敏感操作。原本局部、可控的识别偏差，在高权限执行环境下可能迅速演变为隐私泄露、账户风险甚至资产损失事件。由此可见，侵入式Agent带来的已不只是传统意义上的信息安全问题，更是权限集中背景下的系统性风险放大。 At the same time, since invasive agents can read screen content and trigger subsequent actions, they may also be inherently susceptible to indirect prompt injection. Attackers may embed specific prompts or instructions in webpages, emails, or documents, inducing the agent to perform sensitive actions such as deletion, forwarding, modification, or payment without the user’s full awareness. In a high-permission execution environment, what might otherwise have remained a localized and manageable recognition error can rapidly escalate into privacy leakage, account compromise, or even financial loss. In this sense, invasive agents do not merely create conventional information-security concerns. They materially increase systemic risk under conditions of concentrated authority. 零点击邮件指令触发网盘批量误删 Zero-Click Email Injection Triggers Bulk Cloud-Drive Deletion   05 以双重授权和全链路可审计为基础，构建可信治理框架 A trusted governance framework should be built on dual authorization and full-chain auditability 针对上述问题，白皮书提出，Agent产业的可持续发展不应建立在越过生态边界和削弱信任机制的基础上，而应推动形成以API协同为主、GUI模拟为辅的治理框架。其核心在于建立双重授权机制，即Agent开展跨应用操作，不仅需要获得用户对系统权限的明确授权，也需要获得被调用应用或服务方对具体业务动作的许可。 In response to these risks, the white paper argues that the sustainable development of the agent industry should not be built on bypassing ecosystem boundaries or weakening trust mechanisms. Instead, the industry should move toward a governance framework in which API-based collaboration remains the primary model and GUI simulation plays only a supplementary role. The core principle is the establishment of a dual-authorization mechanism. This means that for an agent to perform cross-application actions, it must not only obtain the user’s explicit authorization for system-level permissions, but also secure permission from the invoked application or service provider for the underlying business action itself.   在具体实施层面，可信治理框架应围绕四个方向展开。其一，清晰界定Agent的权限边界与可代办事项范围，防止能力外溢。其二，对涉及隐私、支付、资产和身份变更等高风险操作设置更严格的动作约束与确认机制。其三，建立覆盖授权、决策、执行与结果反馈的全过程留痕体系，确保关键操作可核验、可复盘。其四，在可审计基础上进一步明确责任归属，为后续争议处理、损失核定和制度约束提供依据。 At the implementation level, the report proposes that a trusted governance framework should advance along four directions. First, agent permission boundaries and the scope of delegable tasks should be clearly defined in order to prevent capability spillover. Second, stricter constraints and confirmation mechanisms should be imposed on high-risk actions involving privacy, payments, assets, and identity changes. Third, a full-process traceability system should be established across authorization, decision-making, execution, and result feedback so that critical actions can be verified and reconstructed. Fourth, on the basis of auditability, responsibility allocation should be clarified to support future dispute resolution, loss assessment, and institutional enforcement.   白皮书最终强调，Agent技术的演进方向不应是以侵入替代协同，而应是在可验证、可约束、可追责的框架内实现跨主体协作。只有在保障生态秩序、商业公平与用户安全的前提下，Agent才能真正释放其对社会效率提升的长期价值。 The white paper ultimately emphasizes that the future direction of agent development should not be one in which invasion replaces collaboration, but one in which cross-entity coordination takes place within a framework that is verifiable, enforceable, and accountable. Only by safeguarding ecosystem order, commercial fairness, and user security can agents fully deliver their long-term value in improving social and economic efficiency.   06 全球竞争不能以透支信任为代价，可信治理将成为中国AI走向国际市场的前提 Global AI competition cannot be pursued at the expense of trust; trusted governance will be a prerequisite for China’s AI industry to expand internationally 白皮书进一步指出，不能片面以创新视角理解侵入式Agent，而应考虑其对产业生态、公众隐私、安全环境、国际竞争的综合影响。Agent并不只是模型能力、产品形态和落地速度的竞争，更是治理能力、信任基础与规则适配能力的竞争。若侵入式Agent以突破权限底线、削弱授权机制、牺牲用户信任为代价，看似抢占了竞争先机，实则可能为整个产业与社会发展埋下更大风险。 The white paper further argues that invasive agents should not be viewed solely through the lens of innovation. Their broader implications for industrial ecosystems, public privacy, security conditions, and international competition must also be taken into account. Competition in AI agents is not only a matter of model capability, product form, and deployment speed. It is also a competition in governance capacity, trust infrastructure, and the ability to adapt to evolving rules. If invasive-agent providers seek short-term advantage by pushing beyond permission boundaries, weakening authorization mechanisms, and eroding user trust, such an approach may appear to secure an early lead, but in reality it may embed greater risks for the wider industry and society.   中美AI竞争格局的演进，不仅关乎技术能力和落地速度，也越来越取决于创新效率、安全约束与生态协同之间能否实现可持续平衡。对中国人工智能产业而言，具备安全、信任与规则兼容基础的Agent体系，将更有助于提升其面向全球市场的长期竞争力；若过度依赖以牺牲信任为代价的扩张路径，不仅可能削弱单一产品或企业的国际合作空间，也可能对中国AI整体的国际信誉带来负面影响，对融入全球主流的AI技术路线与治理体系同样构成负面障碍。 The evolution of China-US AI competition increasingly depends not only on technological capability and implementation speed, but also on whether innovation efficiency, security constraints, and ecosystem coordination can be balanced in a sustainable way. For China’s AI industry, an agent framework grounded in security, trust, and compatibility with emerging governance norms will be more conducive to strengthening long-term global competitiveness. By contrast, an expansion path built at the expense of trust may not only narrow the international cooperation space available to individual products or companies, but may also negatively affect the broader international credibility of China’s AI industry and create additional obstacles to its integration into mainstream global AI technology and governance systems.

NEWS On September 12th, Frost & Sullivan Frost & Sullivan China Center of Excellence (COE) officially opened in Zhangjiang Science City, Pudong New Area , marking a crucial step in the company's platform-based, high-end, and global strategic layout. At the opening ceremony, Dr. Wang Xin, Global Senior Vice President of Frost & Sullivan, Co-Chair of Asia-Pacific Region, and Chairman of China, along with industry peers and ecosystem partners, gathered in Zhangjiang Science City to witness the new development milestone of Frost & Sullivan Frost & Sullivan China Center of Excellence' innovation empowerment and ecological coexistence, and to jointly explore new paths for intelligent industrial inspection innovation. Dr. Liu Jun, Chairman of YUAN CAPITAL, and Dr. Wang Xin, Chairman of Frost & Sullivan, unveiled the center The newly launched YUAN CAPITAL (Shanghai) Innovation Center is located in the core area of Zhangjiang Science City, enjoying significant geographical advantages and a strong innovation atmosphere. The center integrates research and innovation, industrial cooperation, and strategic investment and mergers and acquisitions, forming a modern composite innovation venue for technology development, external exchanges, and internal empowerment. Integration, Accumulation, Empowerment UNICOMP WORLD Leveraging Shanghai's geographical advantages, YUAN CAPITAL (Shanghai) Innovation Center will deeply integrate into the Zhangjiang innovation ecosystem, collaborating with YUAN CAPITAL Asia Artificial Intelligence Center, adhering to "horizontal expansion, vertical deep cultivation" development strategy, playing a core hub role, coordinating the group's global strategic layout, the transformation of AI inspection technology results, and domestic and international operations, empowering platform-based innovation , continuously cultivating new growth trends for the company. In the future, YUAN CAPITAL will continue to build a leading platform-based technology enterprise in the industry, providing global customers with multi-modal integrated intelligent inspection solutions, empowering high-quality development for all industries with its strong innovation capabilities.

Introduction AI can read a large number of reports, web pages, and databases in a very short time, but "reading more" does not automatically mean "using it more reliably". When the conclusions in reports are broken into fragments, compressed into answers, or even used by intelligent systems for subsequent actions, industry research must re-answer a question: How can knowledge remain accurate after leaving the report? AI is entering industry research at a rapid pace. It can read reports, retrieve data, compare companies, generate summaries, and combine multiple materials to form a well-structured industry assessment. The time-consuming information search and preliminary sorting in the past are being reduced. What is more significant is that AI is changing the way knowledge is used. When market size data is extracted from a report, when a trend judgment enters a company's knowledge base, and when an expert opinion is re-expressed by a model, do these still retain their original time, scope, and applicable conditions? If new evidence emerges and old conclusions are corrected, can old knowledge already included in the model's responses, customer materials, and decision-making processes be identified promptly? If constraints are lost during transmission, who should explain, maintain, and correct them? In the past, industry research was mainly responsible for one report. In the future, it will also be responsible for the state of key knowledge in the report after leaving the original text and being repeatedly called upon by people and machines. This is the starting point of our proposed "Knowledge 3.0" research framework. From "remaining and being found," to "reliable use with accountability" "Knowledge 3.0" is not a historical phase with an academic consensus, nor a completed technical standard. It is a research framework we propose to understand the changes in knowledge in the AI era. In this framework, Knowledge 1.0 addresses "remaining". Text, publishing, archives, and corresponding professional systems enable knowledge to escape individual memory and be stably recorded, cited, and passed on. Knowledge 2.0 addresses "being found". The Internet, search engines, databases, and network collaboration enable scattered content to be connected, discovered, and jointly produced across regions and organizations. Knowledge 3.0 faces a new question: When machines begin to participate in the selection, transformation, and use of knowledge, can a piece of knowledge continue to be reliably used by people and machines? Can it be traced, questioned, corrected, and have someone accountable for it? These three capabilities are not mutually replaceable. Reports, books, and databases will not lose their value just because Knowledge 3.0 appears. On the contrary, complete documents remain important carriers for preserving the argument process, institutional context, and complex judgments. What needs to be added is another capability: after key knowledge in a document leaves its original position, it should still be able to move with the necessary context and return to the original text for inspection. Why don't more reports automatically become more reliable industry knowledge Industry research has long accumulated a large number of reports, data, and expert judgments. These contents are important foundations for training industry models, building company knowledge bases, and conducting retrieval-enhanced generation, but the scale of data does not equal the quality of knowledge. First, explicit knowledge may be distorted during transmission. The same "market size" may refer to sales volume, shipment volume, or terminal transaction amount respectively; the year in the same annual report may be the release date, the data benchmark year, or the start of the forecast period; the same "expected rapid growth" may be the model's calculation result or include analysts' comprehensive judgments on policies, competition, and supply conditions. When people read a complete report, they can often restore these boundaries from chapters, tables, and footnotes. When a machine only obtains a fragment, the boundaries are easily lost. Secondly, many experiences that determine the quality of research have not been included in formal results. An industry forecast may come from public data, corporate interviews, model calculations, and analyst judgments, but the final report may not record how researchers handled abnormal samples and contradictory interviews, why a certain scenario was chosen, and what changes would trigger a re-estimation. These experiences cannot be fully encoded; if no traces of judgment are left, machines may misinterpret professional judgments as unconditional facts. Therefore, a report database can preserve "what the industry has said before", but reliable industry knowledge also requires explaining: specifically who, at what time, according to what scope, based on what relatively independent evidence, formed which claim; under what conditions does this claim apply, and is it still valid now. What vertical models first need is not just more reports, but also better knowledge units. From document fragments to knowledge objects If the entire document is too large and isolated sentences are too thin, what the machine truly needs is a unit of expression between the two. We call it "knowledge object". A knowledge object is not the shortest text fragment, but the smallest knowledge unit that can be recognized and called upon with the necessary context and return to the original material for inspection. It can be a fact, an indicator, a definition, a method, or a conditional professional judgment. To avoid leaving only a seemingly precise field in the knowledge object without context, Knowledge 3.0 proposes five types of questions that cannot be forgotten. The content layer explains what it is about and under what conditions it applies; the experience layer retains the context, trade-offs, and exceptions of the judgment; the evidence layer connects direct sources, methods, and conflicts; the responsibility layer records who provided, transformed, inspected, released, maintained, and corrected; the lifecycle layer explains versions, status, alternative relationships, and reasons for changes. The five layers are not five forms requiring all knowledge to be mechanically filled out, nor does it mean the more fields, the more reliable it is. They are more like five groups of continuous questions: what we are using, why we can use it, where we cannot use it, and who to find when changes occur. Knowledge objects do not replace reports. Complete documents preserve the whole picture, and knowledge objects support precise calls; users can still return to the original text to restore the context after obtaining key conclusions. The two must be mutually accessible. How will Knowledge 3.0 change industry research This change affects industry research not simply by adding a delivery format, but by redefining the responsibilities of research institutions in the knowledge chain. First, reports will shift from being the end point of delivery to the starting point of the knowledge lifecycle Traditional research projects usually completed their main deliverables when the report was released. In the AI environment, a conclusion will continue to be retrieved, relayed, and combined. Research institutions need to identify which knowledge is frequently used, prone to failure, or has high error costs, and establish versions, status, and update triggers for them. When new market data appears, it should not just silently replace a number on a web page. What is really important is to explain: what claims have changed, whether the changes come from new evidence, scope adjustments, or corrections, and which existing judgments may be affected. Second, research quality will move from "the entire report being reliable" to "why a specific claim holds true" The brand and professional reputation of research institutions are still important, as they let users know why a material deserves priority attention. However, an institution's reputation cannot automatically turn predictions into facts, nor can it replace the evidence and boundaries of specific claims. Future industry research needs to distinguish facts, explanations, predictions, and suggestions, and ensure that evidence corresponds to specific claims. Similar views in multiple reports do not necessarily mean there are multiple independent evidences; they may just be repeated statements from the same source. The value of research institutions will be more reflected in discovering original evidence, distinguishing scopes, presenting differences, and explaining uncertainties. Third, analyst experience will transform from "individual ability" into "organizational traces of continuous learning" AI cannot replace researchers' judgments in complex situations, nor can it replicate an expert through a few fields. But organizations can more consciously preserve the origin of judgments: what signals were seen, what explanations were considered, why a certain scenario was chosen, whether the result met expectations, and what exceptions were later discovered. These records are not meant to turn experience into rigid rules, but to let machines and later scholars know that a judgment does not appear out of thin air, nor is it unconditional in all scenarios. For industry research institutions, this will turn personal experience into sustainable learning, review, and correction organizational capabilities. Fourth, research results need to face both human readers and machine users Reports for humans pursue complete arguments, expression rhythm, and reading experience; knowledge for machines also needs a stable identity, clear boundaries, corresponding sources, version status, and calling methods. Future industry research will not have to choose between "writing reports" and "doing data interfaces", but will need to establish two interconnected paths. People can read complete reports to understand complex logic, and machines can call upon key knowledge objects to complete retrieval, comparison, and auxiliary judgment; once risks increase or action is prepared, the system should be able to return the formal original text, reveal uncertainties, and hand over decisions to those with appropriate authority and responsibility. Fifth, research institutions will shift from content producers to important nodes in the knowledge responsibility chain In the Knowledge 3.0 environment, the role of research institutions may further expand: producing and organizing evidence, transforming knowledge objects, inspecting sources and scopes, maintaining the lifecycle of important knowledge, and providing entry points for explanation and correction. This does not require an institution to take full responsibility. Knowledge may pass through multiple entities from creation to use, so a visible responsibility chain needs to be established: who completed what action, who has the right to make decisions, and who will notify and correct after changes occur. Where should industry research institutions start Knowledge 3.0 should not start with "re-structuring all reports". A more cost-effective path is to first select those knowledge that are frequently individually called upon, change rapidly, or have high error consequences. Frequently used market size, business indicators, industry definitions, and key forecasts can prioritize supplementing sources, time, scope, statement type, and boundaries; content relying on expert judgments should retain context, exceptions, and re-estimation conditions; continuously changing content should record knowledge differences between versions, rather than just file modification dates. At the same time, the path for knowledge objects to return to the complete report should be preserved. Objectification is not discarding the original text after cutting it up, but connecting precise calling and complete reading. These practices need to be compared and verified: whether knowledge objects improve version selection, evidence correspondence, and conflict handling? Whether the lifecycle mechanism reduces old knowledge from entering new answers? Whether the additional maintenance costs are lower than the reduction in errors and reviews? Only through real-scenario comparative studies can Knowledge 3.0 move from a theoretical framework to a usable capability. Conclusion: Keeping knowledge reliable after leaving the report AI will not make industry research lose its value. It is forcing industry research to build its value in a deeper place. Future competitiveness of research institutions depends not only on who can publish reports faster and accumulate more content, but also on who can keep key knowledge accurate after leaving the original text: it can be called upon, evidence and boundaries can be found, versions and uncertainties can be seen, and the responsible person can be found when changes occur. When knowledge crosses organizational boundaries, facts, evidence, judgments, and inspection capabilities are still distributed among different entities. Connecting these knowledge, evidence, changes, and responsibilities may form a public credible knowledge network. It is not a central database or "unified truth", nor is it the only implementation of Knowledge 3.0, but a path to be jointly verified. Knowledge 3.0 is still an open research proposition, and it cannot be completed by a single institution alone. We look forward to communicating with three types of partners: technical partners in large models, knowledge engineering, data infrastructure, and industry application fields, to jointly verify whether the knowledge structure can bring measurable system increments; content partners such as research institutions, industry associations, enterprises, experts, and professional data providers, to jointly explore the production, expression, and continuous maintenance of high-quality knowledge; governance partners in standard research, independent inspection, legal compliance, and public governance, to jointly discuss the boundaries of evidence, responsibility, correction, and publicness. This is not an recruitment for a predefined platform, nor a prior commitment to a set of standards. We hope to start from real industry problems, jointly define issues, conduct comparative verification, and explore an industry knowledge infrastructure that can be reliably used by people and machines.

Frost & Sullivan Insight From an industry perspective, the core barrier to computing power leasing is the comprehensive ability to acquire resources, operate clusters, provide customer service, and manage funds. Enterprises do not need to build their own infrastructure; long-term leasing or hosting third-party resources is also feasible. The key lies in whether substantial control over resources is achieved, and whether responsibility for scheduling, operation, and SLA delivery is independently held, with primary responsibility for service quality. The industry distinguishes service providers from intermediaries mainly based on resource control rights and delivery responsibilities: service providers actually control resources and bear operational risks, with revenue coming from computing power and value-added services; the barriers lie in scale and technology. Intermediaries facilitate supply and demand to earn commissions, with low risk but thin profits, and barriers depend on channel efficiency. The fundamental difference between the two is whether they bear asset risks and add technical value. An interview with Frost & Sullivan China's Senior Partner and Managing Director, Jia Pang, was conducted by a reporter from The Times Weekly regarding the entry barriers and business models of the computing power leasing market. Q: From an industry perspective, what are the core barriers to computing power leasing? What key capabilities are required for an enterprise to enter the computing power leasing market? Jia Pang Senior Partner and Managing Director of Frost & Sullivan China The core barrier to computing power leasing is whether a stable, controllable, and sustainable computing power supply capability can be established, mainly including four aspects: First, the ability to acquire and control computing power resources , enabling stable acquisition of key resources such as GPU servers, cabinets, networks, and electricity; Second, the ability to operate and deliver computing power resources , including the ability to pool resources, manage clusters, schedule operations, monitor and measure, and handle failures, ensuring service availability and SLA; Third, the ability to acquire and serve customers , continuously meeting customer needs and improving the utilization rate of computing power resources; Fourth, the ability to manage funds and assets , effectively managing risks such as resource procurement, equipment depreciation, idle resources, and technological iteration. Therefore, the competitive barriers in the industry mainly reflect the comprehensive ability of "resource acquisition and control + operation and delivery + customer service + fund and asset management". Q: Do computing power leasing enterprises must have their own computing power infrastructure? If they mainly rely on third-party procurement or leasing of GPU resources and then provide services downstream, does this model have long-term commercial value? How is the scale barrier measured? Jia Pang Senior Partner and Managing Director of Frost & Sullivan China Not necessarily. There are both heavy-asset models where enterprises purchase and build their own computing power infrastructure, and models that obtain computing power resources through long-term leasing, hosting, and cooperative procurement. Using third-party resources does not affect the enterprise's ability to engage in computing power leasing business. The key is whether stable and continuous control over relevant resources can be achieved, and whether responsibility for resource allocation, scheduling, pricing, billing, operation, and customer service is independently held, with responsibility for final delivery quality and SLA. If only short-term resource matching or reselling is done, the business stability and customer stickiness are usually limited. Currently, there is no unified threshold for the number of GPUs or revenue scale in the industry. To determine whether an enterprise operates on a large scale, it is more appropriate to comprehensively evaluate indicators such as stable available GPUs or equivalent computing power scale, resource utilization rate, customer and contract stability, computing power service revenue, and renewal status, rather than solely measuring by the number of own GPUs. Q: How does the industry define "computing power leasing service providers" and "computing power intermediaries/resource integrators"? What are the essential differences between the two in business models, profit margins, and competitive barriers? Jia Pang Senior Partner and Managing Director of Frost & Sullivan China There is no unified legal classification in the industry. In practice, it can be distinguished mainly by resource control rights and final delivery responsibilities. Computing power leasing service providers usually have relatively stable control over relevant computing power resources and are responsible for resource allocation, scheduling, measurement and billing, operation, and customer service, bearing primary responsibility for service availability and SLA; computing power intermediaries or resource integrators mostly carry out functions such as supply and demand matching, resource matching, and transaction coordination, usually not continuously controlling underlying computing power resources, and their responsibility for final delivery is relatively limited. In terms of business models , computing power leasing service providers mainly charge fees for computing power leasing and related operation, deployment, etc., while bearing risks such as resource lock-in, equipment depreciation or leasing, idle resources, and price fluctuations; intermediaries or resource integrators mainly earn income through commissions, service fees, or transaction price differences, with relatively low asset investment and resource risks. The former usually has higher revenue scale and single customer value, but profitability depends more on resource utilization rate, procurement costs, and pricing ability. In terms of competitive barriers , computing power leasing service providers rely more on stable resource acquisition and control, cluster operation and scheduling, SLA delivery, customer base, and fund and asset management capabilities; intermediaries or resource integrators rely more on supplier and customer channels, supply and demand information coverage, and matching efficiency. The core difference between the two models is whether the enterprise has substantial control over computing power resources and bears primary operational and delivery responsibilities for customers. *This interview was published in The Times Weekly. The author is Zhu Chengcheng. The original title was: Exclusive | Gaole Co., Ltd.'s 6.7 billion yuan computing power deal mystery: The registered address of its Harbin subsidiary is virtual, and more than half of the computing power business team joined less than 3 months ago?

近日（2026年8月）， 上海昂朴生物科技有限公司 （下称“昂朴生物”）与 北京无界进化科技有限公司 （INFevo，下称“无界进化”）正式宣布达成深度战略合作。双方将整合昂朴生物诱导多能干细胞（ iPSC ）疾病模型与多组学数据库资源、无界进化自研AI虚拟细胞大模型OCOO与计算生物学能力，共建适配全球NAMs监管体系的遗传病创新研发平台，开辟以“ 人源数据+虚拟计算 ”替代传统动物实验的新药研发路径。 当前全球生物医药正迎来 “去动物化” 研发范式变革，各国监管机构持续放开新型非动物测试方法学（ NAMs ）的申报应用通道。2025年4月，美国FDA官宣逐步取消单克隆抗体等药物的动物实验硬性要求；2026年3月，FDA发布NAMs使用指南草案，明确药企可提交类器官、器官芯片、AI计算模型生成的数据直接用于非临床安全性评价。国内同步加速产业落地，国家药审中心启动NAMs研究应用“先锋计划”，中国食品药品检定研究院设立创新方法评价中心，一套兼容人源细胞、数字模拟技术的全新审评体系正在成型。 在此产业转型关键窗口期，昂朴生物与无界进化的强强联合，将搭建国内自主可控的“实体人源细胞+AI虚拟模型”一体化研发底座，填补遗传病领域数字化非临床评价体系空白。   昂朴生物： 国内领先的iPSC智能制造与遗传病数据台 昂朴生物成立于2013年，是一家专注于遗传病领域和诱导多能干细胞（iPSC）技术的高新技术企业。公司已通过CAP、CMA和CNAS等多项权威质量体系认证。 在细胞模型层面，昂朴生物构建了国内领先的工业级iPSC制备分化平台，拥有500㎡工厂级实验室及1,400㎡GMP厂房，可定向分化为多种细胞模型，批次稳定性优于行业平均水平。在数据资产层面，昂朴生物建设了中国领先的遗传病iPSC库（iBGD），当前已积累近1,000株iPSC细胞株、300TB+多组学数据，覆盖罕见病目录内重点疾病。这一资源库对标日本RIKEN、美国 CIRM 等国际级iPSC生物样本库，填补了中国在遗传病iPSC领域的产业基建空白。   无界进化：从基准 SOTA 到临床一线的AI虚拟细胞引擎 无界进化（INFevo）成立于2025年，是晶泰科技（2228.HK）深度孵化的前沿AI生物企业。2026年6月，公司完成数千万元人民币天使轮融资，由顺为资本、红杉中国与松禾资本共同参与。 无界进化专注数字生命与计算生物学，以自研虚拟细胞基础模型 OCOO 为底座，结合强化学习构建面向药物发现的自主科学发现系统TPP（The Popper Project），将模型预测、实验设计、实验验证与结果回流纳入统一研发工作流，通过持续迭代不断提升模型能力。公司正加速建立覆盖新靶点发现、虚拟药物筛选与评价、临床方案优化的全功能AI生物学平台。   双向印证：“虚拟评价—人源验证—结果回喂”的正向研发闭环 本次合作实现标准化人源疾病多组学数据与高精度AI虚拟细胞模型深度耦合，完美契合全球NAMs监管政策导向，形成行业独有的闭环研发体系。 AI虚拟细胞的预测结论要成为研发决策依据，离不开大规模、高质量人源数据的独立验证；而人源细胞实验要形成规模化产出，同样需要足够强的计算与排序能力与之匹配。本次合作正是让二者深度耦合：OCOO虚拟细胞平台在数字空间完成潜在治疗靶点挖掘以及高通量化合物评价，精准推演不同遗传背景、不同人体细胞类型下的药物药效与毒副作用，并依托海量人源数据进行交叉验证；昂朴生物iBGD资源库输出大批量、可溯源、匹配真实人体病理特征的基因型—表型—药物响应多组学数据，既为模型预测提供独立的交叉验证基准，也在湿实验层面对结果给出校正。 两者互为印证、彼此校正，落地“虚拟评价—人源验证—结果回喂”的全新研发范式：以iPSC分化的人体功能细胞替代传统实验动物，以大数据与AI模拟替代高成本、低效率的经验试错实验。每一轮湿实验缩小模型探索空间，进一步模型又压缩下一轮所需的湿实验量，研发成本与周期随迭代持续下行。双方由此打通实体细胞实验与虚拟评价的链路，为药物创新提供符合国内外审评标准的基础设施，并推动新药的开发落地。 三个真实困境，一套共同解法 新药研发中代价最高的并非失败本身，而是失败发生得过晚。以下三类困境分处药物发现、疾病研究与临床决策的不同环节，却指向同一个技术缺口——缺少一个兼具人体生物学真实性与大规模计算能力的中间层。 当一个分子在动物模型中有效，却在临床阶段失效。 这是创新药研发中代价最高的系统性偏差，也是“去动物化”变革的根本动因。联合平台以“靶点虚拟研究与化合物虚拟评价＋人源细胞实验验证”的双层漏斗推进研发，快速形成带证据链的靶点与化合物优先级清单，以及符合FDA、CDE NAMs申报口径的人源数据，直接支撑项目决策与申报。将淘汰环节前置至成本最低处，意味着更短的靶点发现至临床前候选化合物（PCC）周期、更低的动物实验成本与合规风险，以及在全球监管转向中的先发位置。 当一种疾病全国仅有数十例患者，研究难以形成通量。 样本稀缺、动物模型缺失或失真、单中心难以支撑系统性筛选，是遗传病与罕见病研究长期的结构性困境。iBGD已覆盖罕见病目录内重点疾病的iPSC细胞株，可定向分化为运动神经元、皮层神经元、心肌细胞、星形胶质细胞、视网膜色素上皮细胞等疾病相关细胞类型；配合虚拟细胞模型，同一突变背景可在数字空间完成大规模扰动推演，产出疾病机制假说的计算生成与优先级排序、候选干预方案的预测结果，以及可直接在人源细胞上落地的实验设计。病例数由此不再构成研究通量的上限，候选药物的临床转化亦得以提速。 当临床存在明确的用药选择难题，却缺少可计算的解法。 真实世界的治疗决策多依赖经验与有限证据，难以复用与横向比较。虚拟细胞模型与患者来源细胞、类器官体系可形成双向迭代：AI提供可计算、可比较、可排序的预测框架，真实样本提供反馈与校正，产出面向具体患者群体的疗效与耐药预测、机制解析与潜在治疗靶点线索，并可沉淀为标准化的研究与分析范式。 三者分处研发链条的不同位置，制约却同源：足够真实的实验难以规模化，足够快的计算不足以还原人体。无界进化与昂朴生物的联手，让人源细胞与AI虚拟细胞的耦合补上真实性与通量之间的这一层，这也是本次合作的技术起点。   共启新篇：构筑国产研发底座，持续开放平台合作 当全球监管体系加速告别动物实验，当“人源数据+AI计算”成为药物研发的新坐标，昂朴生物与无界进化的战略合作，正是对这一时代命题的回应。 昂朴生物的遗传病iPSC数据银行，提供了中国自己的、可规模化获取的人源疾病模型与数据资产；无界进化的AI虚拟细胞模型，则提供了在计算机中预测药物响应的智能引擎。这套融合体系，对产业、药企、患者具备多重核心价值：对于创新药企，整套研发链路产出的数据符合FDA、CDE NAMs申报要求，有效压缩临床前研发周期、削减动物实验相关成本；对于遗传病领域，解决传统动物模型病理模拟失真、转化效率低下的行业痛点，加速疑难遗传病全新疗法开发；对于国内生物医药产业，搭建不依赖海外样本库与海外计算模型的自主研发底座，夯实国产创新核心竞争力。 未来，双方将持续扩充遗传病标准化细胞模型库，迭代优化遗传病专属AI虚拟细胞解决方案，完善适配国内外审评标准的NAMs技术方案。双方合作将聚焦于遗传病相关方向的新药研究与开发，并积极向全行业开放联合研发平台能力，积极推动在遗传病方向的治疗突破，为患者带来有效的治疗。

Frost & Sullivan Insight Recently, the State Council issued the "15th Five-Year Plan for Carbon Peak Action", clearly proposing to improve the energy conservation and carbon reduction standards for computing power facilities, promote energy conservation and carbon reduction reforms for non-compliant computing power facilities, and orderly phase out backward and inefficient technologies and equipment, thereby enhancing the energy efficiency level of computing power facilities. The 2026 Government Work Report indicates that the proportion of green electricity used in new data centers at national hubs should exceed 80%, and it is encouraged to achieve 100% consumption of green electricity. What are the relatively mature energy conservation and carbon reduction approaches currently available in the industry, ready for large-scale implementation? For example, technologies such as liquid cooling, waste heat recovery, and CPO (Co-packaged Optical) are in what application stages? Currently PUE , CUE, and unit computing power carbon emissions ("computing power carbon efficiency") are multiple evaluation indicators. Do you think a unified national green computing power evaluation system will be needed in the future? What are the challenges in implementing a unified standard at this stage? In the next three to five years, which aspects do you believe the largest changes in the green computing power industry will manifest? Will the development of AI technology itself change the energy conservation methods for data centers? For instance, optimizing cooling scheduling through AI and predicting loads to reduce idle energy consumption—how is the progress of this "using AI to save energy with AI" approach? Zhou Mingzi, Partner of Frost & Sullivan China, interviewed by The Times Weekly to discuss how green computing power opens a new paradigm of "mutual pursuit". Securities Daily Q: What are the relatively mature energy conservation and carbon reduction approaches currently available in the industry, ready for large-scale implementation? For example Liquid cooling , waste heat recovery, and CPO (Co-packaged Optical) technologies are in what application stages? Zhou Mingzi Partner of Frost & Sullivan China From the perspective of industry commercialization, the technology that has entered the stage of large-scale application in energy conservation and carbon reduction for computing power is cold plate liquid cooling, with a market share of over 80%. Currently, not only leading brands such as Inspur and Huawei have natively supported servers with this technology, but new intelligent computing centers are basically configured with it. PUE can be stabilized between 1.15 and 1.25, making it the only technology path that can be "delivered and used immediately" at present. Although immersion liquid cooling offers better performance, it faces practical challenges such as high initial investment and the need to rebuild the operation and maintenance system, so it still relies on zero-carbon parks and benchmark projects from finance and operators. The industry predicts that it will take about 3 to 5 years before large-scale implementation. Waste heat recovery technology is mature, but its commercial closed loop is severely limited by geographical conditions: in the north, projects like Ningxia Zhongwei Ningxia Zhongwei Frost & Sullivan, Frost & Sullivan China, LeadLeo, LeadLeo Research Institute, LeadLeo, YUAN CAPITAL, SULLIVAN TELE-TREND CLOUD TECHNOLOGY, TradeGo, MagnaTEC, Automotive & Mobility, Environmental Protection & Energy Saving Technology, Logistics & Supply Chain, MATERNAL AND INFANT, Education & Training, Real Estate & Property, Catering & New Retailing, Advanced Materials, Healthcare & Life Sciences, Semiconductor & Chip, COMMERCIAL AVIATION, Technology, Media and Telecom, LANDSCAPING, Big Data & AI, Infrastructure Construction & Utilities, Culture & Entertainment, AGRICULTURE, FORESTRY ANIMAL HUSBANDRY AND FISHERY, Food & Beverage, Fintech, SHIPPING AND PORTS, Dual Carbon & New Energy, Mining & Metals, Public Sector, Cross-Border E-commerce Trade, Building Technology, Construction & Decoration, Beauty & Fashion, Smart Homes, Digital Infrastructure, Enterprise Services, Consumer Electronics, and other projects have already integrated waste heat into the green electricity direct supply solution; in the south, they can only connect with scattered scenarios such as greenhouses and pools, with a long recovery period, and no independent replicable business model has yet been established. CPO is the most significant differentiation: the exchange side will start production in the second half of 2026. Nvidia's Quantum 3400 CPO switch has been scheduled for mass production, reducing transmission losses by 60%; however, scale-up optical interconnection between chips will not be scaled up until the Feynman architecture is mass-produced in 2028, when the market is expected to see a new wave of applications. Q: Currently, multiple evaluation indicators such as PUE, CUE, and unit computing power carbon emissions ("computing power carbon efficiency") exist. Do you think a unified national green computing power evaluation system will be needed in the future? What are the challenges in implementing a unified standard at this stage? Zhou Mingzi Partner of Frost & Sullivan China From the perspective of industry development, it is necessary to establish a national unified green computing power evaluation system, but achieving "complete comparability" at this stage is not realistic. The limitations of a single PUE indicator are already very obvious: it does not reflect the energy structure—a coal-based data center with PUE 1.1 may have several times higher carbon emissions than a green electricity data center with PUE 1.3; it also does not measure computing power output efficiency—a cluster with PUE 1.15 but a GPU utilization rate of only 30% is essentially "false efficiency"; moreover, scope three emissions such as construction period, equipment manufacturing, and decommissioning are completely outside the coverage of PUE. Policy efforts are already shifting: the 2026 Government Work Report included "computing power and electricity coordination" in the new infrastructure project for the first time. The National Data Bureau and the National Energy Administration have clearly defined two constraints: "green electricity proportion of new computing power facilities ≥ 80%," and "PUE ≤ 1.20 in the west region ≤ 1.25 in the east region." The China Academy of Information and Communications Technology is also promoting two new indicators for AI business outputs—computing power energy efficiency and Token energy efficiency—these two new indicators mark the shift of the industry's evaluation focus from "data center efficiency" to "business output efficiency". On the other hand, implementing a unified standard faces three substantial obstacles: first, the criteria for green electricity traceability have not been unified, with differences in the recognition standards for green certificates, green electricity trading, green electricity direct supply , and self-generation and self-use photovoltaic paths; second, there is a lack of industry consensus on the accounting method for scope three carbon emissions; third, the differences in regional resource endowments are too large. Western hubs like Ningxia can meet the requirements by relying on wind and solar power plus green electricity direct supply, while the Yangtze River Delta needs cross-provincial green certificates and energy storage support to achieve 80%. Forcing a comparison using the same score will cause certain concerns in both eastern and western industrial markets. Therefore, the more likely evolution path in the future is that the state will formulate a comprehensive framework, and industry associations will introduce hierarchical leading values, similar to the gold and silver grading logic of LEED certification, to achieve better implementation results. Q: In the next three to five years, which aspects do you believe the largest changes in the green computing power industry will manifest? Zhou Mingzi Partner of Frost & Sullivan China Over the next three to five years, Frost & Sullivan believes that changes in the green computing power industry will focus on four interconnected main lines. The first is that "green electricity 80%" has shifted from policy advocacy to rigid constraints, fundamentally restructuring the investment logic for data centers. Among the eight major computing power hubs, only Ningxia, Gansu, and Inner Mongolia have mature green electricity direct supply channels, while the green electricity proportion in most other nodes is less than 30%. The 2026 Government Work Report has explicitly included this proportion as a constraint. This means that 2026 to 2028 is a window period to seize green electricity direct supply channels, sign long-term power purchase agreements, and secure source-grid-load-storage resources. Projects that miss this window will find it difficult to pass energy assessment. The second is that the industrial layout will further refine from "east data, west computing" to a new map of "green electricity hinterland training and storage, coastal reasoning, and hub transit". Training clusters above ten thousand units will basically be locked in the west region. The third is that liquid cooling combined with zero-carbon parks will shift from demonstration to a required standard. The "15th Five-Year Plan" explicitly requires the construction of 100 national zero-carbon parks, and the green and low-carbon transformation of computing power facilities has been separately listed. In the future, new large-scale clusters will basically feature a combination of liquid cooling, green electricity direct supply, and waste heat utilization. The fourth is that computing power and electricity coordination will generate new trading products. Data centers, as adjustable loads, will participate in grid peak shaving, green electricity spot trading, and non-real-time tasks across regions. These activities will make "computing power carbon efficiency" and "green electricity traceability" tradable assets. The return on investment for projects certified by CCER will be higher than pure hardware investment. Frost & Sullivan believes that in the next three to five years, what will truly differentiate companies is not hardware parameters, but the ability to sign green electricity procurement agreements, operate carbon assets, and schedule computing power assets—while hardware can be obtained by everyone, green electricity channels, carbon certification, and credible computing service assets are scarce resources. Q: Will the development of AI technology itself change the energy conservation methods for data centers? For example, optimizing cooling scheduling through AI and predicting loads to reduce idle energy consumption—what is the progress of this "using AI to save energy with AI" approach? Zhou Mingzi Partner of Frost & Sullivan China Frost & Sullivan believes that the path of using AI technology to optimize the energy consumption of data centers actually progresses faster than most market expectations. However, a key boundary must be clarified: AI optimization is the most cost-effective lever for existing infrastructure transformation, but it cannot save old air-cooled data centers with PUE above 1.3—those data centers need to replace liquid cooling hardware first, and AI can only add value on this basis. Technological evolution has gone through three stages: the first stage is pre-optimization based on load prediction to avoid temporary peak impacts; the second stage is global optimization, combining optimization of chillers, water pumps, cooling towers, and CDU to extract an additional 10% to 30% energy efficiency from the existing energy-saving framework; the third stage is computing power and electricity coordination, linking cooling strategies with the timing of green electricity output and computing power migration. When green electricity is abundant, pre-cooling is done in advance, and batch processing tasks are run more frequently; when green electricity is scarce, non-real-time loads are actively reduced. Frost & Sullivan believes that the product form with real premium potential in the future may be a complete package of "AI optimization plus green electricity timing matching plus cross-regional load migration," rather than simply selling AI energy-saving software. A high-end equipment manufacturing company in Wuxi has entrusted EasyEnergy to manage the energy and computing power in its park through intelligent operation management. EasyEnergy uses AI to coordinate servers, cooling, photovoltaics, and energy storage systems, predicting computing power loads, equipment temperatures, and green electricity output, and dynamically adjusting cooling strategies, server operating states, and task timing to achieve integrated optimization of computing power, cooling, and electricity. When green electricity is abundant, the system can pre-cool and increase batch processing and reasoning tasks; during peak power periods, frequency reduction of servers, task off-peak scheduling, and energy storage adjustment are used to reduce peak energy consumption. This scenario upgrades traditional single-cooling energy saving to global optimization of computing power, cooling capacity, and power resources, reflecting the acceleration of "using AI to save energy with AI" from technical verification to actual operational scenarios. *This interview was published in Securities Daily. The author is Wang Jingru. The original title is: Green Computing Power Opens a New Paradigm of "Mutual Pursuit"