高盛最新中国AI报告：原生应用势头强劲，AI智能体商业化模式初现

Core Insights - Goldman Sachs highlights the strong growth momentum of AI-native applications in China's software market, transitioning from foundational integration to higher-level LLMs for specific use cases by Q2 2025 [1] Group 1: AI Agents and Multi-modal Models - AI agents are evolving from concepts discussed internally to commercially viable products, serving as new user interfaces for enterprises and knowledge workers [1] - Private company Manus launched a general-purpose AI agent in March 2025, capable of complex tasks like resume screening and property searches, priced at $39 per month [1] - Software vendors are accelerating the integration of AI agents into their platforms, with Kingdee launching five specialized AI agents in May 2025 [1] Group 2: Enterprise Demand and Consumer Challenges - Strong demand for AI applications in enterprises is evident, particularly from state-owned enterprises, schools, and government clients, with significant project wins reported [2][3] - The deployment of private AI models, exemplified by DeepSeek R1, is driving demand for AI model deployment projects, which often include integrated hardware solutions, increasing contract values [2] - Despite robust enterprise demand, consumer-facing AI tools face significant monetization challenges, with low conversion rates typically ranging from 3% to 13% [3][4] Group 3: Domestic Software Development and EDA - The acceleration of domestic EDA (Electronic Design Automation) software development is a clear growth area, with local suppliers rapidly releasing new products to capture the increasing demand from domestic clients [2][3] - Companies like Huada Jiutian and Jiutongfang are launching upgraded tools for mixed-signal SoCs and digital simulation, respectively, indicating a focus on enhancing local capabilities [3] Group 4: Cost and Quality Improvements - AI firms are making significant strides in multi-modal models, focusing on lower costs, higher quality, and longer context capabilities, with recent upgrades from companies like Stepfun and ModelBest [2][4] - The pursuit of professional-grade content generation is evident, with advancements in image, video, audio, and 3D content creation, showcasing the broad potential of multi-modal technologies [4] - Companies are working to reduce model parameters and computational costs, exemplified by Stepfun's Step1X-Edit model, which maintains high efficiency and quality with only 19 billion parameters [4]