Goldman Sachs' latest China AI report: Strong momentum for native applications, initial emergence of AI intelligent body commercialization models

Goldman Sachs pointed out that the Chinese software market continued its strong development momentum for AI-native applications in the second quarter of 2025. In the fields of AI foundational models and applications, AI vendors are transitioning from the foundational integration stage to higher-level LLMs that use specific use cases.

According to news from the Chasing Wind Trading Desk, on July 7, Goldman Sachs published a research report indicating that intelligent agents (AI Agents), multimodal AI models, and model deployment have become the three core growth engines of the Chinese software industry:

• Strong momentum for AI innovation, but direction is diverging: AI agents and multimodal models are moving from concept to application, becoming the core driving force for enterprise software upgrades. This indicates that continuous investment in technology and product iteration will be the norm.
• Solid demand on the enterprise side (ToB), with government and state-owned enterprises as major buyers: Bidding data for AI projects shows that the real buyers are concentrated in state-owned enterprises, government agencies, and schools. In particular, privatized model deployment represented by DeepSeek is becoming a sizable and clearly growing market.
• Huge monetization challenges on the consumer side (ToC), with low conversion rates still a shortcoming: Despite the proliferation of AI applications, data shows that the payment rates for AI tools aimed at individual users are generally low (mostly between 3%-13%), contributing limitedly to total revenue.
• Acceleration of domestic substitution, with EDA software becoming a certain growth track: In specific fields, such as EDA (Electronic Design Automation) software, domestic demand is driving local vendors to accelerate product releases and market expansion, forming a strong growth point.

Rise of AI Agents: A New Generation of Interaction Interfaces for Enterprise Software

The report indicates that AI agents are rapidly evolving from concepts discussed on edge devices and within enterprises to commercially viable products.

This is not just a simple "AI+", but is expected to become a new type of user interface for enterprises and knowledge workers, with core capabilities to respond to and even proactively adapt to environmental changes, autonomously plan and complete workflows.

Goldman Sachs cited the example of the private company Manus, which launched a general AI agent in March 2025 that can independently complete complex tasks such as resume screening and property searches, offering a pricing plan of $39 per month.

Additionally, software vendors are accelerating the integration of AI agents into their professional platforms.

For example, Kingdee launched five specialized AI agents for finance, travel, recruitment, and more in May; Meitu also plans to launch AI agents in the second half of 2025 to optimize user experience in its entertainment and productivity tools.

Solid AI Demand on the Enterprise Side, but Huge Monetization Challenges on the Consumer Side

Goldman Sachs analyzed the bidding situation for enterprise application projects from late April to the present and found that the project momentum in the second quarter of 2025 is robust.

A notable trend is that the release of foundational models represented by DeepSeek R1 has significantly stimulated the demand for state-owned enterprises, schools, and government clients to deploy privatized AI models.

From specific projects, the overall scale of AI model deployment projects is generally larger than that of other ERP or system upgrade projects, as they often include integrated solutions such as computing hardware, which drives up contract value. ** Examples of specific winning cases in research reports:

• The LLM model project of China Steel Corporation, with an amount as high as 43.5 million RMB.
• The ERP upgrade project of China National Petroleum Corporation, with an amount of 19.7 million RMB.
• The ERP system project of China Tobacco Fujian Industry, won by Kingdee, with an amount of 24.4 million RMB.
• In addition, several universities and government big data centers have purchased deployment services based on DeepSeek, with contract amounts ranging from 2 million to 15.7 million RMB.

Despite the solid demand for AI applications on the enterprise side, data provided by Goldman Sachs reveals the severe challenges faced in commercialization on the consumer side:

• Low payment rates: Among various consumer-facing (ToC) AI tools, the payment conversion rates are mostly unsatisfactory. Even for the better-performing Kingsoft WPS AI, the payment rate is only 13%. The payment rate for Meitu Xiuxiu is 5%, while most tools hover between 3% and 6%.
• Limited revenue contribution: In terms of contribution to total revenue, the revenue generated by AI features remains very small. The revenue contribution of iFlytek's Spark is only 1%, and Meitu Kaipai is also 1%. Only a few mature products (such as Wondershare's Filmora, with an AI feature revenue contribution of 46%) perform outstandingly, while most new AI applications are still exploring effective profit models.

Upgrading multimodal models, accelerating localization of EDA/IP software

Goldman Sachs points out that AI vendors are continuously pushing multimodal models towards lower costs, higher quality, and longer context, while the acceleration of EDA software localization has become a certain growth track among AI vendors.

AI vendors have made significant progress in multimodal AI models, with companies like Stepfun, Minimax, and ModelBest releasing upgraded multimodal AI models in recent months. These models focus on the generation and editing of images, videos, audio, and 3D content, developing towards professional-level standards while reducing computational costs and lowering entry barriers, enhancing quality:

• Professional-level content generation: Model capabilities are evolving from basic text and image generation to professional-level generation and editing of images, videos, audio, and even 3D content. For example, Wondershare's "Sky Screen 2.0" large model can achieve cinematic camera effects; Desay SV and ModelBest have collaborated to develop an AI audio solution for smart cockpits, showcasing the broad potential of multimodal technology in different application scenarios.
• Lower cost thresholds: Vendors are committed to reducing model parameters and computational costs. For instance, the Step1X-Edit image editing model launched by the private company Stepfun has only 19 billion parameters while ensuring high efficiency and quality.
• Higher quality output: Pursuing high-definition image quality and precise physical simulation has become a new direction. Stepfun's Step1X-3D model focuses specifically on 3D content generation At the same time, as the core tool for chip design, local EDA suppliers are accelerating the release of new products in the second quarter of 2025 to seize the growing localization opportunities among analog/digital fabless and foundry customers.

Huada Empyrean released an upgraded parasitic parameter analysis tool and a Liberal IP tool for mixed-signal SoCs; Jiutongfang (private company) launched digital simulation and verification tools and announced a trial period for specific customers in June 2025.