Key points of investment:

Distributed vector database Hippo and big model operation management software Sophon LLMOps launched the enterprise-level vertical question-and-answer model “Infinity”, and the big model for data analysis “Search”. ① Wuya is an enterprise-level vertical question-and-answer model based on a self-developed pre-trained model, vector database Hippo, and graph database StellarDB.

In the financial field, “Infinity” incorporates rich financial reports of listed companies and industrial chain map data, which can provide financial institutions with comprehensive and in-depth investment research and analysis tools. In the field of laws and regulations, “Infinity” incorporates common legal and regulatory knowledge such as trading rules and regulatory requirements of major exchanges. Users can ask questions about specific provisions of laws and regulations, regulatory rules, trial measures, etc., and “Infinity” will provide legal risk warnings and suggestions for dealing with them. ② “Search” can use natural language to describe complex business requirements involving various data models, and allows non-professional users to freely query data as needed through natural language without learning and mastering database programming languages. ③ By the end of 2023, orders for the company's AI Infra related products had exceeded 54 million yuan, which were implemented in various industries such as finance, government, and energy.

Further extending from Data Infra to AI Infra, we are building AI infrastructure for enterprise customers, creating a complete tool chain from corpus processing, model training, and knowledge base construction to help enterprises quickly establish large industry models and quickly use AIGC. Starlink Technology integrates big data, artificial intelligence and other technologies to launch the knowledge platform TransWarpKnowledge Hub (TKH) to provide enterprises with a comprehensive, efficient, and intelligent data processing and knowledge management solution. It can achieve the goal of knowledgizing all data within the enterprise, and can also control the enterprise's Infraal computing power, corpus, models and applications to help enterprises build a safe and controllable AI Infra and help enterprises quickly use AIGC.

First, TDH for AI creates a new generation of one-stop multi-model digital docks. The multi-mode base of the big data infrastructure TDH 9.4 accelerates AI analysis. The distributed vector database Hippo supports the storage, indexing, and management of 10 billion vector data after conversion, such as text, images, audio and video. The distributed graph database StellarDB provides trillion-level graph data storage, and the Starlink distributed analytical database ArgoDB supports standard SQL syntax and provides multi-mode analysis and real-time data processing. The Starlink distributed time series database Timelyre supports the storage and processing of massive time series databases.

Second, corpus processing tools and map construction tools help enterprises establish high-quality models and applications. Starlink Technology has released Transwarp Corpous Sudio (TCS), a one-stop multi-scene corpus platform, which covers the entire life cycle of corpus acquisition, cleaning, processing, management, application and management. It has a variety of flexible collection and construction methods, and can efficiently process large amounts of corpus in a distributed manner. Enterprise users can use Transwarp Knowledge Studio for LLM, a star ring map construction tool provided by TKH to transform internal corporate data, personal experience data, and public information data into knowledge, making the data platform more intelligent. Third, the basic services of the big model are Infinity and LLMOps, so that the big model can be implemented quickly. Starlink Technology launched Transwarp SophonllMops, a large model operation platform, to provide a one-stop enterprise-level large-scale model production and application development tool chain to help enterprises complete the entire life cycle from pre-training to fine tuning, to reinforcement learning, to continuous model evaluation. Infinity, the Infinity model independently developed by Starlink Technology, has autonomous and controllable characteristics to ensure data security. Through 0-1 pre-training, it can tailor its own large models to various industries, providing strong intent understanding, semantic recall, data processing and analysis capabilities.

Based on the Wuya Big Model, the Starlink knowledge platform TKH has created native AI applications such as Wuya “Insight, Wuya” Questionnaire, Wuya” Finance, and Wuya Engineering, which can be widely used in various fields such as finance, energy, manufacturing, and engineering. Wuya Insight supports rapid storage of multi-modal data such as audio, video, and graphics of unlimited length, and supports automated document slicing and vectorized processing. With the self-developed RAG framework, it can achieve accurate recall of knowledge, and can be used in rich business scenarios such as market research and analysis, enterprise supply chain analysis, legal risk warning, and intelligent writing. In addition, Wuya” Question Count is based on the big model of Star Ring data analysis, and combines data analysis subjects, indicators, label design, data development and governance to form a complete process of changing from natural language to data query language and returning data tables or data charts.

Profit forecast and investment suggestions: On March 14, 2023, the company issued the “2023 Restricted Stock Incentive Plan Implementation Assessment Management Measures”. The assessment targets for operating income are: based on 2021 revenue, the 2023 revenue growth rate is not less than 60%; the 2024 revenue growth rate is not less than 125%; and the 2025 revenue growth rate is not less than 200%. Furthermore, when the trigger value is reached, the company level can also obtain an 80% ownership ratio: based on 2021 revenue, the 2023 revenue growth rate is not less than 48%; the 2024 revenue growth rate is not less than 100%; and the 2025 revenue growth rate is not less than 160%.

We believe that facing the impact of artificial intelligence in the industry, as well as product agility and popularization improvements, the company's products are also iterating towards multi-modality, intelligence, agility, and popularization. The company works both internally and externally, continuously improves ecological construction, builds a partner structure system, adopts different cooperation strategies, improves the talent system internally, and launches stock incentive plans. In addition to this, Xinchuang needs to be launched urgently, and the company is expected to benefit from being on the database circuit. In this context, we believe that the company is expected to reach the target value of equity incentives in 24-25. The company's main revenue consists of software products and technical services, and other businesses account for a relatively small share. Combined with equity incentive targets, we expect the company's software product and technical service revenue growth rates to be 35%, 47.80%, and 49.23% respectively in 24-26; the company's revenue growth rates will be 33.1%, 44.6%, and 45.8%, respectively. The gross margin of the company's software products and technical services continued to increase, with 71.99%, 72.99%, and 73.99% respectively in 24-26; the company's overall gross margin was 61.1%, 63.1%, and 65.1%.

We expect the company's revenue for 2024-2026 to be 0.653/0.945/1.378 billion yuan, net profit to mother of -0.232/-0.174/0.006 billion yuan, respectively, and corresponding EPS of -1.92/-1.44/0.05 yuan respectively. Based on the current industry stage and the company's growth rate and barriers, the company was given 8-10 times the 2024 dynamic PSs, with a 6-month reasonable value range of 43.25-54.07 yuan, giving it a “superior to the market” rating.

Risk warning: Product iteration is less rapid than expected; market competition intensifies; accounts receivable account is relatively high.