Progress of AIPC side cooperation: On May 9, 2024, at the 2024 HP Commercial AI Strategy and AIPC New Product Launch Conference, Starlink Technology and HP released three Z series AI workstations, focusing on the AI knowledge base, smart vision and large model service platforms. The strong combination of software and hardware empowers AIPC.

[TKH helps enterprise users work efficiently]: In order to help enterprise-level users analyze various documents safely and efficiently and quickly build knowledge question and answer applications, Starlink Technology has deployed the knowledge platform Transwarp Knowledge Hub on HP Z series workstations. Through TKH's accurate knowledge management and powerful computing power, it can meet the needs of different types of knowledge applications within enterprises in various fields such as finance, energy, manufacturing, and engineering, thereby effectively improving enterprise business efficiency and competitiveness.

[AI tools help enterprises quickly implement large models]: Starlink Technology and HP Z series AI workstations have jointly launched a large model application and development service all-in-one machine. The solution combines the product capabilities of the Starlink big model development and application platform Sophon LLMOps, and can provide full-link tool capabilities including detailed processing and storage of corpus knowledge, corpus data labeling, large model fine-tuning training, big model evaluation and deployment to large-scale model application construction and application deployment. Through private deployment within the enterprise, the security of corporate private data can be guaranteed. The cost-effective solution allows small and medium-sized enterprises to have a convenient and efficient full-process experience of large model development and application implementation.

Major overseas model upgrades bring industry prospects: On May 14, Beijing time, OpenAI launched GPT-4O at the spring press conference, and also released the desktop version of ChatGPT. The significant increase in the ease of use of GPT-4o also marks the continued progress of AI large model applications towards more natural human-computer interaction, and end-side intelligence is expected to be significantly upgraded.

[Smooth for real-time voice conversation] The GPT-4o model can perform real-time inference in audio, visual, and text, and can input or output any combination of text, audio, and images. The model can respond to audio input in as little as 232 ms (an average of 320 ms), and the response speed is close to that of a human conversation.

[Anthropomorphic emotion interpretation, emotional value filling] GPT-4o can detect the speaker's emotions and give real-time feedback. At the same time, the model can recognize real-time video content to guide users to solve equations and complete tasks such as interpretation, code analysis, and chart understanding.

[Desktop version of ChatGPT creates efficient work] GPT-4O can correctly understand the code and fully describe it, open the application directly on the desktop, and understand, refine, and summarize the input content.

We think:

1. Represented by GPT-4o, large models at home and abroad are expected to achieve double iteration in terms of model performance and user interaction experience for some time to come. In segmented fields, the commercial side may prioritize the effective implementation of functions such as AI assistants.

2. Model upgrades can effectively improve the end-side experience and effectiveness. AIPC, as a powerful breakthrough on the end side, can optimize performance by learning user behavior patterns, and provide more accurate information retrieval, content recommendation, and automated task processing to help users live and work.

3. AIPC requires large end-side model products, as well as vector databases to transform content such as text, images, and videos into high-dimensional vector analysis and processing.

Starlink Technology provides:

1) Distributed vector database Hippo

2) Large model operation management software Sophon LLMops

3) End-side big model products: enterprise-level vertical question and answer big model “endless” and its derived big data analysis model “search”.

4) TKH (TKH has comprehensive, efficient, and intelligent data processing and knowledge management tools, and can provide enterprise users with key capabilities such as multimodal knowledge construction, multimodal knowledge storage and service, knowledge engineering, knowledge rights management, and knowledge question and answer assistants. (The platform has a built-in knowledge question-and-answer application built on Starlink's self-developed model - Wuyai·Knowledge, which supports users to upload multi-source data such as private documents, tables, images, audio and video, and perform automated knowledge engineering to ensure that data such as corporate documents and personal experience documents are not out of the domain. At the same time, it also supports connecting to external databases or real-time information to build a secure and intelligent enterprise private domain model.) By the end of 2023, orders for Big Model Era AI Infra related products provided by the company exceeded 54 million yuan.

Profit forecasting and valuation

We expect the company to achieve operating income of 6.62/8.72/1,078 billion yuan in 2024 to 2026, a year-on-year growth rate of 34.84%/31.72%/23.72%; net profit to mother in 2024-2026 will be -123/0.62 billion yuan, respectively, corresponding to 8.41 times P/S in 2024. Maintain a “buy” rating.

Risk warning

1. Risk of AI technology iteration falling short of expectations; 2. Release of AI commercial products falling short of expectations; 3. Risk caused by policy uncertainty; 4. Risk caused by downstream market uncertainty.