Key points of investment:

24H1 performance: Although revenue has not started, the asset side may indicate growth potential (1) Profitability: Affected by adverse factors in the supply chain such as the “list of entities”, the company 24H1 achieved operating income of 64.7653 million yuan, YOY -43.42%. The realized net profit to mother was -0.53 billion yuan, a year-on-year decrease of 14.719 million yuan, and the loss narrowed by 2.70%. (2) Balance sheet: By the end of 24Q2, inventory reached 0.235 billion yuan, an increase of 83% month-on-month, and prepayments of 0.55 billion yuan, an increase of 169% month-on-month. We expect the company to have sufficient supply and is expected to release significant results in 24H2. We look forward to driving the company's revenue to a new level after launching new products.

The client progresses faster and gradually establishes customer advantages

In the Internet field, 24H1 products have been shipped in batches in natural language application scenarios. The company's current measurement capabilities and iterative expectations of chip products meet customer needs.

In the field of big models, the company helped leading domestic algorithm companies explore and implement big model applications in their respective vertical fields. (1) In the field of large visual models, the company has reached strategic cooperation with leading customers to provide solid computing power support for the online commercial application of large visual models. (2) In the field of large language models, the company's products were adapted to large models with leading customers and obtained multiple product compatibility certifications with this customer, which met the customer's demand for artificial intelligence computing power and accelerated the entire process of the customer's large model from training to deployment.

Ecological construction continues to be strengthened to adapt to mainstream AI models

In terms of ecology, the Cambrian training software platform was updated and iterated on PyTorch 2.1/2.3 support, supporting multiple training and reasoning business scenarios, and implementing a long-term mechanism to quickly follow up on community versions. Furthermore, the Cambrian training software platform implemented native support for MLU from the Transformers and Accelerate communities, added the Cambrian backend of Triton 3.0.x, and open-sourced the cross-platform AI compiler front-end Triton-Linalg, improving the development efficiency of adapting the industry's artificial intelligence applications to the company's products. In terms of AIGC business adaptation, the company's reasoning software platform follows the application needs and development of the industry, supports and optimizes a series of literary and biological models such as the Llama series and the Qwen series, as well as various mainstream multi-modal models such as StableDiffusion3, achieving coverage of mainstream AIGC business scenarios such as Wenshengwen, Wensheng map, Wensheng video, and multi-modality, and promoting the expansion of the company's application ecosystem.

Profit forecasting and investment advice

We believe that the asset side may indicate sufficient supply for the company, and 24H2 is expected to drive performance growth.

We forecast the company's 2024-2026 revenue of 1.482/2.28/3.497 billion yuan (previous value:

1.299/2.017/3.206 billion yuan), corresponding to the current P/S multiple 72/47/31 times. As a top priority for domestic replacement, AI computing power has broad scope for future development, and the company is expected to continue to benefit as a domestic AI chip leader. Maintain a “buy” rating.

Risk warning

Risk of AI demand falling short of expectations, risk of high customer concentration, and risk associated with supply chain stability.