Research Report Direct Access -- 2.3

1. Market Overview [Taogu Ba]

On February 3rd, the A-share market experienced a volatile upward trend, with the three major indices collectively strengthening. By the close, the Shanghai Composite Index rose 1.29% to close at 4067 points, the Shenzhen Component Index increased by 2.19%, and the ChiNext Index gained 1.86%. The total trading volume across both markets was 2.54 trillion yuan, a slight decrease compared to yesterday. Throughout the day, 4856 stocks rose, while 532 declined.

Today’s rebound may largely stem from a mood recovery after the sharp decline in the previous trading session. Given that trading volume has not significantly increased, the market may still harbor doubts about short-term recovery. If external commodity prices continue to fluctuate, funds might flow back into defensive and low-volatility assets. Additionally, there is cautiousness toward cyclical sectors that were more deeply impacted earlier. Furthermore, the phase of announcements and earnings disclosures is becoming more intensive, and market attention to earnings realization or the exhaustion of negative news has significantly increased. The previously strong sectors may need to be supported by actual performance to withstand volatility.

2. Rockets Can Do More Than Launch Satellites; Space Mining Is Not a Dream!

Event: Recently, China Aerospace Science and Technology Corporation announced that during the “14th Five-Year Plan” period, it will carry out major project demonstrations for “Tiangong Kaiwu,” focusing on breakthroughs in key technologies such as small celestial body resource exploration, intelligent autonomous mining, and low-cost transfer and transportation. This move directly catalyzes the entire commercial space chain: significantly increasing demand for heavy-lift rockets and low-cost transportation; providing clear application scenarios and scalable possibilities for space photovoltaic technology in mining bases; and also promoting the development of related frontier tracks such as in-orbit manufacturing and processing, deep space communication, accelerating the commercialization process of space resource development.

3. Cambrian: Major Opportunities for Domestic Computing Power Under the Intense Competition of Large Tech Giants

The company clarifies related rumors: R&D progress is smooth, and operations are steadily advancing. 1. The information circulating online about the company organizing small-scale exchanges recently is false. 2. The company has never issued any annual or quarterly revenue guidance data. 3. Currently, R&D is progressing smoothly, and operations are steadily advancing.

Q4 has established a performance growth inflection point. If we exclude impacts such as bonus provisions, the actual business growth rate is expected to be even higher.

The biggest beneficiaries of increased AI infrastructure investment by large tech companies: Tencent, Alibaba, and Baidu have all issued bonds for debt financing, totaling over $30 billion, which are expected to be used to expand overseas cloud computing businesses. Alibaba is ramping up investment in AI and cloud computing, with future three-year investments possibly increasing from 380 billion yuan to 480 billion yuan. As a leading domestic computing power company, Cambrian leads in hardware performance and software ecosystem, with significant revenue growth validating its market position. It is expected to fully benefit from the domestic AI infrastructure boom.

3. AI Applications

On January 13th, DS published a paper about the Engram architecture. ⁠ ⁠ To put it simply, the core points of this paper are: ⁠ ⁠ By optimizing model architecture design (software), the hardware can achieve: ⁠ 1. Transferring 20%~25% of GPU computing power to CPU calculations ⁠ 2. Moving 20%~25% of GPU memory (HBM) to system memory (GDDR/DDR/LPDDR) for calculations ⁠ 3. The result of this approach is that the model’s accuracy is almost unaffected, and may even improve. ⁠ ⁠ CPUs are much cheaper than GPUs, ⁠ and system memory is much cheaper than GPU memory; ⁠ transferring 20%~25% of GPU resources means performance is increased by 20%~25%. ⁠ The overhead for CPUs is minimal, ⁠ while additional costs for system memory are estimated at 5%~10% (HBM is much more expensive than DDR). ⁠ The key point of this model scheme is that it can be used for both cloud AI model inference and edge AI model inference.