Aofei Data released reports for the first quarter of 2022 and 2023. In 2022, the company achieved revenue of 1,097 million yuan, a decrease of 8.9% over the previous year; the net profit of the mother was 166 million yuan, an increase of 14.58%; after deducting the net profit of the non-return mother was 109 million yuan, a decrease of 28.8% over the previous year. Looking at a single quarter, 4Q22 achieved revenue of 271 million yuan, a year-on-year decrease of 3.5% and a decrease of 2.7% from the previous quarter; Guimu's net profit was 54.53 million yuan, an increase of 157.5% over the previous year and an increase of 49.6% over the previous quarter. In 1Q23, the company achieved revenue of 320 million yuan, an increase of 18.8% over the previous year; achieved net profit of 0.4 billion yuan, an increase of 4.5% over the previous year; and realized deduction of non-net profit of 42 million yuan, an increase of 36.5% over the previous year.
Intelligent computing centers are the general trend. In the past, the computer industry assumed that “digital processing will become cheaper,” but considering the increasing complexity and competitiveness of research, Christopher Manning, deputy director of the Stanford Institute for Artificial Intelligence, said that training costs for cutting-edge models are still rising. We believe that AI computing systems are facing challenges such as optimized design of computing platforms, computational efficiency in complex heterogeneous environments, high degree of parallelism and expansion of computing frameworks, and computational performance of AI applications. The development of computing power will pose a greater challenge to the overall computing demand, and it is urgent to improve the efficiency of the entire AI computing system.
The intelligent computing center is expected to drive an investment of 10 billion yuan and more than 100,000 high-power cabinets. The computing power of a single intelligent computing center is about 2.69 petaFLOP/s. According to our estimates, the training computing power requirement for a single large model is about 280.67 PetaFLOP/s. We think “big models with small models” will become a trend in the future. Assuming that the demand for computing power increases 50 times (many small second- and third-tier Internet companies connect to big model API training), the total computing power demand expands to 14033.5PetaFLOP/S. At this point, about 5217 intelligent computing centers (5,2170 cabinets) are needed. However, since the above calculation is based on 39 kW of power, the density is usually 2 to 4 times that of an ordinary high-power cabinet, and if the amplification factor is 2, 104,000 cabinets are required.
Based on the calculation of the number of GPUs required for ChatGPT, nearly 75,000 high-power cabinets will be needed in the future.
According to Titanium Media, GPT 3.5 requires nearly 30,000 A100 chips for complete training, corresponding to nearly 3,750 Nvidia DGX A100 servers. Considering that the maximum power of a single Nvidia DGX A100 server is 6.5 kw, assuming an average power of 4 kw/unit, the 10kw power cabinet can carry 2.5 AI servers. To meet 3,750 AI servers, 1,500 cabinets need to be built. Assuming that computing power requirements expand 50 times, a total of about 75,000 high-power cabinets are needed.
Investment advice: We are optimistic that demand from downstream enterprises will recover in the context of the “Eastern and Western Computing” +AI large-scale parameter model. We believe that with the smooth implementation of the company's multi-location data center expansion, the diversified customer structure is also expected to enhance the large-scale effect of the company's revenue and bring about a more stable operating cash flow. It is estimated that from 23 to 25, the company's revenue was 1,47/ 20.0/2.78 billion yuan, and the EBITDA was 5.4/7.8/1.08 billion yuan, corresponding to the closing price of EV/EBITDA 25/18/13x on May 4, and net profit of 2.1/33/ 550 million yuan, corresponding to the closing price of P/E 48/30/18x on May 4. First coverage, giving a “recommended” rating.
Risk warning: Market competition heightens risks, project delivery risks, and overseas business expansion falls short of expectations.