As the number of visits to large models such as Kimi continues to increase, the length of long text increases, and the number of model parameters and training tokens increases, the demand for inference computing power and training computing power is expected to continue to increase. In the wave of computing power, localization of computing power is a long-term trend, and demand is expected to continue to rise.
The Zhitong Finance App learned that Orient Securities released a research report saying that with the popularity of the big model led by Kimi, the number of visits, etc. continues to rise, it is imperative for various companies to expand their inference computing power. Also, as the length of the context of long texts increases, the consumption of computing power will surely increase. With the continuous iterative upgrading of large models such as Kimi, the number of model parameters and the amount of training data are expected to continue to increase, and the scaling law will continue to work to drive computing power demand. Domestic AI chip manufacturers, domestic AI chip line and Nvidia line server manufacturers, computing power service providers such as computing power leasing, IDC, and server operation and maintenance are all expected to continue to benefit.
Related targets:
Domestic computing power: It is recommended to focus on Haiguang Information (688041.SH), Cambrian-U (688256.SH), Hi-Tech Development (000628.SZ), Shenzhou Digital (000034.SZ), Firefly Communications (600498.SH), Softcom Power (301236.SZ), Radio and Television Transport (002152.SZ), etc.
Nvidia Hashrate Chain: It is recommended to pay attention to Wave Information (000977.SZ), Ziguang Co., Ltd. (000938.SZ), etc.
Computing power service providers: It is recommended to pay attention to Yunsai Zhilian (600602.SH), Runze Technology (300442.SZ), Yakang Co., Ltd. (301085.SZ), Unicode-W (688158.SH), Capital Online (300846.SZ), Shenzhen Sanda A (000032.SZ), Tongniu Information (300895.SZ), Huatie Emergency (603300.SH), etc.
The main views of Orient Securities are as follows:
Kimi's traffic continues to increase, and the Dark Side of the Moon has expanded five times and the Kimi website is still down
Orient Securities believes that demand for computing power is expected to continue to increase as the volume of large model traffic increases and the length of long text context increases. This week, Kimi's announcement of support for 2 million words of ultra-long, lossless context attracted attention. The number of visits, app downloads, and daily usage is impressive. On the afternoon of March 21, Kimi's app, applet, and web versions all showed that they were not working properly. The company said that since observing an abnormally high increase in traffic on March 20, 5 such expansion work has been carried out, and inference resources will continue to cooperate with the expansion to carry out the continuous increase in the number of users. Meanwhile, Baidu and Ali are preparing to open up long text capabilities. As large model traffic increases and the length of long text context increases, the demand for computing power is expected to continue to increase.
Step Star released a preview version of the STEP-2 trillion parameter MoE language model. The trillion parameter model places extremely high demands on computing power
On March 23, Step Star released the Step series of general-purpose large models, including the Step-100 billion parameter language large model, the Step-1V 100 billion parameter multi-modal large model, and a preview version of the Step-2 trillion parameter MoE language large model, providing an API interface for some of its partners to try out. According to Step Star, models with large trillions of parameters require at least a single cluster of 8 million cards equivalent to A8 million cards and efficient and stable training. Orient Securities believes that demand for computing power is expected to increase further as large-scale models such as Step Star continues to rise along with Scalinglaw's trillion-parameter and multi-modal large models.
Shanghai announces 2024-2025 computing power Pujiang Plan, setting targets such as the scale of intelligent computing and localization of intelligent computing
On March 21, 11 departments of Shanghai jointly issued the “Implementation Plan for the High-Quality Development of Intelligent Computing Infrastructure in Shanghai “Computing Power Pujiang” Intelligent Computing Action (2024-2025). The plan sets a number of intelligent computing construction goals: by 2025, the scale of intelligent computing power in Shanghai will exceed 30 eFlops, accounting for more than 50% of the total computing power; new intelligent computing domestic computing chips will account for more than 50%, etc. The government continues to actively promote the construction of intelligent computing and the localization of computing power, and the domestic computing power line is worth paying attention to.
Risk warning:
The development of industries related to the AI model falls short of the expected risks; the implementation of policies related to computing power falls short of the expected risks; the development of computing power falls short of the expected risks.