AI偏向科普性报告：围绕算法、算力、数据和应用

Equity Research·Industry Research·Computer Computer Industry Research 1 / 6 东吴证券（香港） 请务必阅读正文之后的免责声明部分 [Table_Main] Popular science report of AI: Algorithms, Computing Power, Data and Applications Outperform (Maintain) Investment Thesis ◼ As a new paradigm of AI development, big models are a milestone for artificial intelligence towards general intelligence. Big models refer to models that can adapt to a series of downstream tasks after training on large-scale general data, which are essentially based on statistical language models, but the "emergent capability" gives them powerful reasoning capabilities. The existing big model framework is essentially the same, and almost all large language models with parameter scales exceeding 100 billion adopt the GPT mode. However, different types of enterprises give their own advantages in the field they are in, and the developed big models still have differences in function. Technology plays a decisive role in the effect of big models, so the future competition pattern also depends on technological breakthroughs. ◼ Calculating power is the "oil" of the AI era. The training and inference of big models will use the calculating power support of AI chips. Under the same data and algorithms, calculating power is the key to the development of big models, and it is the "oil" of the AI era. Assuming one-month training time of GPT-3, we estimate that 843 Nvidia A100 chips are needed. We calculate that ~16255 Nvidia A100 chips are needed if the DAU of GPT-3 reach 50 million. As GPT-4 uses multimodal data, we estimate that its calculating power demand is >10 times of GPT-3. China's leading internet companies have successively invested in big models, and we expect that only ten leading factories may add ~200,000 A100 demand in one year. In the long run, the demand is expected to exceed 2 million chips, and the new calculating power demand is likely to double the size of the calculating power market. In 2021, Nvidia occupied more than 80% of the market share in the China acceleration card market, and the performance of domestic AI chips is still lower than that of foreign ones. We expect the launch of domestic big models to drive the development of domestic chips. ◼ Data resources are one of the important driving forces for the development of the AI industry. Datasets, as the core component of data resources, refer to data specially designed, collected, cleaned, labeled, and managed for the training of AI algorithm models. The performance of large-scale language models is strongly dependent on parameter scale N, dataset size D and computing power C. The training data mainly comes from Wikipedia, books, journals, Reddit social news sites, Common Crawl, and other datasets. GPT4 relies on a large amount of multimodal data for training. In the future, the competitiveness of AI models may be reflected in the quality and scarcity of data. The development of data elements market can promote the further opening of relevant public, corporate and individual data, providing important support for China's AI development. ◼ AI empowers all walks of life, and the future is promising for AI applications. AI is comparable to the fourth technological revolution, and the most direct application for this time is in the field of content creation, opening the imagination boundary of the industry. We should look for areas where the application functions are significantly improved, customer stickiness is significantly improved, and the market size is greatly increased with the empowerment of AI, mainly including content creation, office software, ERP, robots, and chip design. At present, some big model manufacturers have started industrial applications, but computing power is still the main reason for limiting large-scale commercialization of AI. Once solved, the information industry may be directly benefited from AI+, so we are optimistic for leaders in the information industry. ◼ Recommendations: Algorithms: we recommend large-model companies that have a first-mover advantage, such as 360 Security Tech, Iflytek, Hithink Royalflush. Computing power: we recommend Changsha Jingjia Micro, Changsha Jingjia Micro, Digital China. Data: we recommend leading information companies in each sub-track, such as Sichuan Jiuyuan Yinhai Software, Anhui Ronds Science&Tech, and Zhejiang Supcon Tech. Applications: we recommend companies with "killer" application potential, such as Beijing Kingsoft Office Software, Yonyou Network Tech, and Hundsun Tech. ◼ Risks：Lower-than-expected policy implementation, Intensified industry competition. [Table_PicQuote] Industry performance [Table_Report] Related reports 《华为盘古大模型产业链梳理》 2023-03-27 《数据安全，为数据要素市场发展保驾护航》 2023-03-24 Soochow Securities International Brokerage Limited would like to acknowledge the contribution and support provided by Soochow Research Institute, and in part