Shared todayAI seriesIn-depth research report: "AI large model industry in the first half of 2024 Investment strategy: The trend of multi-modalization of large models is significant, and the AI application side is accelerating and prosperous》.
(Report produced by: Dongguan Securities)
Total report: 30 pages
1. Media industry market and performance review
1.1 Industry trends diverge in the first and second half of the year
Shenwan Media ranked second in terms of growth during the year. In the first half of 2023, benefiting from multiple favorable events such as the continued optimization of epidemic control and the rapid iteration of large AI models, the Shenwan Media Index rose in stages. After reaching a high point in June, affected by factors such as AI application implementation falling short of expectations and weak market trading sentiment, the industry entered a relatively deep correction stage. In late October, segmented hot topics such as AI education, real-person interactive video games, and micro-short dramas boosted the overall mood of the media, and the media index rebounded in November. Overall, as of November 30, the 2023 Shenwan Media Industry Index has increased by 20.10%, ranking second among the 31 Shenwan Primary Industry Indexes, outperforming the Shanghai and Shenzhen 300 Index in the same period by 29.80 percentage points. From the perspective of industry segments, as of November 30, the six secondary sub-industries of Shenwan Media have increased across the board in 2023, all outperforming the Shanghai and Shenzhen 300 Index in the same period. Among them, the publishing industry performed outstandingly, with a cumulative increase of 41.72% this year, ranking first among all sub-industries.
The current valuation is below the recent 5-year average, with some room for upside. In terms of valuation, according to iFind data, as of November 30, 2023, the PE TTM (excluding negative values, the same below) of the SW media industry was 24.87 times, which is at the 43% percentile of the past five years and lower than that of the past five years. The annual average is 6.89%, which has a certain room for growth. In terms of sub-sectors, the PE TTM of the SW gaming sector is 26.05 times, the PE TTM of the SW advertising and marketing sector is 28.78 times, the PE TTM of the SW film and television theater sector is 51.73 times, the PE TTM of the SW digital media sector is 48.54 times, and the PE TTM of the SW publishing sector is 48.54 times. 14.95 times, SW TV broadcasting sector PE TTM is 32.19 times.
1.2 The performance of the media industry in the first three quarters has recovered to improve, and the growth rate has improved compared with H1
The performance of the media industry has recovered significantly in the first three quarters, and the growth rate has increased compared with H1. With the marginal improvement of the content review environment and the recovery of macroeconomics and consumption, the prosperity of the media industry has rapidly rebounded in the first three quarters of 2023, and industry performance has continued to recover. In the first three quarters, industry revenue and profits have improved year-on-year, and the performance growth rate is higher than that of H1 promoted. On the revenue side, under the holistic method, the operating income of the media industry in the first three quarters of 2023 was 362.961 billion yuan, a year-on-year increase of 5.73%, and the growth rate was 1.03pct higher than that of H1. On the profit side, the industry's net profit and deductions attributable to the parent company in the first three quarters of 23 Non-post-owned net profits were 34.354 billion yuan and 29.557 billion yuan respectively, a year-on-year increase of 21.90% and 21.51% respectively. The growth rate was 7.27pct and 9.43pct respectively higher than that of H1. The industry's profit side increased significantly in the first three quarters, mainly due to factors such as the increase in investment income from gains and losses from changes in the industry's fair value and the decrease in industry asset impairment losses. In terms of profitability, the industry's gross profit margin in the first three quarters of 23 was 32.47%, a year-on-year increase of 0.85pct; the net profit margin was 9.78%, a year-on-year increase of 1.36pct, and the industry's profitability has improved.
In the first three quarters, our ability to control expenses has been enhanced, and operating cash flow has increased significantly. On the expense side, in terms of period expenses, the industry's period expenses in the first three quarters of 2023 totaled 70.532 billion yuan, a year-on-year increase of 5.51%; the period expense rate was 19.43%, a slight year-on-year decrease of 0.04pct. In terms of R&D expenses, the industry's R&D expenses in the first three quarters of 2023 were 11.528 billion yuan, a year-on-year decrease of 1.26%; the R&D expense rate was 3.18%, a year-on-year decrease of 0.22pct. In terms of cash flow, the industry's net operating cash flow increased significantly in the first three quarters of 2023, increasing from 30.901 billion yuan in the first three quarters of 22 to 41.131 billion yuan in the first three quarters of 23, a year-on-year increase of 33.11%. In the first three quarters of 2023, the industry's net operating cash flow/operating income ratio was 11.33%, an increase of 2.33ct compared with the same period in 2022, and the industry's income liquidity has been enhanced.
Among the sub-industries, film, advertising and marketing, games, publishing, and digital media have outstanding performance. In terms of industry segmentation, based on Shenwan's secondary industry classification standards, we eliminated ST companies and combined the company's business conditions to separately classify seven major industries: games, movies, dramas, publishing, advertising and marketing, television broadcasting, and digital media. segmented areas. In the first three quarters of 2023, the performance of the film, advertising and marketing, games, publishing, and digital media industries was outstanding, with revenue and net profit attributable to shareholders both achieving positive year-on-year growth. Among them, the film industry recovered strongly driven by the supply of high-quality content, and revenue Achieved double-digit growth, and net profit attributable to the parent company achieved triple-digit growth.
Performance ranking of the first three quarters of 2023 by segment:
Sorted by year-on-year growth in operating income: Movies (+46.13%) > Advertising and Marketing (+5.61%) > Games (+4.51%) > Publishing (+4.14%) > Digital Media (+3.14%) > Television Broadcasts (-1.43%) > Episodes (-7.73%).
Sorted by year-on-year growth rate of net profit attributable to parent companies: Movies (+236.44%) > TV Series (+33.28%) > Advertising and Marketing (+15.93%) > Publishing (+13.56%) > Games (+7.95%) > ; Digital Media (+6.76%) > Television Broadcasting (-31.1%).
2. The rapid development of multi-modal models promotes the accelerated prosperity of the application layer
ChatGPT has set off a global AI large model arms race. In November 2022, penAI launched ChatGPT, a new chat robot based on GPT-3.5. It quickly gained widespread attention from the market with its excellent natural language interaction capabilities and powerful content creation and other functions. The number of users exceeded 100 million after 2 months of launch. The emergence of ChatGPT led a new wave of generative AI technology development. Subsequently, domestic and foreign technology giants such as Google, Meta, Baidu, 360, and Alibaba have invested in starting an AI large model arms race. According to the "China Talent Intelligent Large Model Map Research Report" released by the China Institute of Science and Technology Information, as of the end of May 2023, a total of 202 large models have been released globally, including 79 large models with more than 1 billion parameters released domestically. Ranking second only to the United States, the total number of large models in China and the United States accounts for nearly 90% of the global number of large models. China currently ranks among the top in the world in terms of the number of large models.
The development trend of large model multi-modality is significant, opening up a wide range of application scenarios. In March 2023, OpenAI launched the GPT-4 large model, which supports the input of different data forms such as graphics and text, achieving multi-modal development, and the answer accuracy is significantly improved compared to GPT-3.5. In September, OpenAI officially released the multi-modal model GPT-4V, which added voice and image interaction functions and continued to make efforts in the multi-modal direction. In terms of the image interaction function that attracts more attention, in the official case of OpenAI, the user uploaded a picture of a bicycle and asked ChatGPT how to lower the seat. During the process, he used a drawing tool to circle a specific position and asked for details. In the end, ChatGPT understood the user's multiple rounds of questions. The text and picture content and detailed operation steps are given, demonstrating the powerful multi-modal understanding ability of GPT-4V. Other domestic and foreign technology giants are also rushing to enter the multi-modal field, including Google's PaLM-E, Meta's AnyMAL, Microsoft's KOSMOS-1, Baidu's Wen Xinyiyan, Alibaba's M6 large model, etc. Large model multi-modality The trend of cultural development is significant. From the perspective of model technology, single-modal models are mainly aimed at the input of one kind of data. They have limitations and incompleteness in terms of interaction, and their applications are usually limited to a single data type task. Multimodal models can process two or more data inputs at the same time, capture complex cross-modal data relationships, comprehensively understand and analyze information, and thus participate in more complex and deeper tasks, with broader application scenarios.
With the rapid upgrade of multi-modal model capabilities, games, publishing, film and television, and marketing are all expected to benefit. In terms of application progress, domestic large models have been applied in many fields such as finance, office, entertainment (including game development, marketing, film and television production, etc.), education, medical care, transportation, etc. From the perspective of industry penetration, according to the "2023 Comparative Study of AI Large Model Applications in China and the United States" released by Titanium Media, the penetration of large models in the entertainment and education fields exceeds 50%, ranking third and fourth respectively. For financial and government application scenarios. Specifically, there are currently two major development directions in the integration exploration of generative AI technology based on multi-modal models and application scenarios in the media field: On the one hand, AIGC can empower content production to reduce costs and increase efficiency, in games, publishing, film and television , advertising marketing and other related fields involving text, pictures, audio and other related fields to bring innovations in application experience. On the other hand, in terms of products, it is expected to open up the commercial value space of software and hardware AI products by combining it with AI technology. For example, there are already related application products in the field of education with strong willingness to pay, including smart education in software. Intelligent learning machines in terms of platform and hardware, etc. With the rapid iteration of multi-modal large models and the continuous upgrading of multi-modal capabilities, the generative AI industry space is expected to accelerate, and the gaming, publishing, film and television, and marketing fields of the media industry are all expected to benefit.
3. Games: The supply side continues to release, AI helps reduce costs, increase efficiency and innovate gameplay
Version numbers have resumed normal issuance, and the game supply side is expected to continue to improve. According to the National Press and Publication Administration, as of November, a total of 873 domestic game versions have been issued in 2023, far exceeding the number issued throughout the previous year (468). Judging from the rhythm of domestic game version number issuance, except for August, which is a blank period for version number issuance, the remaining months are steadily issued, and the number of each batch is 86 or more. In terms of imported game version numbers, 27 and 31 imported game version numbers were issued in March and August 2023 respectively, for a total of 58 games. Compared with the 44 games issued in 2022, the number has also increased significantly. As the issuance of game version numbers returns to normal, the game supply side is expected to continue to improve.
The size of the game market is showing a growth trend. With the normalization of version number issuance, the number of new games online has increased and the pace has accelerated. The number of new games entering the iOS best-selling list has also made breakthroughs, and new games have maintained good performance. According to Gamma data, in August 2023, a total of 23 games entered the iOS daily best-selling list TOP200, which is the month with the largest number of products on the list so far in 2021. Benefiting from factors such as the continued development of new games and the maintenance of high-quality operations of existing games to increase user activity, the overall size of the domestic game market is showing a growth trend. From January to August 2023, the domestic game market size was 202.1 billion yuan, a year-on-year increase of 7.16%; among which, the market size in August was 29.226 billion yuan, a year-on-year increase of 46.08%, and the year-on-year growth rate increased by 8.59pct compared with the previous month. At present, there are still many games that have obtained licenses but have not been launched. After subsequent launches, it is expected to boost the overall prosperity of the industry, promote the further expansion of the game market, and increase the performance of related companies.
The game integrates text, voice, images, videos and other content forms, and is expected to make better use of multi-modal capabilities to empower the three major aspects of game production and production, marketing reach and player experience optimization, and achieve cost reduction and efficiency improvement. At the same time, it brings innovative experiences to players, thus opening up space for growth in the game industry.
Production and production links: In terms of game planning and programming, the text generation model is applied in basic code writing, world view framework construction, and gameplay plan planning. It can help R&D personnel simplify repetitive work, save more conception time, and thus more More energy is invested in core research and development links. In terms of art design, the current image generation model performs well in intent drawing and conceptual design. Art designers can enter text prompts in models such as Midjourney and convert them into visual images in a few seconds, helping art designers quickly Browse different drawing styles to draw inspiration, which greatly reduces early design costs and time. For example, Giant Network previously cooperated with Alibaba Cloud artificial intelligence platform PAI to build the AI painting platform iMagine internally. After applying AI painting, the company as a whole has achieved 50%-70% cost savings in original paintings of characters and scenes. In terms of UI and There is more than 80% efficiency optimization in icons, and AI application efficiency gains are considerable.
Marketing reach link: Traditional information collection and public opinion analysis require manual collection of massive information through crawlers and other methods, and then interpret the content after complicated analysis and processing processes, which has relatively high threshold and cost. By applying AI models and connecting to the Internet through plug-ins and other methods, it is possible to collect, analyze and interpret multi-channel and multi-lingual player information with high efficiency, and quickly gain insight into market trends. In addition, AI technology can be used in the purchasing and delivery process, greatly optimizing the entire process of "buying material production - delivery effect analysis - purchasing and delivery execution" to achieve cost reduction and efficiency improvement.
Game experience link: AIGC is expected to accelerate the implementation of game play innovation and intelligent interaction applications. By enriching game content and play methods, it can further enhance players’ immersion and realism in the game. At present, a representative case in China is the application of intelligent NPC system in the "Nishuihan Mobile Game" released by NetEase in June. In this game, NPCs no longer wait for players to trigger fixed lines at a specific location. Intelligent NPCs have independent personalities and behavioral logic, and can even actively "chat" players for chat conversations, and give reasonable answers by understanding the players' natural language input. feedback to provide players with a more realistic interactive experience. In addition, "Ni Shui Han" has also added "AI poetry writing", "intelligent face pinching" and other gameplay methods, which significantly enhances the richness and freedom of the game.
4. Publishing: Steady performance growth, Al+ Education opens up new value space
4.1 Steady performance growth, cash reserves, high dividend ratio and dividend rate
The book retail market recovered in the first three quarters and the short video channel maintained positive year-on-year growth. From the perspective of industry data, according to open book information statistics, the scale of the book retail market in the first three quarters of 2023 fell by 1.04% year-on-year, compared with 6.55% in the first quarter and 2.41% in the first half of the year. The decline continued to narrow, and the market recovery trend is improving. . Looking at offline sales from different sales channels, the sales volume of physical stores in the first three quarters of 2023 dropped by 22.56% year-on-year, a slight narrowing of 0.99 percentage points compared with the first half of the year, and the decline was still significant. In terms of online sales, platform e-commerce and vertical and other e-commerce sales in the first three quarters of 2023 maintained negative year-on-year growth, down 6.13% and 9.76% respectively; emerging channel short video e-commerce is still the highland for traffic gathering, maintaining a year-on-year growth rate of 55.5% Positive growth, the growth rate expanded by 8.14 percentage points compared with the first half of the year.
Industry performance grew steadily. Judging from industry performance, due to the relatively rigid market demand in the publishing industry, industry performance has remained stable despite the impact of the epidemic. In 2020, the industry's performance suffered a certain decline due to the epidemic. From 2021 to 2022, the industry's revenue returned to positive year-on-year growth, and the profit returned to a level higher than that of 2019 before the epidemic. In the first three quarters of 2023, the operating income of the publishing industry was 103.477 billion yuan, a year-on-year increase of 4.14%; the net profit attributable to the parent company was 12.794 billion yuan, a year-on-year increase of 13.56%. The industry achieved double growth in revenue and net profit, and the operations of enterprises in the industry continued previous stability.
The cash flow situation of the industry is good and the cash reserves are sufficient. In terms of cash flow, in the past five years, the average operating cash flow of the publishing industry has shown a net inflow, and the average net operating cash flow of the industry is higher than the overall average level of the media industry. The cash flow situation of the publishing industry is good. In terms of monetary funds, since 2019, the average monetary capital of the industry has been above 2.5 billion yuan, and has shown a trend of increasing year by year. In the past five years, the average monetary capital of the industry has also been higher than the overall average level of the media industry.
The dividend ratio and dividend rate are relatively high. In terms of dividend ratio, the industry average cash dividend ratio has shown an overall upward trend in the past five years, from 35.87% in 2018 to 46.31% in 2022. Among them, in 2022, the cash dividend ratios of New Classic, Central South Media and Guangdong Media are all higher than 70%. The higher dividend ratio reflects that companies in the industry have stable cash flow and good profitability, and have a strong willingness to give back to investors. In terms of dividend rate, the industry average dividend rate also shows an increasing trend, increasing from 2.10% in 2018 to 2.55% in 2022. Among them, in 2022, the dividend yields of Chinese Media and Central South Media are both higher than 6%. A higher dividend rate means that investors can get more cash flow returns, and companies in the industry have good investment value.
4.2 Model training requires massive amounts of data, and the value of high-quality data in the publishing industry is highlighted.
Large models accelerate iteration, and the demand for high-quality data continues to grow. The training of large models requires massive data support. According to the "Artificial Intelligence White Paper" released by the China Academy of Information and Communications Technology in 2022, the number of parameters and data scale of current model training are growing at a trend of 300 times per year. In addition to quantity, the quality of pre-training data is also a key factor affecting the performance of large models. At present, Internet data content is complex and the overall data quality is relatively low. Simply collected raw data needs to be professionally cleaned before it can be used for model training. The cleaning process is not only costly, but also causes a waste of computing power. Directly collecting high-quality data can improve the accuracy of the model and enhance the generalization ability and stability of the model while saving cleaning costs and reducing model training time. With the iterative development of large models, the demand for high-quality data will continue to grow.
High-quality Chinese data content is scarce. The current "pain point" in the development of the Chinese version of large models is that there are fewer Chinese databases that can provide training. For example, according to the language-specific proportion results of GPT3 training data previously disclosed by OpenAI on Github, English words account for as high as 92.65% in the training data set, while Chinese words account for only 0.1%. Chinese public data resources are far less than those in English. In the context of the rapid development of large models, Chinese data content has become an important and scarce resource.
Book publishing companies have rich and high-quality Chinese data resources, and the industry is expected to usher in a revaluation. The data sources currently used for large model training include web pages, books, encyclopedias, etc., among which books account for a large proportion. As the domestic publishing industry has a long history, a large audience, and the publishing process requires strict “three reviews and three proofreadings”, domestic publishing companies have accumulated a large number of systematic and high-quality Chinese data resources. A representative company is China Science and Technology. On the one hand, the company holds 15% of the shares of Wanfang Database, one of the three major Chinese academic databases in China. Wanfang Database covers a large number of text data resources such as journals, academic papers, and conference minutes. Can be used for model training. On the other hand, the company relies on the Chinese Academy of Sciences to publish a large number of scientific and technological books and journals. Due to their strong professionalism, scientific and technological books can provide more rigorous data resources for model training, which is conducive to improving the effect of model training. With the emergence of domestic large-scale models and the support of national policies for high-quality data, the value of high-quality Chinese data resources owned by domestic book publishing companies is expected to be fully released, and the publishing industry is expected to usher in a reshaping of valuations.
4.3 AI reshapes the publishing industry and opens the second growth curve of “AI + education”
AI technology can be applied to all aspects of the traditional publishing industry to help save production costs and time, and improve content production efficiency and quality.
1) In the topic selection and planning process, editors used to spend a lot of time manually collecting data from various fields and platforms, cleaning and integrating the data, and then making analysis and judgment. Nowadays, AI technology can be used to quickly analyze and process massive data, gain an in-depth understanding of market trends and readers’ needs, and assist publishing companies to more accurately predict market trends and make topic selection decisions.
2) In the content editing process, in terms of text, the emergence of natural language processing models such as Chatgpt can provide relevant writing frameworks and references according to instructions, improving the author's creative efficiency. In terms of pictures, image generation models such as Midjourney can provide support and assistance in illustrations and cover design.
3) In the content proofreading process, traditional proofreading requires a lot of manual work. Now the use of AI proofreading tools can automatically detect grammar, spelling errors, etc. in the article and provide modification suggestions, which can improve the quality and accuracy of the text while reducing labor costs.
4) In the marketing and promotion process, AI tools can help quickly analyze user behavior data, accurately identify potential target customers, and promote based on their behavior and interests to improve marketing efficiency.
The demand for AI education is strong and its implementation is accelerating in overseas regions. Due to the strong adaptability of AI to education and the high willingness of superimposed users to pay for such services, AI education has become one of the fastest-growing scenarios for AIGC, especially in overseas regions. Several online education and learning platforms in the United States have been connected to the GPT-4 model, including language learning platforms Duolingo and Speak, teaching platform Khan Academy, online homework tutoring platform Chegg, etc. These platforms use AI technology to develop personalized learning strategies to help improve student learning efficiency. From the perspective of demand, according to similarweb statistics, educational AI application traffic increased significantly month-on-month in September, mainly driven by strong downstream demand during the school season. Among them, the monthly visits of CheggMate, an AI learning service platform launched by Chegg Company, increased significantly by 110.8% month-on-month, and the monthly visits of Q-Chat, an AI education tool launched by Quizlet Company, increased by 88.2% month-on-month. With the subsequent acceleration of AI model iterations and the continuous development of downstream intelligent hardware equipped with AI functions, the demand for AI education is expected to further increase.
There are three main types of enterprises in China engaged in the field of "AI + education", namely local publishing state-owned enterprises, education and training enterprises, and education informatization enterprises.
1) Local publishing state-owned enterprises: With multiple advantages such as content, channels, and funds, the layout of "AI + education" is expected to open up the second growth curve. In terms of content, local publishing state-owned enterprises have a large number of high-quality educational texts, images, videos and other resources, providing rich content data for the development of AI education. In terms of channels, publishing groups, especially local publishing state-owned enterprises, have the qualifications to publish and distribute local textbooks and teaching aids. Over the years, they have established close cooperative relationships with local education departments and schools, accumulated stable channel resources, and can effectively assist AI. Publicity and promotion of educational products. In terms of funding, most publishing groups have strong financial strength and stable cash flow, and can afford capital expenditures in research and development and promotion of AI education products. At present, among domestic local publishing companies, Phoenix Media, Zhongnan Media, Wanxin Media, Southern Media, etc. have a large presence in the field of AI education. The main product forms include: intelligent hardware equipped with AI functions (such as: Wanxin Media and Huawei jointly inclusive learning machine), smart education service platforms (such as: "Guangdong Education Xiangyun Platform" created by Southern Media) and related AI education products (such as: Guangdong Education iFlytek E Listening jointly created by Southern Media and iFlytek Say) etc. Local publishing state-owned enterprises are expected to rely on their own advantages to take the lead in the field of AI education, expand the commercial value space of software and hardware AI products, and open up the second growth curve.
2) Education and training enterprises: They have certain education, teaching and research content resources and R&D technical talent reserves, and have strong productization capabilities. They mainly focus on the extracurricular education market. For example, TAL launched its self-developed 100-billion-level large-scale model “MathGPT” in the field of mathematics based on its experience in K12 education and training in August. The core function of this product is problem solving and algorithm explanation. Users only need to upload math questions through text or pictures to get conversational answer feedback.
3) Educational informatization enterprises: Focus on providing intelligent educational software and hardware products to schools and educational institutions, and have accumulated certain channel resources in the early stage. For example, iFlytek released the T20 series of iFlytek AI learning machines equipped with the Spark cognitive large model in May this year. In October, it released version 3.0 of the Spark cognitive large model. Version 3.0 added new heuristic dialogue capabilities and learning machine capabilities. AI Q&A has been deeply integrated.
Total report: 30 pages