This article is from "McKinsey China Financial Industry CEO Quarterly", and the copyright belongs to McKinsey. This quarterly magazine mainly focuses on the theme of generative AI (hereinafter referred to as "GenAI"). Through 4 chapters and a total of 8 articles, it comprehensively and in-depth analyzes GenAI's impact on major industries, value chain investment opportunities, China's GenAI market status and future trends, and How companies can deploy GenAI to truly tap its value.
With the popularity of ChatGPT, GenAI has become a topic of concern and hot discussion in all walks of life. Global technology giants and AI manufacturers have withdrawn one after another, fearing to miss this technology feast. Industry leaders and media compare the GenAI wave to the mobile Internet opportunities of the past, believing that it will have a profound impact on the global economy and various industries, and that enterprises will also usher in major transformation opportunities.
GenAI is booming. The scale of the industry is also growing rapidly, and investors have entered the industry one after another. GenAI market revenue will be US$40 billion in 2022, and is expected to reach US$399 billion and US$1,304 billion in 2027 and 2032 respectively, with a compound growth rate of 42% from 2022 to 2032.
The Chinese market will be worth approximately RMB 66 billion in 2022, with a compound growth rate of 84% from 2020 to 2025. In 2025, the Chinese GenAI market will account for 14% of the global market size (US$217 billion).
Since the content is quite large and totals 178 pages, the following mainly introduces the current situation and future trends of China’s GenAI market.
Market size
The overall revenue of the global GenAI market will be US$40 billion in 2022, and is expected to reach US$399 billion and US$1.3 trillion in 2027 and 2032 respectively, with a compound growth rate of 42% from 2022 to 2032.
By 2025, GenAI will account for 10% of all data generated (up from just 1% in 2021).
It is expected that the increase in global market size from 2022 to 2035 will mainly come from training side hardware, advertising applications and software. Among them, the increase in training-side hardware is approximately US$444 billion, and the increase in GenAI infrastructure services (GenAI as a service) alone is US$244.8 billion, with a compound annual growth rate of 60%;
In terms of advertising applications, the related annual compound growth rate will reach 125%, with an increase of US$192.4 billion; and in terms of software, of the approximately US$280 billion increment, GenAI assistant software will have a considerable increase, reaching US$89 billion, with an annual compound growth rate of US$89 billion. The growth rate is expected to reach 70%.
China's GenAI market size will be approximately RMB 66 billion in 2022. It is expected to have a compound growth rate of 84% from 2020 to 2025, and will account for 13% of the global market size (US$217 billion) by 2025.
At the same time, the popularity of the domestic GenAI financing market has continued to rise in the past two years, with the total scale of the top ten financing events reaching US$870 million.
Among them, leading companies such as a leading end-to-end AI pharmaceutical company and a large model startup company have received more than US$200 million in single financing, and the cumulative scale of multiple rounds of financing has exceeded US$300 million and US$250 million respectively.
technology stack
(1) Special hardware: Infrastructure resources are limited, but the localization process is accelerating.
Due to the inability to obtain high-end GPU (graphics processing unit) chips A100 and H100, domestic large model R&D institutions have encountered a bottleneck in computing power. In order to cope with the above difficulties, China's local chip R&D manufacturers have strengthened scientific research and achieved important breakthroughs in computing power technology.
Currently, the A800 GPU chip available in the Chinese market has a transmission speed of only 70% of the top product A100. Since the development of AI technology is highly dependent on advanced GPUs, it requires the use of chips for a large number of model training and expansion. Some domestic scientific research institutions and technology companies have been greatly affected by this, especially in the current situation where the parameters of large models are rapidly expanding and the demand for computing power is significantly increasing. Down.
In order to cope with the computing power bottleneck, leading domestic computing power chip companies strive to achieve breakthroughs on the product side. For example, a domestic mass-produced commercial artificial intelligence chip manufacturer has implemented some business scenarios in a knowledge-enhanced large language model product owned by a leading cloud service provider.
A high-tech company specializing in the development and sales of general-purpose GPU chips and solutions and another computer graphics chip design company hope to develop products with performance close to high-end chips.
The artificial intelligence processor owned by the world's leading provider of ICT (information and communications) infrastructure and smart terminals is used for internal large model development and training.
At present, the computing power of domestic GPUs is mostly within 1000TOPS. The above-mentioned computer graphics chip design company has been able to reach 2000TOPS, but there is still a certain gap between the computing power of the internationally leading H100 chip and the 4000TOPS.
(2) Cloud platform: The global competition landscape is converging. In the future, domestic GenAI manufacturers will rely on cloud platforms to complete model training and adjustment.
In the field of cloud platforms, both domestic and overseas markets present a monopoly pattern in which leading manufacturers occupy an absolutely dominant position. In 2022, the four leading domestic manufacturers will account for 79%4 of the domestic market share; in overseas markets, as of the first quarter of 2023, the top three manufacturers will account for 65% of the global market share.
General large models require massive data for training, but the number of Chinese websites globally accounts for only 1.4% (English websites account for 54%). Public Chinese corpora (including text, pictures, videos, etc.) can be used for training materials) are often limited in quantity and of uneven quality.
At the same time, most of the massive data generated by domestic users on websites and mobile apps has not been used in large model training due to the protection of user privacy, which has adversely affected the efficiency and accuracy of model training.
Despite many limitations in training data, China's general large model technology continues to catch up with the international leading level, and the parameter scale keeps pace with the international leading level and achieves rapid improvement.
Compared with general large models suitable for multiple fields and tasks, industry large models rely more on developers’ understanding of vertical scenarios and massive industry data support. In the context of limited computing power, industry large models are more likely to become The winning track for Chinese GenAI manufacturers.
As far as industry large models are concerned, the domestic market is showing a differentiated competition pattern: the industry large model of a leading Internet company is rooted in multimedia fields such as games, advertising, and content creation, and has become a productivity tool to improve the efficiency of advertising and game scene creation;
The industrial-level knowledge-enhanced large-scale model developed by the top three ultra-large-scale cloud service providers in China relies on its strong understanding of the Chinese language to empower after-sales service and knowledge base construction in the automotive industry, document identification in the medical industry, and copyright protection and protection in the social science industry. Entry management, etc.;
A global leading provider of ICT (information and communications) infrastructure and smart terminals has large Chinese models including NLP, CV, multi-modal and scientific computing that are widely used in industrial logistics, new drug research and development, weather forecasting and other fields.
(3) Basic models: There are a considerable number of large general models, and large industry models may be the way out for domestic GenAI.
In terms of general large models, as of May 2023, China has released 79 large models with more than 1 billion parameters, ranking second in the world in terms of the number of large model releases, second only to the United States. Among the top ten GenAI large model manufacturers in the world in terms of the number of models released, Chinese R&D institutions and manufacturers occupy 4 seats.
General large models require massive data for training, but the number of Chinese websites globally accounts for only 1.4% (English websites account for 54%). Public Chinese corpora (including text, pictures, videos, etc.) can be used for training materials) are often limited in quantity and of uneven quality.
At the same time, most of the massive data generated by domestic users on websites and mobile apps has not been used in large model training due to the protection of user privacy, which has adversely affected the efficiency and accuracy of model training.
Despite many limitations in training data, China's general large model technology continues to catch up with the international leading level, and the parameter scale keeps pace with the international leading level and achieves rapid improvement.
(4) Model libraries and tools: Focusing on open source models, independent manufacturers of large model tools have emerged at home and abroad.
In the overseas market, a number of independent vendors have emerged around open source models that can freely adjust the code, and can realize segmented technical functions such as model training and fine-tuning, model deployment, and model application development. As the domestic GenAI market continues to develop, various development and maintenance tools will gradually mature.
(5) Application: Chinese startup companies are still in their early stages, and their focus areas are relatively concentrated.
Startups in China's GenAI field have advanced financing rounds, concentrated between angel rounds and Pre-A rounds. Most of the financing amounts are less than 100 million yuan. They are in the early stages of development and have huge industry potential.
Among them, an artificial intelligence chatbot owned by a world-leading multinational technology company, as a quasi-unicorn enterprise, completed a 1 billion yuan A+ round of financing in November 2022, with a post-investment valuation of approximately US$1 billion.
In contrast, in the overseas GenAI industry, scientific and technological innovation companies are relatively large in scale, and many unicorn companies have emerged in niche application fields. Among them, a global AI research company headquartered in London is valued at US$3.8 billion.
China's GenAI vertical application fields mainly focus on text, image and audio and video generation. In overseas markets, a large number of GenAI-based development platforms, data analysis platforms and code writing platforms have emerged outside the above fields. The reason is that overseas use of early programming languages (such as COBOL) There are many systems to be written, and many companies face high programming labor costs, so they have a high demand for programming assistance software.
At the same time, the current target customers of Chinese GenAI companies are mostly end users, while there are large-scale enterprise customer groups in overseas markets. As China’s SaaS market matures and companies’ willingness to pay increases, it is expected that domestic text generation and image generation start-ups will rapidly expand to enterprises. customer market.
Commercial application
The commercial application of China's GenAI industry shows two major characteristics: first, the industry distribution is concentrated, mainly concentrated in advantageous industries with relatively mature commercial development; second, most companies are still in the early stages of actively exploring their own business models.
(1) China’s GenAI applications focus on advantageous industries with mature commercial development
Chinese GenAI companies take advantage of domestic vertical scenarios to prioritize the application of GenAI in emerging industries with relatively complete commercial infrastructure; while their European and American counterparts use the mature local SaaS market to apply GenAI to high-tech, communications and various traditional industries ( Such as medical care, education, etc.), with a wider range of applications.
In China, the fastest growing fields of GenAI applications include e-commerce, media, entertainment and games, especially digital virtual humans and e-commerce video marketing, while most traditional industries (such as finance, energy, education, etc.) are still in small-scale pilot projects stage.
GenAI applications are able to flourish on Internet e-commerce platforms because China has a high-quality e-commerce and supply chain ecosystem as well as a large consumer base, which provides opportunities for the implementation of GenAI applications.
Typical industry application cases include a generative AI 3D short video content manufacturer in the video industry, a leading GenAI video large model R&D company in the e-commerce industry, and a game and AI research and application institution under an Internet platform in the game industry.
Among the above-mentioned companies, generative AI 3D short video content manufacturers can generate 3D video content based on text and promote it through short video platforms such as Douyin, Kuaishou, and Bilibili, which greatly improves the creative efficiency of short video content producers;
The GenAI video large model R&D company is deeply involved in the e-commerce industry, using AI to generate virtual human anchors to empower various industries and brands of e-commerce; game and AI research and application institutions use artificial intelligence to create AI companions and AI competitive robots, and Use natural language technology to give non-player characters personalities so that they can interact and talk with players in the game.
(2) The GenAI business model in the Chinese market is still being explored
At present, most GenAI startups in the Chinese market have just completed the output of standardized products and have begun to enter the preliminary commercialization exploration stage. The mainstream business models in the market include cloud resource sales, model API calls, SaaS charging, material charging, etc.
As domestic enterprises are not willing to pay for software, the market needs to be further cultivated, and enterprises have data security concerns about SaaS deployment methods, the business model for large-scale application of GenAI still needs to be explored.
In the European and American markets, the SaaS paid subscription model has basically matured, and a number of overseas GenAI companies have built sustainable SaaS business models.
Risk Management
While GenAI empowers all walks of life, it also brings negative impacts and challenges such as damage to fairness, infringement of intellectual property rights, information leakage, malicious use, security threats, model illusions and third-party risks. Among them, the three major risks of model illusion, malicious use, and information leakage deserve special attention.
Malicious use: Improper use of model output may bring negative impacts such as fraud and the spread of false information. China's GenAI has rich applications and a large user base, and the risk of malicious use is more prominent. However, due to the country's lack of real-name authentication, content compliance, and Implementing measures such as combating false information will help trace the source of malicious use, thereby preventing and controlling related risks.
Information leakage: Private or sensitive information used for model training may be used as output content by GenAI applications. At the same time, personal information in the artificial intelligence model storage system may also face cyber attacks and leaks, so data needs to be deployed internally and under strong supervision. Key industries and enterprises under the epidemic situation need to proactively prevent the risk of information leakage.
Security threats: Not only does responsible AI (RAI-Responsible AI) receive regulatory and corporate attention, but safe AI (SAI-Securing AI) has also gradually entered the field of vision of companies and the public, especially the full life cycle security protection of AI systems. Effective AI threat management directly affects the prevention of information leakage and malicious use. Recently, the National Information Security Standardization Technical Committee has drafted relevant evaluation specifications and standards, providing algorithm providers with life cycle security guidance for machine learning algorithms7, and can also provide reference for regulatory assessments.
Model Illusion: Model output does not conform to reality. Model hallucination is common in all GenAI models and applications at home and abroad. However, because there is still room for improvement in the research and development experience and technical strength of domestic large models, and the Chinese corpus is more complex than English, the problem of model hallucination has become more difficult.
In order to standardize the application of GenAI, China has begun to lay out AI-related supervision as early as 2022. For example, in March 2022, it released the Internet Information Service Algorithm Recommendation Management Regulations, and in November 2022, it released the Internet Information Service In-depth Synthesis Management Regulations.
In July 2023, the Cyberspace Administration of China jointly issued the "Interim Measures for the Management of Generative Artificial Intelligence Services" in conjunction with 7 departments including the National Development and Reform Commission, the Ministry of Education, and the Ministry of Science and Technology. Clear provisions have been made on the protection of personal privacy and portrait rights, and the industry regulatory framework has been further improved.
Looking overseas, the National Institute of Standards and Technology released an AI risk management framework in January 2023, focusing on providing risk control best practices for AI development, governance, and operations. The European Union passed the authorization draft of the Artificial Intelligence Act in June 2023, dividing the responsibilities and obligations of different levels of artificial intelligence, achieving layered governance through risk classification, and striving to promote cross-industry artificial intelligence legislation. , to further prevent the multiple risks that artificial intelligence may bring.
Conclusion
The global artificial intelligence market is currently in a stage of rapid growth. This article provides an in-depth and simple analysis of the development status and trends of the GenAI industry at home and abroad from the four dimensions of market size, technology stack, commercial application and risk management.
Although there are currently large differences between domestic and foreign markets in areas such as large model development, application layout, and business models, we see that China’s GenAI industry is constantly catching up with the international leading level, and it is expected that China’s GenAI related technologies and applications will gradually mature in the future. , and further explore a business model suitable for its own development.
Introduction of the four authors
Qu Xiangjun is a senior partner of McKinsey & Company, head of financial institutions consulting practice in China, and is based in the Hong Kong branch;
Han Feng is a global managing partner of McKinsey and Company, based in the Shenzhen branch;
Hu Yirong is an associate partner of McKinsey and Company, based in the Shanghai branch;
Wang Zhechen is a consultant at McKinsey & Company, based in the Shanghai branch.