Large model technology has led the field of artificial intelligence to a new level of development and has received widespread attention around the world. Large models are an important development opportunity for enterprise users and artificial intelligence manufacturers.
Recently, AiAnalysis has observed that large models are no longer limited to the scope of technical discussions, but have entered the stage of commercial application. Therefore, AiAnalysis conducted in-depth research on issues of concern to both supply and demand parties such as the large model market size, business model, and commercialization progress, and looked into the future trends of the large model market.
Key findings
- China's large model market will usher in a commercial explosion in 2024, with the market size12 billion yuan.
- In the early stages of large-scale model implementation, commercialization mainly comes from the model layer. Large model + computing power is the most mainstream charging method. As the large model ecosystem matures, the application layer will play the main role in the commercialization process in the future.
- The speed at which large models are implemented varies significantly across industries.The two major industries that have invested the most aggressive budgets are energy and finance. The main reason is that these two industries are dominated by central state-owned enterprises and have strong data, computing power and AI foundation.
- With the in-depth development of large-scale models, the application scenarios and ecology of large-scale models have grown rapidly, and many new needs have arisen. Market research has revealed preliminary clues, includingLLMOps,large model all-in-one machineWait.
- The price reduction of large model services will reduce the input costs of end enterprise users and promote the inclusive application of large models.
- Open source large models represented by LLaMa 2 are an important force in the large model market. The comprehensive capabilities of open source large models are generally lower than those of closed source large models. LLaMa 2 and GPT3 are roughly equivalent in terms of comprehensive capabilities. The value of open source large models does not lie in the construction of underlying capabilities, but in the construction of upper-layer applications. Open source will reduce the development threshold and cost of upper-layer applications of large models, thereby accelerating the penetration of large model applications and expanding the commercial market space.
01Definition of research scope
Large model technology has led the field of artificial intelligence to a new level of development and has received widespread attention around the world. Large models are an important development opportunity for enterprise users and artificial intelligence manufacturers.
Recently, AiAnalysis has observed that large models are no longer limited to the scope of technical discussions, but have entered the stage of commercial application. Therefore, AiAnalysis conducted in-depth research on issues of concern to both supply and demand parties such as the large model market size, business model, and commercialization progress, and looked into the future trends of the large model market.
The large model market panoramic map is divided into four levels, namely the basic layer, the model layer, the intermediate layer and the application layer. The basic layer includes vector databases, cloud platforms and other vendors, such as Tencent Cloud and Alibaba Cloud; the model layer includes general large models, industry large models and other vendors, such as Baidu, Zhipu Huazhang; the middle layer includes large model application development tools, LLMOps and other vendors , such as LangChain, etc.; the application layer brings together product and solution providers, covering a variety of application software and terminal equipment, related manufacturers such as UFIDA, Sematic, etc.
Figure 1: Panoramic map of China’s large model market
In China, all four levels of the large model have entered the commercialization stage. Among them, the model layer and application layer have huge commercialization potential and are the focus of this report.
The market size in this report refers to the budget amount on the enterprise user side, and its investment includes three components: hardware, software and services. This report focuses on the analysis of software and services.
In terms of business model, it refers to how the products and services of large model manufacturers are sold to corporate users, including charging methods and pricing.
In terms of commercialization progress, it refers to the specific situation of large-scale model commercialization, including enterprise user industry distribution, purchasing tendencies and application scenarios, etc.
02Progress in commercialization of China’s large model market
2.1 China’s large model market size is expected to reach 12 billion yuan in 2024
AiAnalysis estimates based on public bidding winning information and first-hand survey data that China's large model market size will be approximately 5 billion yuan (RMB, the same below) in 2023. Among them, the pure hardware procurement part accounts for about 65%, the service part accounts for about 20%, and the software part accounts for about 15%.
In 2023, enterprise users' procurement of large models will be characterized by more demonstrations and less procurement, and most of the budgets will not exceed one million. Enterprise users will begin to release a large number of large model budgets in 2024. Large models are planned to account for about 10% of the AI budget, and the budget scale is mostly in the millions of yuan.
Ai Analysis estimates that China’s large model market will reach 12 billion yuan in 2024. In terms of composition, the pure hardware procurement segment will decline, while the software and services segment will increase. Specifically, the pure hardware procurement part accounts for about 60%, the service part accounts for about 23%, and the software part accounts for about 17%.
Figure 2: China’s large model market size and composition in 2023 and 2024
2.2 “Large model + computing power” is currently the most mainstream charging method
There are three types of charging methods in the large model market, namely large model, large model + computing power, and large model + application.
1) Large models: Using large models as sales targets is the simplest commercialization method. Enterprise users can directly buy out large model products permanently, or rent large model services.
2) Large model + computing power: The training and application of large models require a lot of computing power, so some manufacturers sell large model products or development platforms in combination with computing power.
3) Large model + application: Vendors sell upper-layer applications that incorporate large model capabilities to enterprise users. Vendors may require enterprise users to pay software licensing fees or require them to pay software development service fees.
AiAnalysis has compiled statistics on large model bidding winning messages from the beginning of 2023 to early August 2023. There are 60 pieces in total. The data shows that large models + computing power are the most mainstream charging methods, accounting for 62.3% of the total amount.
Note: The statistical caliber here is different from the market size mentioned above.
Figure 3: Proportion of the amount of three types of charging methods in the large model market
Case 1: In May 2023, Chongqing University of Posts and Telecommunications released bidding information. The subject was a large model training all-in-one machine, and the project budget was 2 million yuan. Brief description of procurement content: Artificial intelligence development platform adopts container + image management model, which can realize end-to-end full-process development, integrates large AI training models and image libraries, and supports centralized management and allocation of GPU resources; integrated computing power platform configuration fourth generation The Xeon CPU platform has NVIDIA optimization certification, the computing unit CPU is no less than 80 cores, 8 TESLA H800 80G GPU cards, and an independent management login unit.
In the large model market, there are significant price differences between different charging methods. The average project amount for large models + computing power is the highest, exceeding 10 million yuan. The price of large model + application is the lowest, even lower than that of large model. The reason is that this charging method usually means calling the large model API, and the unit price per customer is low.
Figure 4: Average prices of different charging methods in the large model market
Note: The amount in the above picture is in RMB 10,000.
2.3 The energy and financial industries are at the forefront of commercialization progress
Energy and finance are the two industries with the largest amounts in the large model market.
The enterprise user industries of large models are widely distributed, including energy, finance, education, etc. The data shows that the energy and finance industries accounted for the highest proportions, accounting for 40.9% and 16.9% respectively. The reason why the energy and finance industries have become the leading industries in large-scale model construction is mainly due to the dense distribution of central state-owned enterprises. Central state-owned enterprises have complete data infrastructure construction, high investment in computing power, many AI application scenarios and strong foundation. These reasons promote the rapid integration of central state-owned enterprises and large models.
Case 2: In June 2023, the Industrial and Commercial Bank of China recruited potential suppliers for a procurement project on the capability research and application of knowledge-enhanced NLP large models in financial scenarios, requiring suppliers to have hundreds of billions of natural language modules that can support privatized deployment. model products.
Figure 5: Proportion of monetary contribution of different industries in the large model market
Ai Analysis Research found that another application direction of large models in 2024 is the Government Affairs Bureau, and the citizen service hotline is the main scenario. Most of the relevant projects in 2023 are in the systematic testing stage. Once the testing is successful, there will be a large budget in 2024.
Data analysis is the application scenario with the fastest progress in the implementation of large models.
Large model application scenarios include data analysis, customer service, marketing, office and other application scenarios. The survey results show that data analysis is the application scenario with the fastest progress in the implementation of large models.
Figure 6: Implementation progress and potential value of large model application scenarios
Case 3: In the data analysis scenario, the operations analysis and management committee of a leading securities company put forward requirements for large models in the data analysis scenario, hoping to obtain analysis results of some indicators based on natural language. A BI software company provided a "BI + large model" product to the securities company, with a project amount of 400,000-500,000 yuan.
03 Commercialization trend of China’s large model market
Ai Analysis looks forward to the commercialization trend of China's large model market and proposes six trend perspectives, covering commercialization, leader, new demand, price, open source, overseas expansion, etc.
Figure 7: Overview of commercialization trends in China’s large model market
Trend 1: Replicating the development trend of artificial intelligence commercialization, the application layer will become the main force in the commercialization of large models in the future.
Good commercialization opportunities exist at every level of the large model ecosystem. In the early stages of large-scale model implementation, commercialization mainly comes from the model layer. As the large model ecosystem matures, the application layer will play the main role in the commercialization process in the future. This trend has gradually emerged.
Looking back at the commercial development trend of artificial intelligence in the past, the early model layer completed commercialization through model licensing and other methods. With the booming development of artificial intelligence applications, subsequent commercialization mainly relies on the application layer, and commercialization is completed through project development fees for application construction, application licensing fees, etc. Large models will replicate the same commercialization trend.
At present, with the emergence of more and more open source large models, the main trend of commercialization has emerged from the model layer to the application layer. For example, Meta released Llama 2, and Baichuan Intelligence also open sourced Baichuan-13B. The future commercialization of these large open source models mainly relies on the applications built on the models.
Trend 2: The two major industries of energy and finance are actively investing in large-scale budget implementation models, and central state-owned enterprises are bravely shouldering the pioneering mission as leaders.
The implementation speed of large models varies significantly among various industries, and the two major industries with the most aggressive investment budgets are energy and finance. The main reason is that these two industries are dominated by central state-owned enterprises and have strong data, computing power and AI foundation.
First, the data infrastructure of central state-owned enterprises is well established. Central state-owned enterprises in the energy and financial industries have always attached great importance to data capability building and even have supporting data strategies. In order to support the implementation of the strategy, the data management department has also been upgraded to a first-level department parallel to the information technology department in the organizational structure. For example, the State Grid Big Data Center is a professional support organization that supports the company's digital transformation. The person in charge of the data department is generally held by the core management of enterprise users.
Secondly, central state-owned enterprises have high investment in computing power. Large models consume huge amounts of computing power. Especially when manufacturers need to deploy large models privatly, the initial computing power investment reaches tens of millions. Central state-owned enterprises such as finance and energy have sufficient digital budgets and are not limited in investing in large-scale model computing power. For example, China Southern Network invested more than 30 million yuan in large model computing power in 2023.
Third, central state-owned enterprises have many AI application scenarios and a strong foundation. In the early stage of the implementation of large models, the focus was on enhancing the original AI scenarios, and there were a large number of AI implementation scenarios in central and state-owned enterprises. ICBC's AI application penetration scenarios have exceeded 1,000, and large models can quickly find feasible scenarios from them.
In addition to energy and finance, Panmut is also one of the target industries of the large model market. Fanhu's main payment method is to call large model APIs to improve customer experience through capability internalization.
Trend 3: With the development of the large model market, a series of new demands such as LLMOps and large model all-in-one machines are about to emerge, providing broad prospects for further exploring commercialization opportunities.
With the in-depth development of large model implementation, large model application scenarios have grown rapidly, and many new demands have arisen. Market research has revealed preliminary clues, including LLMOps, large-model all-in-one machines, etc.
There are many application scenarios for large models within enterprise users, and different scenarios require the use of different large model capabilities. Therefore, enterprise users need the ability to call multiple large models at the same time, which in turn creates the need for unified management and unified operation and maintenance. Looking forward to the future, large enterprise users and government departments will introduce LLMOps one after another to purchase multiple large models while adding a management platform.
As large models gradually mature, enterprises need to train large models and develop upper-layer applications more conveniently and efficiently. This is the main reason for the rise of large model all-in-one machines. Large model all-in-one machines have significant advantages in rapid deployment, convenient management, and improved efficiency. The large model all-in-one machine integrates underlying GPU, storage and network resources and is specially designed for upper-layer large model pre-training or inference applications to achieve rapid deployment and convenient management, thereby providing support for efficient promotion and stable operation of artificial intelligence applications. In addition, the large-model all-in-one machine can improve the coordination of software and hardware, improve training and inference efficiency, and reduce the consumption of computing resources. At present, several large-model all-in-one machines have been commercialized. In June 2023, Zhongke Wenge released a large model all-in-one machine called "Yayi large model all-in-one machine". In August 2023, iFlytek and Huawei jointly released a large-model software and hardware all-in-one machine called the "Spark All-in-one Machine". Looking forward to the future, the number of enterprise users of large-model all-in-one machines will continue to increase, and the corresponding product supply will also become increasingly abundant.
Trend 4: The price of large model services is gradually declining, accelerating penetration into small and medium-sized enterprise users, and promoting the vigorous development of the large model market.
The reduction in the price of large model services will reduce the investment costs of end enterprise users and promote the inclusive application of large models. In the context of the current macroeconomic downturn, one of the core issues that hinders enterprise users from investing in large model applications is ROI, of which the price of the large model service itself is an important cost item. As the price of large model services drops, more application scenarios will have the opportunity to achieve higher ROI and have investment value; at the same time, as the price of large model services drops, more small and medium-sized enterprise users will have the opportunity to become paid users of large models. Enterprise users, expand the large model commercial customer base.
At present, the price of large model services is mainly affected by cost and manufacturer's pricing strategy. It is normal for prices to continue to decline. First of all, in terms of cost, the computing power cost of the training and inference phases has gradually declined. Take Nvidia's two GPU products H100 and A100 as examples. According to public data, the computing power of H100 is about 6 times higher than that of A100, but the price is only about 3 times higher, and the cost per unit of computing power has dropped significantly. Secondly, in terms of pricing strategy, with the increase in the number of large model manufacturers and the open source trend of large models, competition in the large model market has intensified. It is expected that large model manufacturers will adopt appropriate low-price strategies in pricing strategies to accelerate the development of downstream applications and ecological development.
The risk that needs to be paid attention to in the future in terms of price is that with the development of the large model application market, there may be vicious competition at low prices in the large model application solution market, and the transformation of commercialization into customized project development based on human service charges, etc., which is not conducive to the healthy development of the market. .
Trend 5: Open source large models accelerate application penetration, and the commercialization of large models accelerates
Open source large models represented by LLaMa 2 are an important force in the large model market. The comprehensive capabilities of open source large models are generally lower than those of closed source large models. LLaMa 2 and GPT3 are roughly equivalent in terms of comprehensive capabilities.
The value of open source large models does not lie in the construction of underlying capabilities, but in the construction of upper-layer applications. Open source will reduce the development threshold and cost of upper-layer applications of large models, thereby accelerating the penetration of large model applications and expanding the commercial market space. For domestic large model manufacturers, open source will weaken the competitive barriers of technology and require more attention to comprehensive capabilities such as data, computing power, services, and ecology.
Open source large models will accelerate the penetration of large model applications, create a commercial application ecosystem based on open source, and accelerate the commercialization of large models. On the one hand, open source large models will reduce the development threshold and cost of large model applications and accelerate the penetration of downstream applications. Since open source large models can be easily obtained and used for free, large model application layer manufacturers and enterprise users with strong technical capabilities can quickly start developing industry applications based on open source large models and accelerate the implementation of large model applications. Especially since the current large model market is in its early stages of development, the selection range of large models is small and the cost is high. However, its business benefits need to be verified, and open source large models can be used for cold start and application exploration. On the other hand, manufacturers of the large model application layer can develop related products and services based on the authorization of open source large models and solve security, service and other issues for end enterprise users, achieve commercialization, and form a commercial application ecosystem based on open source.
Trend 6: The opportunities for China's large model manufacturers to go overseas are mainly in cross-border e-commerce, games, social media and other pan-entertainment fields. Progress depends on the pace of large model capabilities catching up with OpenAI.
The application of large models in overseas markets has progressed rapidly and has been commercialized. Chinese large model manufacturers also have opportunities for overseas commercialization.
There are three main opportunities for China's large model manufacturers to go overseas for commercialization:
1) Industries and scenarios that already have a good foundation for overseas expansion, such as cross-border e-commerce, games, social media and other pan-entertainment fields, are mainly 2C. In these industries, large models can find clear application scenarios and successfully implement commercialization based on existing scenarios.
Case 4: A large model manufacturer in China has developed a "conversational e-commerce" product based on the conversational interaction capabilities of large models. It provides one-stop services such as marketing interaction, shopping guide and checkout, and customer service. It has served a certain chain brand coffee.
2) Based on our capabilities in Chinese processing, we provide services in areas with certain Chinese markets such as Southeast Asia (such as Malaysia and Singapore). China's large model manufacturers have natural advantages in the Chinese language and have differentiated competitiveness.
3) Leverage the “Belt and Road” policy to provide services to the “Belt and Road” regions. Chinese manufacturers have become important suppliers to local governments in the Middle East and Central Asia. Large models will be commercialized as new product services for local governments, especially to provide privatized deployment and large model applications for government and enterprise users in these regions. Overall solution model.
When Chinese large model manufacturers go overseas, they mainly face direct competition from OpenAI. The core competition point lies in the capabilities of large models. OpenAI's large models represented by GPT 3.5 and GPT 4.0 are significantly ahead of domestic manufacturers in terms of capability performance. The current capabilities of domestic large models have basically reached the GPT 3.0 level, and are expected to reach the GPT 3.5 level by the end of 2023 to early 2024.
Specifically, China's large model manufacturers focus on catching up in AI engineering capabilities and data. AI engineering capabilities include full-stack capabilities at the model layer, framework layer, and chip layer. In terms of data, it includes public data sets, user feedback data and fine-tuning data for specific industry scenarios. Public data sets are easy to obtain, but user feedback data and fine-tuning data for specific industry scenarios depend on the number of downstream enterprise users and the scale of the application ecosystem.
04 Conclusion
The rise of China's large model industry marks a new chapter in the field of artificial intelligence. Driven by continuous innovation, it is rapidly integrating into all walks of life, bringing unprecedented opportunities to the business world. Looking to the future, we can optimistically foresee that China’s large models will continue to lead the wave of innovation.
In the future, we can look forward to the emergence of more innovative application scenarios, deeper business cooperation, and the continued evolution of technology. At the same time, international development and policy support will also bring a broader stage to China's large model industry. We firmly believe that China's large model industry will continue to contribute to the world's technological progress and commercialization process, and become one of the key forces leading the future. Let us look forward to and actively participate in this exciting era together and contribute to the future development of China's large models.