(Report Producer/Analyst: Caitong Securities, Cheng Bing, She Weichao)
1. Machine tool faucet, force data pre-training and reshaping growth momentum
Huizhou Intelligent is the leading traditional machine tool in mainland China. The company's machine tool business started, and its holding subsidiary Qi Zhong is the main body of high-end equipment manufacturing business such as machine tools, and after 73 years of development, it has become a leading traditional machine tool enterprise in mainland China.
In order to meet the AI wave, we will focus on the AI large model data pre-training business and reshape new growth momentum. In order to meet the wave of AI development, the company actively deploys the AI large model data pre-training business through acquisitions. In 2019, it acquired Changhua Culture and obtained control of Rere Culture, as the main operator of AI data and training business, and at the same time, Zhongke Huizhou Digital Business, a holding subsidiary, as the main body of technology research and development of AI data pre-training business.
1.1 Consolidate the foundation of heavy industry, and machine tools are the ballast stone for development
The product matrix of the company's machine tool business is perfect. Founded in 1950, the company is one of the key construction projects in the first five-year period of the country, with a rich high-end manufacturing heritage. After 73 years of accumulation, it has become an important CNC machine tool production base in mainland China. Up to now, a machine tool production base of 380,000 square meters has been formed, including 10 categories, 26 series, and more than 600 varieties of machine tool product matrix.
The company's many technologies make up for the gap at home and abroad. Up to now, the company has more than 400 products with independent property rights to fill the domestic gap, vertical lathe processing diameter can be as small as 0.5 meters, up to 25 meters to fill the international gap, horizontal lathe processing diameter can be as small as 1 meter, up to 6.3 meters to fill the international gap. It is widely used in downstream industries such as ships, automobiles, and wind power.
"Qiyi" brand has significant advantages. "Qiyi" CNC machine tool is a well-known machine tool brand in China, and some high-end machine tool products have been successfully exported to more than 30 countries and regions such as Europe, America, Japan and South Korea, with a market share of 40% to 50% for heavy-duty lathes and 100% for heavy-duty deep hole drilling and boring machines as of 2023H1.
We will continue to expand R&D expenditure, build a "4+3+N" innovation system, and consolidate R&D advantages. After 73 years of development, the company has rich experience in the research and development of machine tool products. In recent years, the company has made great efforts to build a "4+3+N" innovation system: building 4 scientific and technological innovation platforms, cultivating 3 scientific and technological innovation teams, introducing N scientific research institutes and enterprises, high platform R&D capabilities, and consolidating R&D advantages. As of 2023H1, the company is the national inspection standard setting unit for heavy-duty horizontal lathes, heavy-duty deep hole drilling and boring machines, and heavy-duty vertical lathes in mainland China, and has presided over and participated in the formulation and revision of 77 standards and 302 patents.
The company's experienced technical team continues to expand R&D spending. By the end of 2022, the number of R&D personnel in the company was 187, an increase of 15% compared with 2021, of which 57% were over 40 years old, and most of them had many years of R&D experience. The company continued to expand R&D expenditure, with R&D expenditure of 64.84 million yuan in 2022, a year-on-year increase of 56.05%, and R&D expenditure accounted for 9.23% of revenue, an increase of 4.27pct year-on-year.
Since 2020, the company has made two-way efforts to closely follow the changes in downstream demand and realize the comprehensive upgrade of machine tool products.
The company keeps up with the changes in demand vertically and expands to downstream popular industries. In order to undertake the release of demand from the downstream wind power industry, the company has quickly developed a series of products such as CNC special horizontal lathes, fixed beam vertical lathes, hard lathes and gear hobbing machines for wind power spindles, flanges and slewing bearings, and the new contract value of wind power industry products will account for more than 60% in 2022.
The company horizontally closely follows the technology trend of the industry and upgrades to high-end products. In 2022, the company will focus on the research and development of high-precision machine tools suitable for downstream aerospace, shipbuilding, and energy industries, and complete 26 new high-end product designs, 147 technical preparations, 190 electrical designs, and 221 process designs.
As of the end of 2022, the company still has 8 high-end product research projects. Among them, the ultrasonic micro-forging assisted laser additive manufacturing project is expected to make up for the domestic technology gap, the intelligent machine tool research project based on 5G communication is expected to help the intelligent transformation and upgrading of the company's products, and the CNC heavy-duty horizontal boring machine is leading in China for the inner hole of the wind power spindle, with a wide range of downstream application prospects.
1.2 Focus on AI data pre-training to reshape growth momentum
The company has benefited from the long slopes and thick snow of the large model track, relying on three core barriers: rich technology and product matrix, high-quality customers with high viscosity, and experienced core technical team, and has quickly become a rookie AI data service provider.
The company has laid out the data pre-training business and reshaped new development momentum. In 2019, the company acquired 100% of the equity of Xuzhou Changhua, obtained control over Rere Culture, built an AI data pre-training business entity, and simultaneously set up a holding subsidiary, Renmin Zhongke Digital Commerce, as the main body of AI data pre-training research and development, and undertook the release of downstream data labeling needs in advance of R&D layout.
The company's data business matrix is divided into two major sections: Internet content review and AI data annotation, and data annotation products and technologies have entered the forefront of the industry.
Up to now, the company has realized the annotation of text, picture, voice, video and other data types, including computer vision-related video tracking, dotting, and continuous frame technology, 2D image semantic segmentation, 3D point cloud fusion, continuous frame and other data annotation technologies, which are suitable for downstream automatic driving, humanoid robots and other popular terminal scenarios, and the precipitated resaleable standard datasets cover text, pictures, audio, and video. Develop solutions for urban planning, medical imaging diagnosis and other scenarios.
The content moderation business leads the industry in terms of personnel scale, content coverage, and accuracy. Founded in 2016, Rere Culture has set up five large-scale audit and labeling bases in Beijing, Chengdu, Suihua, Zaozhuang and Jinhua, with a review team of more than 5,000 people, a total of more than 50,000 audit talents, and a total of 2 billion audit data.
The company's core team has its own experience in the AI data pre-training industry, and has a rich background to lay a solid R&D advantage. The R&D entity of data pre-training, Zhongke Huizhou Digital Business, People's Daily Online + Beijing Zidong Science and Technology Center of the Chinese Academy of Sciences, and other parties have taken the lead in establishing a R&D platform, with their own AI data technology background and industrial experience.
Li Gang, general manager of the business entity, is the former technical director of Alibaba Cloud's Internet Business Division, and Li Bing, the company's supervisor and technical team leader, is a doctoral supervisor of the Institute of Automation of the Chinese Academy of Sciences and the chief scientist of the People's Academy of Sciences.
The company has formed a multi-level and highly sticky customer matrix. The company has successfully established a multi-level and high-stickiness customer matrix relying on high-quality services, covering downstream Internet, artificial intelligence technology enterprises, aerospace, heavy industry, finance, government affairs, universities and other industries.
2 Scenario-driven, AI pre-training data is expected to grow
2.1 Pre-training data is the cornerstone of the AI industry chain
Data is the cornerstone of the upstream of the AI industry chain. Looking at the current AI industry chain, upstream data comes from the collection of terminal scenarios, which is the starting point for algorithms to perceive the world, midstream model development is a tool for data application, downstream scenarios drive model algorithm iteration, and supervision is a rigid need throughout the upstream and downstream links.
According to a study by AI analysis company Congnilytica, the data processing process in AI projects accounts for 80% of the time, of which data annotation accounts for 25%, and the effective preprocessing of data in complex scenarios can shorten the time cycle of data identification, integration, enhancement, cleaning, and annotation, and save costs for model development.
Throughout the AI data chain, AI data pre-training is the key to connecting the previous and the next.
The pre-trained data service takes over the upstream data source: Unstructured data can only activate its value after it has been pre-trained. The AI pre-training data service collects, cleans, annotates, and inspects unstructured data such as voice, images, text, video, and point clouds on the scene side to form an effective pre-training dataset that can be directly used by AI models.
Pre-trained data helps reduce costs and increase efficiency throughout the life cycle of downstream algorithm design, training, evaluation, and iteration.
(1) In the algorithm design process, a small batch of pre-training data is used to verify the preliminary design of the algorithm to reduce the directional bias of the model design.
(2) Algorithm training, effectively simplifying the scale of model parameters and saving algorithm development time.
(3) In the algorithm evaluation process, a small amount of manually annotated pre-training data can be used as the control group of the model output results to effectively evaluate the accuracy of the model.
(4) In the algorithm iteration process, the pre-training data is accurately labeled and processed for bugs, and then the model is accurately repaired to effectively improve the model performance.
2.2 The trend of multimodality is significant, which increases the importance of data pre-training
Multimodal pre-training data is the key to solving the long-tail problem of AI applications, and the migration of multimodal technology brought about by the trend of industrial integration of large models and vertical fields will further enhance the importance of pre-training data.
The multimodal trend of models at home and abroad is significant, and the input data has developed from massive language information and text information to multi-modal data in multiple vertical fields. The bottom layer of the overlay model is to establish the key features of different modal data, such as text, voice, video, image and other data, through the understanding of instructions, and establish multi-dimensional mapping. Therefore, the model training and optimization process requires massive amounts of multimodal data. Data pre-training solves the pain points of difficult to effectively identify multimodal data and make in-depth use of semantic information under the combination of industries by extracting, aligning and fusing cross-modal features of unstructured multimodal data.
The model tends to iterate on optimization in a specific direction, and the focus of competition shifts from parameter size to data quality.
With the combination of large models and vertical industries, more models may adopt a similar strong chemistry Xi model to optimize and iterate in specific fields or specific directions, so in the model pre-training and fine-tuning links, high-quality annotated instruction data is the basis for model accuracy and generalization ability.
The focus of competition ranges from the competition of parameter size to the competition of data quality. For example, in the early days of large models, the mainstream view was that parameter size was the core element of model effect enhancement, and the larger the model parameters, the better the performance, but this view is gradually being broken. For example, the parameter scale of Llama-13B is 1/13 of that of GPT-3, and the final performance of common sense reasoning, closed-book Q&A, and reading comprehension is slightly better than GPT-3 depending on the scale of model training data.
2.3 Scenario implementation drives the release of data pre-training requirements
The evolution of models tends to be the supremacy of algorithm functions, data has become an important driving force for scenario implementation, and the demand for AI pre-training data is growing rapidly worldwide. In the early days of ChatGPT, the model pre-training data was the historical stock data as of May 2019, and the annotation demand for new data from massive terminals is expected to be released with the combination of models in vertical fields. According to Cognilytica's forecast, the global AI training data market size will be 39.3 billion yuan in 2022E and is expected to reach 157.4 billion yuan in 2027E, with a five-year compound growth rate of 31.98% from 2022E to 2027E.
Intelligent driving is the application scenario with the largest elasticity in data pre-training in the next five years, with a five-year compound growth rate of 37% from 2022E to 2027E. According to Deloitte's estimates, the market size of China's basic data service industry in 2022E will be 4.5 billion yuan, and it is expected to reach a maximum of 16 billion yuan in 2027E, with a five-year compound growth rate of 29% from 2022E to 2027E. Starting from terminal scenarios, terminal scenarios such as autonomous driving, smart industry, and Internet content currently account for a large market share of the data service industry.
The three factors of model iteration progress, mass production progress, and penetration rate are expected to catalyze the exponential growth of data processing demand.
(1) In the process of vehicle iteration, different sensor configurations require basic data service providers to customize different data solutions.
(2) The mass production progress has led to an exponential increase in the scale of data processing in terminal scenarios.
(3) The improvement of permeability deepens the complexity of scene data, and puts forward higher requirements for data labeling for the processing accuracy of multi-modal data in complex scenes.
Data drives the resonance of application and model iteration, and it is expected to reverse empower the implementation of scenarios in the future. Data pre-training is the knowledge instillation stage of the entire large model training, and the data annotation service provider provides a large amount of label data for the large model to ensure that the model truly learn Xi s the core data knowledge of the industry and further deepens the industry adaptation. High-quality pre-trained datasets are the premise of the self-feedback reinforcement Xi mechanism of large models, which accelerates the operation of large model iteration gears and realizes the resonance of scene landing and model iteration in the long term.
3. Deploy autonomous driving and quickly create industry barriers
3.1 Make efforts to intelligently annotate to help reduce costs and increase efficiency
Intelligent tools assist in annotation to help data pre-training reduce costs and increase efficiency.
Through intelligent pre-labeling, human-computer interaction-assisted annotation, intelligent quality inspection, etc., the company can reduce personnel costs and quickly improve AI model capabilities. At present, the artificial intelligence annotation crowdsourcing platform and the artificial intelligence inspection and review platform have entered the operation stage.
(1) Human-computer interaction assisted annotation: The company has launched the Enable AI intelligent annotation platform, which is a human-computer interaction mode to improve the efficiency and accuracy of data annotation in complex terminal scenarios. For example, AI can recognize the contour of the vehicle with high precision, and automatically annotate and predict the subsequent frames of continuous frame data for 3D point cloud data.
(2) Intelligent pre-labeling: For the data of simpler scenarios, firstly, the auxiliary pre-labeling model is trained by using the manually annotated small-sample data, then the model is pre-labeled with the remaining sample data, and finally the manual quality inspection is carried out.
(3) Intelligent quality inspection: Through the inspection of the AI intelligent quality inspection model, on the one hand, manual labeling errors are found, such as missing or wrong labels, to improve the quality of data labeling. On the other hand, it is necessary to locate the more difficult samples in the sample data, improve the professionalism of the quality inspection personnel, and then improve the quality inspection efficiency.
Intelligent annotation helps data pre-training and model development empower each other, and leading companies at home and abroad have cut in, and the effect of cost reduction and efficiency increase is remarkable. On the one hand, intelligent assisted annotation helps data pre-training service providers improve annotation efficiency and reduce annotation costs, and on the other hand, empowers model developers to improve model performance, improve model R&D efficiency, and achieve closed-loop. At present, the trend of automatic labeling in the industry is significant, and leading companies at home and abroad have cut in, and the effect of reducing costs and increasing efficiency is remarkable. For example, Haitian AAC has launched an integrated intelligent data processing platform with access to open source large models, and Appen's self-developed intelligent assisted annotation platform can improve efficiency by 91.5% through data pre-labeling.
3.2 Cut into autonomous driving and enjoy more industry increments
The company has always closely followed the transformation of AI large model application scenarios, relying on the advantages of technology, products, and R&D, and taking the lead in laying out the autonomous driving track, which will better grasp the initiative of the intelligent driving data market and divide more industry incremental space.
(1) Platform-based overall planning capabilities: The company has successfully accumulated experience in project overall management and data pre-training and processing for autonomous driving scenarios. The current perfect personnel management structure helps the efficient allocation of personnel for large-scale projects, and sets appropriate confidence intervals, algorithm engine voting mechanisms, and confidence intervals for data diversity and complexity to improve the quality and efficiency of data pre-training.
(2) Intelligent level: Annotating the point cloud data obtained by terminal sensors such as lidar and depth cameras can help autonomous driving extravehicular algorithms and service robots to achieve prediction, such as accurate environment perception, efficient path planning, reliable obstacle detection, and ultimately help the algorithm make behavioral decisions. Up to now, the company's intelligent auxiliary annotation tools have realized point cloud continuous frame and point cloud fusion annotation, building a more comprehensive 3D environment model for autonomous driving scenarios.
(3) Technological advancement
The changes in terminal scenarios drive data pre-training enterprises to form their own insights into the trend of terminal software layer algorithms and hardware layer data collection, forward-looking R&D layout, continuous iteration of annotation technology, and continuous maintenance of technological advancement. Up to now, the company has the ability to solve autonomous driving solutions, and has the ability to annotate multiple types of data such as in-cabin voice, extra-cabin images, and videos.
As shown in Table 5, the Enable AI intelligent annotation platform supports continuous frame annotation of sensor 3D point cloud data and point cloud fusion of different data.
Intelligent driving scenarios are expected to take the lead in releasing the demand for pre-training data, and the demand release progress will periodically converge with the time of algorithm iteration and mass production of landing models.
Considering that data processing is located in the upstream of the algorithm development industry chain, and data demand is placed in the terminal scenario, the intelligent driving scenario is expected to take the lead in releasing pre-training data service industry. According to Deloitte's estimates, the demand for AI pre-training data services brought by intelligent driving is expected to reach 8.3 billion yuan in 2027E, with a five-year compound growth rate of 37% from 2022E to 2027E, accounting for 52% of the market share.
In 2022, autonomous driving will be developed and promoted at the L2+ level, and the commercial application of L3 autonomous driving is expected to be realized in 2025, and the gradual implementation of L4 autonomous driving is expected to be realized in 2030.
Therefore, at present, it will continue to benefit from the increase in data demand brought about by the iteration of L2+ to L3 technology, and the basic data demand will begin to converge relatively after 2025. In 2027, the upgrade iteration from L3+ to L4 will begin, the algorithm iteration will increase the complexity of application scenarios, and the demand for AI pre-trained data processing will increase exponentially, and the demand may start a new round of gradual release in 2027.
4 Earnings Forecasts and Valuations
4.1 Profit Forecast
Assumptions of main business income:
(1) Machine tool business: In 2022, the company's business revenue from this part of the business will be 609 million yuan, an increase of 28.61%. The company is located in the first echelon of machine tool equipment manufacturing, and the technical barriers are stable. According to the company's sales data in the past two years, assuming that the sales volume from 2023 to 2025 will be 500, 512, and 515 units, and the corresponding business revenue will be 7.61, 9.29, and 1.133 billion yuan respectively, and the corresponding growth rates will be 24.89%, 22.21%, and 21.87% respectively
(2) Data pre-training business: This part of the business is mainly for network content review and large model pre-training data annotation, which accounts for a small but considerable amount of revenue at present, and is expected to become the biggest driving force for future business growth.
We assume that the growth rate of this part of the business will be 65.13%, 99.15%, and 60.21% in 2023-2025, corresponding to the revenue of 0.50, 1.00, and 160 million yuan in 2023-2025, and the corresponding growth rates will be 65.13%, 99.15%, and 60.21% respectively.
(3) Other business: In 2022, the company's business revenue from this part of the company will show a slight increase trend compared with the previous year, and the change will be relatively stable. Assuming that the company will continue to maintain a steady growth rate in the future, we predict that the company's revenue from this part of the business from 2023 to 2025 will be 0.75, 0.77 and 79 million yuan respectively.
In summary, it is estimated that the revenue from 2023 to 2025 will be 8.86, 11.06, and 1.372 billion yuan, a year-on-year increase of 26.11%, 24.83%, and 24.01%.
Gross margin assumptions:
(1) Machine tool manufacturing business: This part of the company's business is mature and has a scale effect, and the current company is switching horizontally to the mid-to-high-end market, and the gross profit margin is expected to continue to improve, which is expected to be 22.99%, 24.23%, and 25.12% respectively from 2023 to 2025.
(2) Data pre-training business: This part of the company's business is expected to be catalyzed by the application of intelligent driving scenarios, and the business will switch from Internet content annotation to high value-added pre-training data annotation. The estimated gross profit margin is relatively stable, and it is expected to be 21.12%, 22.85%, and 25.16% respectively from 2023 to 2025.
(3) Other business: Assuming that the gross profit margin level changes steadily, it is expected to be 22.86%, 21.73% and 20.92% from 2023 to 2025.
To sum up, considering the proportion of business, the company's comprehensive gross profit margin from 2023 to 2025 is estimated to be 22.87%, 23.93% and 24.88%.
4.2 Valuation
The company's current main business is machine tool business, so it chooses high-end equipment manufacturing leaders Qinchuan Machine Tool, Haitian Precision and Yawei shares. Haitian Precision is the latest forecast data, Qinchuan Machine Tool and Yawei Co., Ltd. are wind consensus forecast data, and the average PE values of comparable companies from 2023 to 2025 are 36.40X, 27.32X, and 21.77X.
The company's technical barriers in the traditional core machine tool manufacturing business are stable and the advantages are obvious. The new entry into the basic data service track benefits from the long slope and thick snow of the large model, superimposed on the company's forward-looking card position intelligent data annotation and automotive autonomous driving, which is expected to divide more industry share.
To sum up, we expect the company to achieve operating income of 8.86/11.06/1.372 billion yuan from 2023 to 2025, and net profit attributable to the parent company of 1.82/2.66/354 million yuan, corresponding to PE of 41.69/28.53/21.45 times respectively.
5 Risk Warning
1. The combination of large model industries is less than expected
The implementation progress of the integration of large model industry is under the pressure of the level of model iteration, the confirmation of industrial data, and the increasing difficulty of industrial data collection and labeling, and the combination of model industry has a long way to go.
2. The competition in the intelligent data labeling market has intensified
With the increase in market volume, many of them have developed their own data annotation models to seize the first-mover dividend, and the competition has gradually intensified.
3. The implementation of the intelligent labeling platform is not as expected
The company's self-developed intelligent annotation model is still in the process of research and development, and there is a risk that it is difficult to overcome technical difficulties and meet expectations.
——————————————————
The report belongs to the original author, only for Xi! If there is any infringement, please delete it privately, thank you!
The report comes from [Foresight Think Tank]