At present, the AI industry has ushered in a new round of entrepreneurship marked by "large models", and large model startups have broken ground and are thriving. However, we should also note that traditional AI companies (such as CV Four Tigers) have also had a similar golden period of development, and now they have more or less encountered development bottlenecks. This article aims to gain an in-depth understanding of the similarities and differences between them in terms of technological innovation, business model, capital operation, market competition strategy, and future development potential through comparative analysis of the two, so as to help gain insight into the development trend of the AI industry and provide valuable insights and strategic references for investors, analysts, and entrepreneurs in the AI field.
CV "Four Little Dragons"
The Four Tigers of CV refer to four leading AI companies in the field of computer vision (CV) in China, namely SenseTime, Megvii, Cloudwalk and YITU. These companies have made remarkable achievements in the research and commercial application of computer vision technology, accounting for most of the share of China's computer vision market, and also have a high reputation and influence in the world.
SenseTime focuses on the research of original computer vision and deep learning technologies, and has become the largest AI algorithm provider in Asia. With face recognition technology as the core, Megvii Technology provides solutions for personal Internet of Things, urban Internet of Things and supply chain Internet of Things. Incubated by the Chinese Academy of Sciences, Cloudwalk is a national team among artificial intelligence companies, especially in the financial field. YITU Technology has gained market recognition in the field of security, and has since expanded into financial, medical, pharmaceutical, and chip markets.
Horizontal comparison of the "Four Little Dragons".
However, the current "four little tigers" are trying to find ways to go through the cycle.
On March 26, 2024, SenseTime released its 2023 results announcement. According to the financial report, the group will achieve a total revenue of 3.405 billion yuan in 2023, compared with 3.808 billion yuan in the same period of the previous year, a year-on-year decrease of 10.6%, mainly due to the contraction of the smart city business. Its core business, generative AI, increased by 200% year-on-year to RMB1.2 billion, accounting for 35% of the group's total revenue. At the same time, SenseTime Group will lose 6.494 billion yuan in 2023, compared with a loss of 6.092 billion yuan in the same period last year, and the loss will expand by 6.6% year-on-year. It can be seen that SenseTime is still the head of the "Four Little Dragons", and while the new business is growing rapidly, the traditional business is also shrinking.
In 2023, Cloudwalk's total revenue reached 628 million RMB, an increase of 19.43% compared to the previous year, which was mainly due to the significant increase in sales of its human-machine collaborative operating system (CWOS)-related software. However, in the same year, Yuncong Technology suffered a loss of 640 million yuan, and the company's total assets also decreased by 19.47% from the previous year. Cloudwalk Technology's revenue is still dependent on traditional business, and it has not seen new growth points for the time being.
According to Megvii's IPO filings, the company's revenue from 2018 to 2020 and the first half of 2021 was 854 million, 1.260 billion, 1.391 billion and 670 million yuan, respectively. During the same period, the net profit attributable to shareholders of the parent company showed continuous losses, which were -2.8 billion, -6.643 billion, -3.326 billion and -1.865 billion yuan, respectively. Megvii's business is mainly divided into three parts: consumer Internet of Things, urban Internet of Things, and supply chain Internet of Things. Based on the revenue in 2020, the revenue of these three business segments accounted for 18.47%, 65.82% and 15.71% respectively. Although the urban Internet of Things is the core business of Megvii Technology, similar to Yuncong Technology, Megvii Technology is also experiencing a certain degree of market reduction.
YITU terminated its IPO a few years ago, and its financial status is not yet known.
AI is a field that requires long-term investment, and ups and downs are inevitable. What are the lessons of the success and failure of the "Four Little Tigers" in the development process for the current AI model startups?
The history of the development of the "Four Little Tigers".
Since 2020, the trend of domestic substitution has accelerated, and the market share of local brands has increased, and the "Four Little Tigers" have seized this opportunity. SenseTime has strengthened its leading position in the supply of AI algorithms; Megvii Technology has advantages in the field of face recognition technology; With its "national team" attributes, Cloudwalk has made remarkable achievements in the financial field; YITU focuses on security and expands into new markets such as healthcare. They have received financial support through multiple rounds of financing, and have participated in the formulation of national standards while continuously improving the competitiveness of their products, jointly promoting the market expansion and technological progress of domestic AI brands.
At present, some of the four companies have been listed, and some are still waiting for hibernation: SenseTime Technology was successfully listed on the Hong Kong Stock Exchange in December 2021, which was one of the world's largest AI company IPOs that year; In August 2020, Cloudwalk Technology went through the counseling and filing registration with the Guangdong Securities Regulatory Bureau, and after a series of processes, it was listed on the Science and Technology Innovation Board in May 2022. The IPO of Megvii Technology's Science and Technology Innovation Board was accepted on March 12, 2021, successfully passed the meeting on September 9, 2021, and submitted for registration on September 30, 2021. YITU had planned to list on the STAR Market and submitted an IPO application in November 2020, but later in June 2021, YITU voluntarily withdrew its listing application, and the listing process was suspended.
The IPB journey shows that the AI market is still in its early stages of development and will require more time and capital to advance commercialization. Taking a deeper look, the "Four Little Dragons" have the problem of homogeneity in the layout of specific business scenarios. Their core technologies are backed by the development of computer vision technology, and their business models are mainly B2B and B2G.
High degree of overlap in the early business scenarios of the "Four Little Dragons" (Qianshan Capital 2019 Research Report)
Therefore, although they enjoyed the growth dividend of the industry in the early days, with the entry of Internet giants such as BAT, market competition intensified, and the capital boom in the AI industry began to retreat, and the four companies had to start looking for differentiated business models and development space to expand to more segments.
Megvii Technology transformed into an IoT operating system provider and entered the field of AI Internet of Things. YITU chose to design its own AI chip to solve the problem of hardware and software integration, and released the first deep learning cloud custom chip "Quest". Cloudwalk focuses on four major business segments: finance, security, transportation, and commerce, and has made breakthroughs in the field of smart business. SenseTime continued to make extensive layouts in many fields, including education, and established a sub-brand "SenseTime Education......
It can be said that the success of CV Four Dragons lies in seizing market opportunities through continuous technology research and development and innovation, clear market positioning, active capital operation, and the use of policy support and cooperation. However, the continuous high R&D investment, the difficulty of commercializing AI technology, the limited market demand, especially the fierce competition in the ToG market and the long business cycle, have led to serious losses. In addition, the rapid iteration of technology and the fragmentation of scenarios also make it difficult for standardized products to form scale effects.
With that in mind, let's now turn our attention to AI model startups and see how they're doing.
Overview of AI large-scale model startups
Large model is a new thing, at present, there are not too many start-up companies in this area in China, and the exposure rate is relatively high: the dark side of the moon, Zhipu AI, Baichuan Intelligence, Shengshu Technology and Power Law Intelligence, etc.
The Dark Side of the Moon is a startup company in the field of general artificial intelligence, focusing on the development of C-end products, and has launched Kimi Chat, an intelligent assistant that supports 2 million word contexts, and the market is very hot.
Zhipu AI is transformed from the technical achievements of the Department of Computer Science of Tsinghua University, based on the GLM series of large models to create ChatGLM conversational robots, contribute to the open source and other pre-trained models such as chatGLM-6B, and cooperate with cloud vendors to promote its AI services.
Baichuan Intelligence is an artificial intelligence company founded by Wang Xiaochuan of Sogou, focusing on the development of large model technology by virtue of its accumulation in search technology and algorithms. At present, the Baichuan series of pedestal models are open-sourced, while providing advanced features, custom development, and enterprise services.
Biodigital Technology is a company focusing on multi-modal large models, benchmarking against Midjourney, hoping to make a difference in the fields of image generation, 3D content generation and video generation.
Power Law Intelligence chose to cooperate with Zhipu AI to launch a legal vertical model based on the 100 billion base model ChatGLM-130B
PowerLawGLM, the model has been implemented in the smart contract business and upgraded to version 3.0.
It is not difficult to see that AI large model startups have their own advantages in choosing directions, which are more diverse than the "Four Little Dragons", and there are their own reasons, let's take a look.
Comparative analysis of AI large-scale model startups and the "Four Little Tigers".
Technical level: Computer vision technology originated in the late 50s of the 20th century, and it mainly focuses on the analysis and understanding of image and video data, including image recognition, object detection, image segmentation and other fields. With the complexity of application scenarios, a single CV technology faces the challenges of large data volume and variable scenarios. AI large model technology has emerged with the development of deep learning in recent years, especially large models in the field of natural language processing, such as BERT, GPT, etc., which have excellent performance in processing large-scale and high-dimensional data. Large AI models are not limited to text and images, but can also process data in various modalities such as video and voice, which is stronger than CV technology in terms of versatility, but also requires more data and computing resources.
Products and markets: Compared with the homogenization of CV Four Dragons in the vertical field, AI large model companies can not only do vertical fields in the future, but also may take the road of "platform services". With the help of pre-training and fine-tuning technology, AI large model companies can use specialized data to help customers train private models in vertical fields, while powerful general-purpose models have the potential to become platform-level products, such as using AI Agents technology to create rich applications based on models. In terms of the market, AI large model companies can not only enter the traditional B2B and B2G markets, but also reach the broader C-end market with their diversified product capabilities. Each startup has its own resource endowment, and it is up to the entrepreneur to judge whether to choose the C-end or the B-side. At present, the domestic manufacturers are basically biased towards the B-end, while the entrepreneurial star companies are basically taking the C-end route.
Business model: CV Tigers mainly make money by providing customized AI solutions to enterprises and governments, which need to have an in-depth understanding of customer needs, provide personalized services, and implement specific projects in actual operations, with a long cycle and high cost.
By providing API interfaces or Platform-as-a-Service (PaaS) models, AI large-scale model companies allow customers to call AI models according to their own needs, which can quickly expand the scope of services, reduce marginal costs, and improve customer stickiness. Once a customer recognizes the brand, there are various ways to monetize, such as subscriptions, cloud service charges, licensing, etc. Theoretically, standardized products are of course more advantageous due to their low cost and quick results, but they are also more competitive. The cost of customized development of large models actually depends on the customer's own degree of digitalization, and it is easier to implement in industries with rich basic data such as transportation and finance, but on the contrary, there are risks such as expanding the scope of the project.
Capital flow: For start-ups, whether they can raise enough funds in the rapid development stage determines the company's future market share. The CV Four Tigers all raised funds to support their development in the early days.
Financing of the "Four Little Tigers" as of the end of 2020
The market share of the "Four Little Tigers" in 2022
Compared with traditional AI companies, due to the high cost of training and applying large models, AI large model startups need more funds in the early stage, and the current investment willingness in the domestic primary market is not strong, and it is expected that only a very small number of large model startups can raise enough funds. For example, the dark side of the moon will have a new round of financing in February 2024, with an amount of more than $1 billion; Zhipu AI has also received a total of more than 2.5 billion yuan in financing in 2023, which can be described as complete and ready to go.
How will AI large-scale model startups develop?
In 2022, the size of Chinese's artificial intelligence software market will be 30.73 billion yuan, a decrease of 6.9% compared with 2021, which is the first negative growth in nearly a decade. Prior to the emergence of the large model boom, the AI market lacked a strong growth driver, resulting in a decline in the overall market. Incumbents can still achieve small growth in a declining market, while growing companies need to continue to provide leading technology or find a track with stable growth potential.
For AI large-scale model startups, taking advantage of this technological revolution, the development of technology, products and business is worth looking forward to.
In terms of technology, AI large model companies tend to focus on optimizing the existing model architecture to improve the performance and efficiency of the model, including adjusting the neural network structure, optimizing the parameter configuration, and improving the training strategy. Later, with the accumulation of technology, leading companies are bound to start to develop new algorithms independently, or propose innovative data processing technologies, such as the dark side of the moon, which focuses on the development of long text processing technology.
In terms of product roadmaps, AI large model companies will launch products with a single function in the early stage, such as image generation or natural language processing services, to verify the feasibility of the technology and market demand. As the technology matures and the market is recognized, the company will expand its services to the platform, where the "platform" is not necessarily a comprehensive solution, but may be a service similar to the application market. Platform-based services can meet a wider range of customer needs, enhance customer dependence, and create a good application ecosystem, such as ByteDance's "buttons" or products like Kimi+.
Business models are not the focus of large model companies, but they can still bring stable revenue through model innovation (API/PaaS/MaaS models), such as the model-as-a-service (MaaS) model, which allows customers to use large AI models in the cloud without having to build and maintain their own infrastructure. Once the product is accepted by the market and the number of users surges, it is necessary to find ways to reduce costs at this time, such as OpenAI limiting the number of API calls in the later stage. After all, AI large model training and inference require a large amount of computing resources, and controlling the cost of computing power is the key to profitability.
In terms of market competition, AI large model startups mostly adopt differentiation strategies, because the market for large models is very broad and there is enough space to accommodate different products. As for the strategies used by the "Four Little Tigers", such as the vertical integration of YITU Technology (vertical integration of algorithms, chips and applications through self-developed AI chips, which helps improve the competitiveness of the overall solution), SenseTime's platform strategy, and the diversification strategy of Yuncong and Megvii, they may not be applicable, because startups are competing with large manufacturers from the beginning.
Special challenges
The bottlenecks of technological innovation, intensified market competition and capital dependence encountered by the CV Four Dragons will also be encountered by AI large-scale model startups. In addition to this, they face other challenges.
Large AI models require a large amount of data for training, especially when private data is involved, how to ensure the security of these data and prevent data leakage and abuse is an important issue. In addition to the security, compliance, cost-effectiveness, and flexibility of cloud services, it is also important for enterprises to consider the risk of vendor lock-in and whether the cloud services can support an effective data governance strategy.
In addition, the cost of obtaining high-quality data is rising, for example, Google has to pay more than $60 million a year for Reddit's data, and startups cannot afford such a cost, so they have to think of other ways out. Once the data is available, the AI model must also be securely processed to ensure that the model does not generate content that contains private and sensitive information.
Different from ordinary companies, large-scale model startups also have an obvious feature, that is, "small number of people, strong culture", and a team of less than 10 people and engineer culture are "standard". How to unite people and create a working atmosphere conducive to innovation is a question that all large-scale startups must consider.
The technological breakthrough and implementation speed of large models and generative AI exceed expectations, and the technical strength and computing power determine everything, and it is expected that manufacturers that do not have real strength will be quickly eliminated. In the future, the market is broad enough, there is a lot of room for differentiated competition, and at the same time, due to the low marginal cost of software, there may also be a winner-take-all situation in the general field. Overall, the opportunities outweigh the risks, and the future is worth looking forward to!