Introduction: Among all the AI technology service providers in China, SenseTime can be regarded as the most special one.

On August 27, SenseTime (00020.HK), an AI technology company listed on the Hong Kong stock market, released its 2024 interim results announcement.

According to the financial report, in the first half of 2024, the company's overall revenue will be 1.74 billion yuan, a year-on-year increase of 21.4%, and among the three major businesses, the revenue from generative AI will be 1.05 billion yuan, a significant increase of 255.7% year-on-year, and the revenue contribution will reach 60.4%, becoming SenseTime's largest source of revenue. Driven by the strong generative AI business, the Group's overall gross profit reached RMB770 million in the first half of the year, up 18.2% year-on-year.

The strong rise of generative AI has become a new revenue pillar and source of growth for SenseTime.

Globally, the AI industry is still in the early stage of development where input far outweighs output, and how to use AI technology to empower traditional business models and achieve sustainable business revenue is a common challenge for the entire industry and even the AI technology revolution.

With everyone struggling to find answers, why can SenseTime achieve significant revenue growth with generative AI? It's a question worth pondering.

01 What does SenseTime rely on?

Since the beginning of this year, with the soaring stock price and performance of Nvidia (NVDA.US) on the other side of the ocean, many people are surprised to find that the demand for computing power and services from the field of generative AI is growing rapidly at an eye-popping rate.

In the Chinese market, such demand is also exploding at an alarming rate. The IDC report predicts that in the next five years, China's AIDC service market will continue to grow at a high growth rate, with an average annual compound growth rate of 57.3%, and the overall market size will be nearly 200 billion yuan in 2028, of which the computing power demand on the training side will increase by more than 10 times, and the computing power demand on the inference side will increase by more than 200 times.

So, what exactly does the "generative AI business" mentioned in SenseTime's financial report include?

To answer this question, we need to go back to business fundamentals.

First of all, there is no doubt that SenseTime is one of the important players in the AIGC field, and its RiRixin large model is one of the most advanced large model products in China.

As a large model service provider, SenseTime provides support for many customers to call large model capabilities, so the revenue generated by this part of the business has naturally become one of the components of SenseTime's generative AI revenue.

However, for SenseTime's business territory, providing large model capability invocation services to the outside world is only a part of the revenue composition of generative AI. Another core component is the SenseCore SenseTime AI device.

What is SenseTime's AI Megadevice?

In short, SenseTime's AI device is a complete set of infrastructure used by SenseTime itself and provided to third-party customers to train AI models, including hardware and infrastructure such as GPUs and AIDC, as well as many software and algorithms developed by SenseTime.

In fact, the entire SenseTime AI device SenseCore consists of three layers: the computing power layer, which includes AI chips and processing cards, as well as AIDC-related facilities for these hardware; The platform layer includes many supporting software and background programs required for the training and data management of large models; The algorithm layer is the algorithm toolbox and development framework provided by SenseTime to customers.

Whether it is the hardware of AI computing power, or the use of various tools and frameworks at the platform and algorithm level, these resources are integrated together to become a unified large model business platform that can be used by customers at will, so that customers can train their own large models in the shortest time and at the lowest cost, which is where the existence value of SenseTime's AI large device lies, and it is also the basis for SenseTime to gain differentiated competitiveness in the generative AI market.

02 AI Combo: Large Model + Large Device

In the view of SenseTime's management team, the company's biggest competitive advantage comes from the organic synergy of large-scale model enlargement devices.

Dr. Li Xu, Executive Chairman and Chief Executive Officer of SenseTime, said, "Generative AI is in a prime period of development, and we are well-positioned for the next wave of growth. SenseTime's core advantage lies in the deep collaboration of 'large device + large model', the ability to build a first-class large model, and the unique advantages of native multimodality, natural interaction of video streams, and low-cost model inference architecture, so as to stand out in the fierce market competition. ”

As part of the "large model + large device" collaborative strategy, SenseTime has been developing SenseCore for a long time - in fact, as early as 2018, SenseTime started the construction of large devices.

In the first half of 2024, the total computing power of SenseTime will reach 20,000 PetaFLOPS and the scale of more than 54,000 GPUs, ranking firmly in the industry and capable of supporting the stable and efficient training and inference of MoE multimodal models with up to 2 trillion parameters.

In terms of platform upgrades and optimizations, in inference scenarios, SenseTime's innovative technical architecture can increase the number of requests per second (QPS) by 4 times with the same computing power and power cost, and support elastic on-demand scaling of inference services, further optimizing the overall cost of large-scale AI inference. In addition, based on the original technological innovations such as the integration of training and pushing, computing and power synergy, SenseTime continues to improve the GPU resource utilization rate of the Vanka cluster to more than 80%, greatly improving the overall operating return rate of AIDC.

As the other pole of the "large model + large device" collaborative strategy, SenseTime is also speeding up the iteration and development of large models.

In the first half of this year, SenseTime's "RiRixin" large model was iterated to version 5.5, with significantly improved comprehensive capabilities, achieving the performance of GPT-4o in terms of multi-modal and real-time interactive experience, and providing the best large model foundation for the development of AI application layer.

According to the financial report, in the first half of the year, the overall call volume of SenseTime's Ririxin model increased by 400%, and the number of users and the average call volume increased significantly; Customers are also in the Internet, smart hardware, electric vehicles, robotics, healthcare, finance and other industries.

SenseTime's large model and large device are forming an increasingly close synergy at the business level: as the underlying computing power foundation of the large model, the large device provides the most basic guarantee for the rapid upgrade and iteration of the new large model every day; The large model iteratively developed from the large device provides customers in different industries and different needs with flexible capability invocation and demand satisfaction, so that industry customers do not need to start everything from scratch, but quickly realize their business needs with the help of the mature capabilities of the new large model, which greatly reduces the time and resource cost of business process reengineering through AI.

In the first half of the year, more than 3,000 customers used SenseTime's AIDC and RiRixin large model business.

03 Technology empowers the industry AI to create value

In the field of generative AI, SenseTime's customers include Internet companies such as JD.com, Xiaomi, and Kingsoft, as well as industry leaders such as the three major telecom operators, Geely, and OPPO, and even large-scale startups.

For those large companies and institutions with strong strength, why are they willing to choose SenseTime as their primary partner to enter the AI field? What can SenseTime's large models and devices bring to these customers?

The answer can be glimpsed through two cases from different industries.

Case 1: PathOrchestra pathology model

In the medical industry, pathological diagnosis is known as the "gold standard" of disease diagnosis, but for a long time, this field has been faced with the problems of long training cycle of pathologists and uneven distribution of high-quality pathological diagnosis resources. Even for tertiary hospitals in China, excellent pathologists are very scarce talents.

However, due to the high resolution of digital pathology slides and the large number of diseases involved, it is almost an "impossible task" to train each disease under the traditional AI model training paradigm of "big data + fine annotation".

In fact, there have been many AI companies in China that want to challenge this problem, but after many surveys and analyses, these challengers have basically chosen to give up.

In order to solve this global problem, SenseTime has joined forces with a number of top medical institutions and university research institutions in China, including the team of Vice Chairman Wang Zhe of the Pathology Branch of the Chinese Medical Association and the team of Professor He Yonghong of Tsinghua University, to jointly start the research and development of AI-based large-scale model pathological diagnosis technology.

The traditional development process of large medical models requires a lot of manpower to classify, label, and organize medical data, and the computing power required in the model training process is also a big problem. However, for SenseTime, it can replace manual data sorting and automatic labeling through the ability of the new large model every day; Then, through the flexible configuration of large AI devices, the computing resources required for model training can be solved, which is lower cost and faster than traditional development methods.

During the 2024 World Artificial Intelligence Conference (WAIC), SenseTime joined hands with Ruijin Hospital, West China Hospital, Xinhua Hospital, Xijing Hospital, the First Affiliated Hospital of University of Science and Technology of China, and Beijing Tsinghua Changgung Hospital to jointly launch the first pathology model PathOrchestra in China, which has become the pathology model with the most extensive clinical tasks in the world at this stage.

The pathology model PathOrchestra released this time is trained by SenseTime and a number of professional medical institutions using the largest dataset of nearly 300,000 full-slice digital pathological images (nearly 300TB data) in China, and integrates multimodal training data such as text and video, and through self-supervised learning of massive data, the model has learned to analyze pathological images of various organs, and has now covered more than 20 organs such as lungs, breasts, liver, and esophagus.

With the joint support of SenseTime's advanced AI large model + large device, PathOrchestra has been able to empower more than 100 clinical tasks, including pan-cancer classification, lesion identification and detection, multi-cancer subtype classification, biomarker evaluation, etc., and has an accuracy rate of more than 95% in nearly 50 tasks such as multi-organ pan-cancer classification, lymphoma subtype diagnosis, and bladder cancer screening.

Many patients will benefit from this groundbreaking work. SenseTime's overall solution of digital and intelligent pathology driven by large models has also brought a steady stream of orders and returns to the company in the medical field.

Case 2: WPS Copilot Pro

As the most senior office software company in China, Kingsoft's WPS is the most commonly used software in many enterprise office scenarios. In April this year, Kingsoft Office officially launched WPS 365, and joined hands with SenseTime to push the office application based on the RiRixin model to a new height.

For modern office software, the analysis, mining, sorting, and visualization output of data is a highly frequent work. After years of development, modern office software usually has built-in powerful data analysis and report generation modules. However, the use and connection of these functions are often extremely complex and require the participation of professional developers, and it is difficult for general business managers to properly use complex APIs if they lack professional code development capabilities.

In view of this, in order to further improve the production efficiency of WPS office software and reduce the workload of customers, Kingsoft Office has joined hands with SenseTime to launch WPS 365. Different from the personal version, WPS 365 focuses on building enterprise brains for customers, and it is divided into three components: AI Hub (intelligent base), AI Docs (intelligent document library), and Copilot Pro (enterprise intelligent assistant).

Among them, AI Hub integrates the core capabilities of a variety of large models and provides a unified call interface. AI Docs upgrades the traditional cloud document library to an intelligent document library with one click, so that the source of intelligent creation has a basis, and the complete document permission system ensures that the information is not overstepped.

As for the most important Copilot Pro (enterprise intelligent assistant), its core function is to help operators use natural language-driven BI products for data analysis and processing. Copilot Pro converts user needs into executable Python code, and calls WPS 365 API and enterprise private API to automatically complete the corresponding data analysis and processing work, and then feedback to users in a visual way.

The WPS dashboard data analysis application based on the code model capabilities extended by SenseTime can complete data understanding and analysis tasks through dialogue

For example, if a senior executive wants to understand and analyze the attendance of employees in different factories of the company, he only needs to create an intelligent assistant and connect to the corresponding human resources API, and then the user only needs to put forward the analysis requirements in the form of natural language, and the intelligent assistant can automatically obtain the attendance data of different factories, conduct comprehensive analysis, and present the final analysis results. The whole process does not require the involvement of technical personnel, let alone tedious code development and maintenance.

In fact, in the past year, the exploration of domestic large models has been in full swing, and code generation capabilities are often regarded as a key dimension to measure the wisdom of large models. Code generation requires a high degree of abstract thinking and logical reasoning, and requires accurate adherence to the specifications of the code language. At present, among the large models in China, the code generation capability of SenseTime's new model surpassed GPT-4 at the earliest, and it has significant advantages in data analysis, mathematical computing, logical reasoning and other subdivisions.

04 A sea of stars belonging to AIGC

What's next for generative AI? Is it infrastructure, applications, or development platforms? No one knows the exact answer yet.

On the other side of the ocean, chip manufacturers represented by Nvidia have already enjoyed growth dividends in this wave; A number of domestic manufacturers are also striving to seize this development opportunity based on their different resource endowments and advantageous businesses.

Among all the generative AI technology service providers in China, SenseTime can be regarded as the most representative one.

Other companies, some focus on AIDC-related businesses and engage in AI computing power rental business, but in essence, they are actually repackaging traditional IDCs; Some are pure large-scale model developers, devoting all their resources and energy to the development and iteration of larger-scale and higher-cost large models, but they are still in a very early stage of commercialization.

And SenseTime, which "walks on two legs", has shown a unique competitive advantage in this market.

On the one hand, compared with AIDC service providers, SenseTime can provide much richer product and service support, so that customers do not need to build large-scale model-related infrastructure from scratch, but can focus on core business logic with the help of SenseTime's mature large-scale model production environment, without worrying about related support and algorithm frameworks.

On the other hand, compared with large-scale model developers, SenseTime is able to provide enterprise customers with more choices that are closer to their business needs, whether it is completely dependent on the capabilities of SenseTime's new large models, or those traditional business giants with more ambitious goals and want to train their own private models, SenseTime can provide full-stack solutions to meet the multi-level and multi-stage needs of customers.

If the mobile Internet has penetrated and affected thousands of industries in the past two decades, then the future will be the era when AI technology will once again reshape and transform the business world. The opportunities contained in it are in the tens of thousands.

No matter where the next development direction of AI technology goes, for a technology-driven enterprise focusing on long-termism, it is the right way to stick to its position in the industrial chain, create value for customers, bring progress to the industry, and consolidate and expand the company's business territory on the premise of ensuring long-term survival and development.

This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com