SenseTime has entered the 2.0 era: the revenue of the generative AI business driven by the "RiRixin" large model took the lead in exceeding 1 billion

On March 26, SenseTime released its 2023 annual financial report, showing its strong growth momentum in the field of generative AI.

In 2023, SenseTime will achieve revenue of 3.4 billion yuan, gross profit of 1.5 billion yuan, and gross profit margin of 44%. Among them, the generative AI business driven by the "RiRixin" large model + large device has grown explosively by 200%, and the revenue has reached 1.2 billion yuan, contributing 35% of the group's revenue.

Thanks to the forward-looking layout of the "large device + large model" strategy, SenseTime has not only made remarkable achievements in AI infrastructure construction, but also achieved breakthroughs in generative AI business. The 2023 report card shows its strong strength and broad prospects in the field of generative AI. With the continuous advancement of AI technology and the increasing expansion of application scenarios, SenseTime will continue to focus on the core business of generative AI and promote the company's continued rapid growth. The success of SenseTime has not only set a benchmark for the industry, but also laid a solid foundation for the advent of the AI 2.0 era.

Generative AI business has become a new engine for growth

According to the annual report, in 2023, SenseTime's generative AI business revenue will reach 1.2 billion yuan, which is a new business with a revenue volume of more than 1 billion yuan at the fastest speed since SenseTime was established ten years ago.

In addition, the total computing power of large devices has achieved a breakthrough growth of 12,000 petaFLOPS and 45,000 GPUs in operation, thus supporting the high-speed iteration of the "RiRixin" large model system to version 4.0 and reaching the ability level to match GPT-4 in multiple scenarios.

Dr. Li Xu, Chairman and CEO of SenseTime, said, "Generative AI has become not only a transformative innovation in the field of technology, but also a core business for SenseTime. The growth of SenseTime's generative AI business is due to the extensive demand for large model training and inference in various industries, which heralds the official start of a new cycle of hard technology investment in China. By deeply integrating generative AI capabilities at all levels of the business, SenseTime is winning new customers and driving overall efficiency and productivity. ”

Looking back on 2023, SenseTime's "RiRixin" large-scale model series has continuously undergone important technological upgrades, and each upgrade has been firmly at the forefront of the industry, demonstrating SenseTime's profound technical heritage. By February 2024, SenseTime will launch "RiRixin 4.0", which will significantly improve the performance of large models by improving the quality of training data without increasing the number of parameters. In multiple scenarios such as code writing, data analysis, and medical Q&A, a level of competence matched that of GPT-4 was reached.

With the support of a solid technical foundation, SenseTime's "RiRixin" model has established application value in many industries, and is widely used in many fields such as finance, medical care, and intelligent terminals.

With the wave of large models continuing, generative AI is gradually becoming the new favorite of all walks of life, and SenseTime, with its deep accumulation in the field of large language models (LLMs), has successfully transformed this cutting-edge technology into practical business applications, so as to take the lead in the fierce market competition.

In 2023, SenseTime has identified three new business segments: generative AI, traditional AI and smart cars, with a focus on generative AI business while maintaining the advantages of traditional AI business and smart cars. Behind this adjustment, SenseTime not only marks its transformative innovation in the field of technology, but also shows the company's firm pace on the road to AI commercialization.

"Large device + large model"

Build the core competitiveness of AI

In the competition for artificial intelligence, technology is the greatest confidence. Behind the rapid revenue brought by generative artificial intelligence to SenseTime, it is actually the harvest brought by SenseTime's advanced layout at the technical level several times.

Before the rise of a new round of artificial intelligence, SenseTime built a collaborative strategy of "large device + large model" early with its forward-looking layout, and took the lead in completing "model as a service".

SenseTime is an early adopter of AI infrastructure, and has been heavily investing in AI infrastructure since 2018, especially in the SenseCore large device, which has laid a solid foundation for the company's long-term development in the AI field. In addition, as the leading AI computing center in Asia, Lingang AIDC, which was vigorously built by SenseTime under the pressure of losses, now not only provides strong computing support, but also provides a strong guarantee for the development and iteration of large models.

SenseTime is also one of the earliest manufacturers in China to develop large models and release large model series products. Its pre-trained base model, InternLM-104B (104 billion parameters), outperformed OpenAI's GPT-3.5-turbo on 35 evaluation sets.

Based on the collaboration of "large device + large model", in the past year's "100 model war", SenseTime has become one of the few companies that can maintain the high-speed iteration of large models.

Based on the industry's leading AI infrastructure and model services, SenseTime's customer model training, fine-tuning, and inference businesses are growing explosively. More than 70% of the business's customers were new to SenseTime in the past 12 months, while the remaining 30% of existing customers also recorded an increase of about 50% in order value. Typical customers come from all walks of life, such as the three major operators, financial institutions such as China Merchants Bank and Haitong Securities, leading Internet companies such as JD.com, Xiaomi and China Literature, and academic institutions such as Tsinghua University and Shanghai Jiao Tong University.

In addition, in the era of large models, SenseTime's forward-looking layout continues. Previously, SenseTime pioneered the "Model-as-a-Service" business model, which focuses on software invocation capabilities, enabling customers to easily fine-tune large devices and invoke various generative AI capabilities without having to manage the underlying infrastructure.

In the "2023 China AI Development Platform Market Report" released by Frost & Sullivan, SenseTime ranked first in China in the comprehensive score of the two evaluation dimensions of growth index and innovation index. This achievement not only demonstrates SenseTime's leading position in the field of generative AI, but also expresses recognition of its innovation capabilities and the speed of market expansion.

Technological innovation

Drive continued business growth

SenseTime also mentioned in its annual report that the current generative AI model capabilities are far from reaching the upper limit, and the capabilities of the "daily new" large models will be significantly improved every three months, so that the application in enterprises will also be upgraded.

At present, SenseTime is working on the research and development of the "Ririxin 5.0" version, which is scheduled to be released in April, which will fully meet the standard of GPT-4-turbo, and at the same time support millions of words lossless context, and its multimodal model capabilities will also be fully benchmarked against GPT4V.

The continuous iteration and optimization of SenseTime's "DailyNew" large model not only reflects the company's technological pursuit and innovation in the field of artificial intelligence, but also continues to reflect its determination and ability to transform cutting-edge technology into practical application value.

For example, in the financial industry, SenseTime has significantly improved the model's data processing and question-answering capabilities by refining the model's Retrieval Enhanced Generation (RAG) technology. Using RAG technology, financial data is efficiently mapped to the vector database, which strengthens the financial knowledge reserve of the model and effectively reduces the incidence of misinformation generation (commonly known as "hallucination" phenomenon). This breakthrough enables the model to respond accurately to professional financial queries, so many well-known banks, including Bank of China, China Merchants Bank, Industrial and Commercial Bank of China, and Bank of Shanghai, have adopted the "Daily New" model to build their "digital employee" and "digital customer service" systems to provide a more intelligent customer service experience.

In the field of smart terminals, SenseTime also mentioned that 2024 is the year of the explosion of device-side large model applications. Based on years of accumulation in smart terminals, SenseTime has built industry-leading small models with different specifications. For example, with the same 7B model, SenseTime's performance is ahead of Meta's Llama2 and Google's Gemma. In 2023, Qualcomm and MTK showcased the application results of SenseTime's subsidiary's Hui Li's generative AI device-side model on its flagship chip at the press conference, and SenseTime's 7B small model achieved an industry-leading inference speed of 16 tokens/s on Qualcomm's latest chip.

These smart terminal models are essential for AI phones, as they can listen to the user's instructions and invoke various applications on the phone to complete complex tasks. SenseTime has partnered with mobile phone brands such as Xiaomi and Honor to jointly develop these innovative features.

In 2024, "technology leadership" will still be the key word of SenseTime.

SenseTime also said that it will further give full play to the in-depth synergy advantages of "large equipment + large model", continue to expand the scale of computing power, and improve the comprehensive service capabilities of infrastructure and large models. We will continue to invest in the iteration of the "daily new" large model system to maintain the industry-leading model capabilities. At the same time, it will also be committed to realizing the "Moore's Law" to improve the inference efficiency of large models, and through innovative optimization strategies such as cloud + terminal combination, the cost of model inference will be rapidly reduced, so as to open up broader application prospects.

Based on its technological leadership, SenseTime's business growth is also expected to be accelerated. SenseTime said that it will accelerate the commercialization process and market penetration of generative AI, and provide users with generative AI solutions with the best cost performance.

SenseTime said in its annual report, "Scaling Law plays a crucial role in our journey towards artificial general intelligence (AGI). The breakthroughs we've witnessed over the past year have only marked the beginning of this journey. The capabilities of generative AI models are not yet at their peak, and we expect these models to continue to grow and enable higher-level applications in the enterprise. ”

(This article does not constitute any investment advice, and the information disclosure content is subject to the company's announcement.) Investors act accordingly at their own risk. ）

Edited by Meng Jintao

National Business Daily