Hi, everyone. Welcome to SenseTime Group's 2024 mid-year results briefing. Our host is Shen Xuwei. Joining us today from management are Dr. Xu Li, Chairman and CEO of SenseTime, Mr. Xu Bing, Co-founder and Executive Director, and Mr. Wang Zheng, Chief Financial Officer. Let me begin by reading the disclaimer. This discussion may contain forward-looking statements. Forward-looking statements involve inherent risks and uncertainties that could cause actual results to differ materially from current expectations.
For a detailed discussion of such risks and uncertainties, please refer to SenseTime Group's latest announcements filed with the HKEX. All forward-looking statements in this briefing are based on assumptions that we believe to be reasonable as of today. SenseTime Group does not undertake any obligation to update these statements unless required by applicable law. This discussion also contains certain non-IFRS financial measures that are for comparative purposes only.
Furthermore, during today's meeting, management will make pre-prepared remarks in Chinese. Interpreters will provide simultaneous translation in English on the English channel. Please refer to our results announcement for details. In the event of any discrepancy between Chinese and English, please refer to the original Chinese statements by management. Dr. Xu Li will begin by providing an overview of SenseTime's performance. Mr. Xu Bing will provide more details on each business progress, followed by Mr. Wang Zheng, who will review the company's financial performance. After the management presentation, we will hold a brief Q&A session. Now over to Dr. Xu Li.
Thank you, Xu Wei. Welcome, everyone, to SenseTime's 2024 mid-year results briefing. Let me begin by highlighting a few key areas for the first half of 2024 . This is the second time we are reporting our results since our strategic pivot towards generative AI last year, and we have witnessed a positive trajectory. Group revenue for the first half reached RMB 1.74 billion, representing a 21.4% increase. Overseas generative AI exceeded expectations, achieving a remarkable 256% growth, generating revenue of RMB 1.05 billion, contributing 60% of total group revenue.
This signifies that SenseTime transition towards Gen AI has been more successful than we expected. The Gen AI market is fiercely competitive, but our deep synergy between large models and SenseCore enables us to rapidly enhance model capabilities, reduce inference costs, and create economies of scale. IDC has recently published two reports regarding the market share.
Our LLM platform and application captured 15% market share, placing us number two in the industry. Our AI business services achieved 15%, ranking third. Gen AI presents numerous commercial opportunities. Many companies are investing heavily in large models, regardless of cost, to seize these opportunities. For first time, we must not only ensure competitiveness, but also actively manage our operating costs and expenses. Our R&D remains our largest expense. We are also taking steps to balance benefits of long-term growth and short-term investment. In the first half, we secured a leading position in large model competition while achieving 26.5% reduction in EBITDA loss and 21.2% reduction in overall loss. Generative AI is one of the hottest investment themes in 2024. The Chinese market is also developing rapidly.
In China, large models and computing power are recognized as new quality productive forces, improving efficiency across sectors. According to the IDC latest report, AI will generate a RMB 1 trillion market in China from 2024 to 2028. Among which, internet industry clients were the first to use large models and intelligent computing. In the first half of this year, more industries began to embrace Gen AI technologies such as intelligent hardware, electric vehicle, robotics, healthcare, and finance, leading to a few massive vertical markets emergence. SenseTime is well positioned to capitalize on the next wave of growth. Our SenseNova large model and large scale AI cloud services are fully operational, and we have acquired a wave of clients driven by the essential needs for training trillion-parameter LLMs, training industry-specific models, and conducting large-scale, expandable, low-cost AI inference.
We provide comprehensive enterprise-level generative AI solutions for the market, covering from cloud to on-premise data centers to edge devices. Let us then further break down into three major business segments, which changed significantly as compared to last year, first half. Firstly, generative AI. Revenue reached RMB 1.05 billion, becoming the largest revenue contributor, accounting for 60% of group's revenue, up from 21% last year. Secondly, SenseAuto revenue reached RMB 158.1 million, doubling from last year, accounting for 10% of group's revenue, up from 6% last year. Third, traditional AI. First half revenue was RMB 520.4 million, contributing 30% of group revenue. The evolution of these three segments reflects our strategy.
Furthermore, our overseas market revenue has experienced healthy growth, reaching a year-on-year growth of 40%, exceeding the group's overall 21% growth rate. Overseas revenue now accounts for 18% of the group's total revenue, reflecting the robust demand for AI in international markets. Next, I will hand over to Xu Bing, who will discuss the three, the business in more detail.
Thank you, Xu Li. Last year, we have pivoted to GenAI. Our generative AI revenue surged by 200%, achieved a total revenue of RMB 1.2 billion for the entire year. In the first half of this year, it continued its amazing growth trajectory from last year, exceeding RMB 1 billion revenue within just six months, nearing the annual revenue of 2023 full year.
This business also has a unique feature. Billing occurs quarterly or even monthly, resulting in favorable cash flow and strong sustainability. In each quarter, we've witnessed clients expanding their usage of model inference and computing power. Over 3,000 industry users are using our products and services. We will share more examples later in our presentation. Generative AI business is experiencing explosive growth. In 2022, it contributed only 10% of group's revenue. In 2023, it became 35%. In the first half of this year, it exceeded 60%. This signifies that our strategic pivot and upgrade towards GenAI has achieved more than we expected. Our core advantage lies in the synergy between SenseCore and large models.
It not only enables our models to iterate faster, but also makes sure our expensive computing resources are fully utilized, keeping the inference cost down, achieving elastic inference, generating economies of scale sooner, and maintaining an advantage in the market. The development of GenAI has created a massive demand for computing power. Intelligent computing has become an essential infrastructure for training and inferencing large models. SenseTime is among the first AI companies in China to invest in the field. Before our listing in 2021, we had already plowing in the field, invested RMB 5 billion R&D investment in Shanghai Lingang AIDC, the largest AIDC in Asia at the time. In the first half of this year, the demand was very strong. We also further expanded our scale and enhanced service capabilities.
By August of this year, the number of GPUs deployed in our AI cloud exceeded 50,000 chips, and the total computing power exceeded 20,000 petaflops. This scale is quite scarce in China. We built the first 1,000-GPU cluster for our own use back in 2019. We were among the first to integrate heterogeneous domestic chips, collaborating with domestic GPU players. We now have the capability to connect tens of thousands of GPUs with a goal of achieving 100,000-GPU connectivity by 2024 to 2025. Our target within this year is to expand our total operational computing power to 25,000 petaflops. The software and toolchains we provide for AI infrastructure are also among the best and the most comprehensive. Scale alone is insufficient. We have also invested heavily in engineering techniques for energy efficiency, stability, and cost effectiveness.
Currently, AIDC are being built around the world, and we have the capabilities to design, construct, and flexibly schedule operations for clusters ranging from thousands to tens of thousands of GPUs. This makes us a key partner for many AIDC investors. This year marks the first year of large-scale model inference. Our original algorithm frameworks, including SenseCarve and OpenPPL, provide inference optimization from cloud to edge. In the first half of this year, we launched a large-scale model inference simulator that predicts the optimal inference performance of different models with various hardware configurations.
Combined with our own self-developed inference engine, we achieved a four times improvement in queries per second at the same level of electricity and computing power consumption, significantly enhancing the cost effectiveness of our inference services. Regarding commercialization, intelligent computing services have vast market potential, and the development is still in very early stage.
According to an IDC report in July, the Chinese intelligent computing services market is projected to grow at a CAGR of over 50% for the next five years, reaching a total market size of nearly RMB 200 billion by 2028. Training size demand is expected to increase more than tenfold, while inference size demand will surge over 200 x. SenseTime concurrently holds a market share of about 15%, ranking third in China. The companies ahead of us and behind us are internet giants, whose CapEx investments far exceed ours.
This shows that our products can meet clients' imminent needs better. Our clients include Jingdong, Xiaomi, Kingsoft Office, three major telecom operators, automakers like Geely, universities like Tsinghua University, and leading LLM startups. Our intelligent computing capabilities are also recognized at the national level. After rigorous review by China Electronics Technology Standardization Institute.
Technology Standardization Institute, SenseTime was awarded the enhanced level in the computing power service capability maturity model assessment, which is the highest honor and the first in China. Now let's look at the performance of our large model. The SenseNova large model series is a combination of the company's cutting-edge R&D resources and tens of thousands of GPU power. It has been iterated from version 1 in late 2022 to version 5.5, released in this July. This has been a race against time. The development of domestic and international models has been incredibly fast in the past two years. To maintain a leading position, we need extensive computing power, the best talent, and feedback from our clients. In this April, the release of SenseNova 5.0 marked a significant milestone for Chinese developed LLMs.
It became the first model to surpass GPT-4 Turbo in third-party benchmarks such as SuperCLUE and ranked first among Chinese LLMs in the AutoArena benchmark, conducted by Alibaba DAMO Academy for Chinese capabilities, also surpassing GPT-4 Turbo. This July, SenseNova was further upgraded to version 5.5, exhibiting a 30% improvement in overall capabilities, and its multimodal capabilities have been comprehensively enhanced, achieving real-time interactive experiences comparable to GPT-4o, making it the first multimodal real-time interactive LLM in China.
SenseNova 5.5 not only enhanced fundamental capabilities, but also focused on synthesizing high-order train of thought data. This significantly improves the model's inference capability. Multimodality has always been a strength of SenseTime. We have accumulated substantial data and know-how, covering numerous scenarios. SenseNova 5.5 even surpasses SOTA level of overseas model in multimodal capabilities.
It also offers an excellent real-time interactive experience with a minimum first batch delay of 16 milliseconds, attracting a large number of intelligent hardware and robotic users. We have distilled various sizes of models based on SenseNova 5.5 to cover cloud, on-premises deployment, and edge devices. We have also expanded our customer base beyond internet companies to include intelligent hardware, electric vehicles, robotics, healthcare, and finance. The overall usage of SenseNova increased by multiple folds since beginning of this year to July, with a substantial increase in both the number of users and average usage per user. We understand the market is very keen to understand our commercialization progress.
For example, in finance, enterprise clients including Bank of China, China Merchants Bank, and Haitong Securities, leverage our financial LLM capability to develop financial vertical applications such as employee assistance, compliance, risk control, coding assistance, digital employees, marketing and sales empowerment, et cetera. In healthcare, major hospitals such as Ruijin Hospital and Xinhua Hospital adopt our SenseDaYi model, making LLM as the intelligent hub of hospitals.
This enables intelligent pre-consultation, post-consultation follow-up, drug consultation, et cetera. For copilot assistants, our coding copilot and office copilot product has become one of the fastest growing copilot products in China. Large enterprise users like China Telecom and Kingsoft Office, along with hundreds of thousands of individual users and developers, have become our users. Anthropomorphic interaction. Our anthropomorphic LLM powers apps like Weibo, China Literature, iQIYI, and IdeaFlow. Daily usage has increased by more than 20 times within the past six months.
On smart device, our edge AI model maintains a leading position in terms of inference speed and model capabilities, delivering smooth experience on smartphones, automobiles, personal computers, and smart speaker. Inference speed, which is up to 100 tokens per second, serving clients such as Xiaomi and Oppo. As of July, the number of enterprise users for SenseNova API increased by more than seven times. With each new version upgrade, we see significant growth. The usage by our top clients is also surging. In the middle is a graph of one of our internet clients. Its average monthly usage of the client from January to July increased by 900%. Daily token consumption reached more than 10 billion. On the right-hand side, a certain mobile phone client. Usage has increased by 3.5 x in the past two months.
The main reason for this growth is continuous penetration into more terminal devices. The penetration starts from high-end devices initially, but as inference costs come down, LLM starts to penetrate into mid to low-end models, leading to a 3.5 x growth. Overall, the decline in inference costs is a crucial factor, driving large-scale ADAS application, and this requires joint optimization of algorithm and hardware, which is our core advantage. Finally, a bonus, our SenseMirage product. At the end of July, SenseMirage launched a trial product called Shupai, which can be searched on WeChat mini program. This product can generate many interesting selfies and group photos within seconds. We reached 1 billion users benchmark in the first nine days of trial operation, achieving over 500,000 DAUs. Now it has accumulated 3 million users and generated 20 million photos within a month.
We invite you to try it out. Our international business is steadily growing, achieving a 40% year-on-year growth, exceeding the group's overall revenue growth rate. In Saudi Arabia, we collaborated with government agencies to provide AI courses to over 2,000 teachers and over 25,000 students. In the UAE, we created an immersive generative AI experience for Yas Island, a popular vacation destination. We launched localized companies in Thailand, Thai language, in Hong Kong and Thailand, respectively. Sorry, Thai language LLMs in Hong Kong and Thailand, respectively. Localized models are better suited to local needs and cultural understanding, and we believe that each country and region will require its own localized LLM. Now let's move on to SenseAuto business.
In the first half of this year, SenseAuto achieved total revenue of RMB 159 million, representing a 100% year-on-year growth, accounting for 10% of group's revenue. Growth was driven by mass production of both Smart Pilot and Smart Cabin products. 710,000 new vehicles were delivered, representing an 82% growth. We have expanded to 15 more new models in the first half, including overseas brands such as Volkswagen and Lexus. We also participated in several overseas projects, enhancing our global footprint and influence. There are three competitive advantages for SenseAuto. Firstly, UniAD based end-to-end intelligent driving, being among the earliest in China to adopt a pure visual technology. Secondly, multimodal Smart Cabin based on SenseNova 5.5, which redefines a new interaction experience. And thirdly, AI cloud services for automakers.
Leveraging the computing power of over fifty thousand GPUs in SenseCore, providing intelligent driving, training, and data services to auto makers. Let's take a closer look at two of our major mass-produced smart auto products. Firstly, in Smart Pilot, SenseAuto leads the development of mass production of end-to-end autonomous driving, known as UniAD, supported by massive computing power and high-quality data. UniAD provides a higher ceiling for intelligent driving capabilities compared to traditional solutions. At the 2024 Beijing Auto Show, UniAD successfully demonstrated real-world testing, navigating complex real-world conditions solely through visual perception with our high-precision maps. This achievement was recognized as China's own FSD. Our solutions are in mass production for GAC, FAW, and Hozon, et cetera.
Secondly, in Smart Cabin, based on SenseNova Multimodal Large Model and text-to-image models, we have established partnerships with Audi, Volkswagen, Honda, BMW, Xiaomi, and Geely, et cetera. To develop in-cabin LLMs, we have deployed SenseNova 5.5 on vehicles, reducing the first batch delay to 50 milliseconds, allowing real-time in-cabin human computer interaction to be incredibly smooth. Some of the more notable vehicles released in the first half include Xiaomi SU7, IM Motors, and Geely's LEVC. Now, I would like to hand over to our CFO, Mr. Wang Zheng, to discuss the company's financials.
Good evening, everyone. I'm delighted to share with you our financial performance for the first half of 2024. As a reminder, our business historically has a degree of seasonality fluctuation. Revenue in the first half has historically accounted for slightly more than one-third of full year revenue. Majority of our business typically occurs in the second half of the year. We achieved a revenue growth of 21.4% in first half. Despite a significant external challenge and ongoing economic uncertainty, our focus on transitioning to generative AI has delivered remarkable results, steering the company back onto a path of healthy growth. Revenue related to Gen AI has become the primary driver of SenseTime revenue growth. This segment continued its rapid growth from twenty twenty-three, further increasing year on year, 266% in the first half.
The proportion of Gen AI revenue has jumped from 20.6% in the first half of 2023, to now 60.4% in the first half. Our core advantage in Gen AI lies with deep synergy between large scale infrastructure and large models, enabling us to rapidly iterate models and reduce inference costs. That is far ahead of the fierce competitive market. This business possesses a high barrier to entry, building through accumulated long-term talent pool, technology, and capital investment. Our intelligent vehicle business also doubled its revenue in the first half of this year. We have already discussed the highlights, and hence, we will not elaborate further here. The only segment experiencing a decline in revenue is traditional AI. This segment certainly faces various challenges, and we have implemented a series of strategic investments.
Particularly, proactively, we further contract traditional AI business related to smart cities. It is worth mentioning our overseas business. Commercialization of GenAI in overseas market generally is lower than in China, yet our overseas business achieved a high growth rate of 40% in first half. Every region, including Northeast Asia, Southeast Asia, and other regions such as Middle East, have all recorded higher revenue growth rates than China. The proportion of overseas revenue to total revenue has rose to 18.5% in the first half. Our gross profit margin reached 44% in first half of this year, on par with full year gross profit last year. Recovery and growth of gross profit, coupled with our continued strong control over operating costs, leading to a significant narrowing of both EBITDA and net profit losses.
In the first half, it reached to narrowing of 26.5% and 21.2% respectively. Pace of narrowing of losses is significantly faster than a year ago. On operating expense, we have maintained a consistent level of control while also making adjustments to ensure our long-term competitiveness. Looking on the right-hand side of the slide, total operating expenses decreased by 1.4% in first half. Overall, selling expenses saw the largest decline, reaching 21.2% or 18% decrease compared to previous half year. This positive trend benefits from the optimization of company revenue structure and business model. Management expenses declined by 7.7% versus first half last year. Excluding equity-based compensation costs, management expenses declined sequentially. Among three major expenses, research and development saw a slight increase of 6.1%.
We strategically allocated more computing resources to our R&D team to ensure our leading position of ten times as SenseNova large language model. Excluding computing-related costs, other R&D costs have decreased overall. Our overall working capital turnover efficiency continued to face challenges. On the left-hand side of the slide shows the number of days for each key working capital indicator, calculating using the end of period value. This metric indicates our cash conversion cycle lengthened in the first half of this year. This change primarily driven by two factors. We strategically made some inventory purchases related to computing power increased inventory days. This is aimed to better support our continued rapid growth of GenAI. Historically, our accounts payable balance also closely related to our traditional AI business.
The decline in this business in the first half resulted in a decrease in absolute value of accounts payable and the number of days. Our trade receivable days have continued to decline, decreasing by more than 100 days compared to a year ago. The bottom right corner of the slide shows our trade receivable recovery remains at a relatively high level. Moreover, as company rapidly transitions to GenAI, our revenue quality will continue improve. The average recovery speed for revenue generated since 2023 has accelerated. Historically, trade receivables with longer aging are associated with smart cities, which face significant challenges due to macroeconomic conditions, which led to increase in proportion of bad debt provision on the asset side of the balance sheet shown in top right corner. However, the rate of increase has begun to slow.
Our capital expenditure in first half of this year mainly focused on building large-scale computing power, but this investment declined slightly year on year and sequentially. Partially, this is because we already made significant investments in computing resource ahead of the curve. With the introduction of increasing number of partners, has given us opportunity to control more computing power with limited funding. Total cash reserves and net cash reached RMB 9.9 billion and RMB 4.6 billion. This definition of cash includes structured deposits, but does not include equity and bond investment balance, which reached RMB 6.7 billion.
This balance on the right is fair value concept, with bond investment managed by third-party professional institutions, including investment-grade U.S. dollar bonds. The fair value of this sector is growing steadily. Therefore, we have sufficient resources to focus on creating and seizing long-term opportunities in generative AI. This concludes our financial review. Thank you.
With the following, we'll go into our Q&A section, and with the first question coming from CICC analyst, Brenda Zhao.
Good afternoon, Mr. Xu Li, Mr. Xu Bing, and Mr. Wang Zheng. Thank you for taking my question. First of all, I would like to congratulate the company on the impressive growth of generative AI. My question is about edge AI. We are seeing that the focus of AI deployment in 2024 is shifting from killer apps to edge deployment. SenseTime has a clear advantage in edge models, in your collaboration with smartphone manufacturers, what are the potential applications for edge AI? How do you view that AI capability is implemented in Apple Intelligence and the Pixel 9, please?
Thank you. Response from Dr. Xu Li: Thank you, Ms. Zhao, for your question. First of all, we are very optimistic about the prospects of edge AI. While the exploration of killer apps for Gen AI is still ongoing, from our perspective, the user base from mass is growing rapidly. We believe that as a user base expands, new application models will gradually emerge. Regarding the application you mentioned, such as those in Apple and Pixel 9 devices, they have indeed provided us with significant insights. However, when we consider edge AI, we do not limit ourselves to smartphones alone.
We also consider the use cases in various IoT devices, such as in-car systems, PCs, smart glasses, headphones, and speakers. These applications will be key drivers of large language model deployment in the market, and we need to strategically plan for them, particularly in China. With the extensive proliferation of IoT devices and our early mover advantage in the mobile internet era, we see substantial opportunities in edge AI. As we have highlighted in previous tech days, we are highly focused on enhancing the inference efficiency on the edge.
We have optimized edge side chips, enabling our LLM to achieve inference speeds of over one hundred characters per second on edge devices, with the latency reduced to just a few tens of milliseconds. Additionally, we have innovated in edge inference architecture, implementing a cloud edge collaboration computation model, where most computations occur at the edge.
While most complex tasks are handled in the cloud, our early adoption of heterogeneous MoE models has also given us a cost advantage. In practical applications, the use of models on the edge is particularly critical for multi-model inputs, especially video inputs. In pure language interactions, there are still challenges in achieving natural interactions. Truly natural interactions should innovate and involve minimal input while enabling strong interaction and information output.
Hence, leveraging SenseTime's strength in the computer vision to develop a comprehensive, real-time, multimodal interaction model is of paramount importance to us. For example, our SenseNova 5.0 multimodal large model is already capable of streamlined interactions, which has driven the development of the next stage of interactive applications. We are very optimistic about the application of real-time interaction, video-based, and multimodal large models.
We believe this will be widely adopted in the applications for downstream devices, especially in the IoT devices. We expect that edge and edge cloud computing will be applied across various IoT devices broadly in the future. As mentioned earlier, the rapid increase in API costs of our large language model in mobile apps is largely due to our success in reducing inference costs, allowing coverage of more edge devices. This is why we are progressively expanding our coverage from high-end to mid-range, and eventually to low-end smartphones, with potential future expansion into speakers, smart home devices, and interactive TVs. All these devices represent ideal application scenarios for edge models, so we remain highly confident in this area, and our optimization will continue to focus on these practical applications.
Thank you.
And the second question comes from Haitong, Mr. Yang Ling.
Good evening, Mr. Xu Li, Xu Bing, and Mr. Wang. My question is regarding to computing power. I noticed that in the first half of 2024, the growth rate of computing power was still very fast, and there is still a global shortage. How are we planning to scale our computing power resources to maintain our core competitiveness? My second question is about domestic computing power. What progress have we made in this area?
The response from Dr. Xu Li: Western companies focus on scaling law, where no major firm plans for less than a hundred thousand GPUs, with everyone aiming for a hundred thousand GPUs and trillion-parameter models. This is a significant trend. If we want to truly benchmark against this, I believe that in China, having a fully connected computing power resource is still crucial for models. Unlike large corporations that can invest heavily in capital expenditures and CapEx, we have to use a more strategic approach.
We operate computing power. This operational software helps improve efficiency, which in turn enhance the resilience of the entire operation. In the first half of the year, we also followed this logic by improving efficiency while expanding the scale of operational computing power. Of course, this is a step-by-step process, and we are currently in a stage that is neither too far off nor too close. This relative asset light model will cause some decline in gross margins, but it will actually reduces cash burn. I think this is beneficial and it strengthens our resilience. We see this as a promising development model. Regarding localization, AI inference volumes are currently very large.
You can see that a single app consumes tens of billions of tokens, and this can quickly reach hundreds of billions, and this is only on the text level. When we move into video and image domains, our experience shows that the demand for computing resources in inference is tremendous. There has been several instances where the servers crashed.
In terms of domestic computing power for inference, the costs are actually quite favorable, and you can also handle some scenario-specific vertical training. So with this training inference model, there is no need for domestic computing power to connect at extremely high scales. The key is to achieve a favorable cost-performance ratio, which is feasible. In certain industrial applications, as long as the cost-performance ratio is on point, it can scale. That's why in our future construction localization-focused applications, there is a large market space for domestic computing power.
Regarding the larger connections exceeding 100,000 GPUs, it's really about the connectivity capability, including topology, scaling, crash resistance, and disconnection handling. These are vertical issues we need to solve urgently. For example, if the system crashes every hour, how can we divide the training into segments to ensure stability and enable multitasking? This is actually the core capability of the software we are advancing. So to answer your question, I'm optimistic about the trend of cost reduction in domestic computing power for inference. In certain vertical field, domestic computing power may actually offer a considerable advantage for training and inference.
Next, question number three from Huatai, Mr. Xie Chunsheng.
"Hello, I am Xie Chunsheng from Huatai. What are the core capabilities of the next-generation large model? This question has several backgrounds The first one is that I saw in your report that your large language model in July was the first in China to achieve GPT-4.0 level capabilities, realizing multimodal interaction. The company's model capabilities are at the leading level in the industry and country, definitely.
The second background is that at this stage of AI development, everyone is very concerned about whether the capabilities of the next-generation model will be significantly improved comparing with the past. In fact, everyone is watching to see if the scaling law has slowed. Everyone is also looking forward to the improvement of the capabilities of large models. But at the same time, it is difficult to describe what the core capabilities of the next generation of large models are. For example, like GPT-5 or Llama-4, the core model capabilities is hard to describe quantitatively.
The third background is that after we tracked these large model manufacturers overseas, whether it is OpenAI or Google, everyone will mention one word when describing their next-generation model capabilities, which is reasoning. But reasoning is very abstract and difficult to understand. So I would like to ask, from an industry point of view, what are the core capabilities of a next-generation model? How does the core capability correspond to application scenarios, please?
The response from Dr. Xu Li: Thank you. This is a big question. First, the performance decline of scaling law is actually a mathematical indicator, which we call loss function. When it comes to real-world subjective issues, is there a ceiling for scaling law? I think it's very possible to encounter. So is it just about expanding the model itself to improve performance, or about constructing higher-level chains of thoughts, data to form better reasoning?
I basically divide data into three layers. Let me first talk about the first two layers. The first layer is knowledge points. When knowledge is input into the model, it can be better memorized, compressed under certain circumstances, and then provide feedback or even generate new content. The second layer is the relationship behind these knowledge points. That is reasoning from A to B. Why is there such a reasoning? Only by reasoning step by step, can we form higher-order intelligence. Whether a model is intelligent or not, depends entirely on whether there is enough high-order data, chains of thoughts data, and whether the chain is complete. In many vertical industries, the improvement of model performance come from constructing good vertical domain data. For the further development of large models, we still need to supplement better data. First, the supplementation of knowledge points is there-...
Current data sufficient, or will the data dry up in the future? This is a realistic question. SenseTime has an advantage in what we have accumulated the visual data as supplementary data, which makes our multimodal capabilities stronger. In a multimodal context, our original perception capabilities, such as act as a translator in processing raw data, enabling large models to learn the underlying knowledge more quickly.
At the same time, the high-order chain of thought data we construct in many vertical industries will join influentially the training of the entire large model. Furthermore, the thinking ability of large models come from data and iteration in vertical industry applications. Such a closed loop is particularly important. Whether a model is intelligent or not, depends entirely on whether the method of constructing the chain-of-thought data is strong enough, and whether this methodology can be sustained and iterated.
Based on isomorphic data, I believe the West will continue to push to the limit of scaling law. Given their access to hundreds of thousands of GPUs, they will certainly reach the boundary, but is this method the most efficient one? I don't think so. There must be better ways to improve model capabilities. Apart from data, the model's architecture is also crucial. The architecture of the model has been adjusted at each stage. In the future, with the integration of trained inference, the model's cost effectiveness will be higher, which is also the improvement point of our next generation model. This includes not only the efficiency of training and inference with individual modules of the model, but also a new type of multi-cluster training or even multi-level training, involving joint training of models at the end, edge, and cloud levels.
The improvement in model capabilities in this area is also highly relevant to industry applications. In summary, first, we need to make the model perform well in various scenarios. Second, we need to make the model affordable for use in those scenarios. Only by advancing on both fronts, can we achieve a competitive advantage in the next stage of model development. From our perspective, we remain highly focused and are pushing hard in these two directions. Thank you.
From Mr. Xie Chunsheng: "You've clearly explained the path to improving reasoning capabilities, and I've learned a lot from it. Thank you."
And next, question number four from analyst from HSBC, Ms. Helen Fong.
Thank you all for giving us the opportunity to ask questions. I would like to extend my congratulations to the company on its remarkable growth in generative AI revenue from the first half of the year. I would also like to seize this occasion to inquire which of our products or services predominantly contribute to this increase. Moreover, considering the long-term perspective, what are future growth expectations from generative AI, and what will be the main driver for this growth, please?
An answer from the management: Thanks for your question. It is indeed the direction that our entire management team currently prioritizes the most, which is the commercialization of our robust technological expertise accumulated over the past years, especially in AI infrastructure and large models. We have actively been exploring commercial expansion and firmly believe that only through successful commercialization can we further advance in the field of generative artificial intelligence.
Let us think about generative artificial intelligence, which has seen a significant breakthrough since the beginning of 2023, often referred to as the ChatGPT moment. Following this breakthrough, we have observed explosive business growth internationally. This surge is evident in AI infrastructure development, represented by companies like NVIDIA, which provides specialized GPUs and chips and have shown astonishing growth.
Additionally, there has been a rapid ascent of new companies to build up large models, scaling from tens of millions to billions in revenue with just two years. This exponential growth in generative AI is electrifying, and its competitive landscape is fierce, with every tech giant vying to capture this wave of opportunity. Domestically, the situation mirrors the international scene, with generative AI and large models being elevated to a strategic national priority, the new quality, productive forces, and receiving high market demand.
The challenge lies in an acute shortage of computing power and resources. Fortunately, our decade-long focus and investment in AI capabilities, particularly computing power, position us strongly. We have developed scarce yet valuable competencies, such as operational expertise in computing, networking, and dynamic scheduling of GPU resources, which are in high demand among investors and builders of new computational facilities.
Secondly, a multitude of domestic enterprises are keen on integrating generative AI to redefine next-generation products and user experiences, from AI-powered smartphones and computers to the intelligence of electric vehicles and various internet applications. These are highly competitive markets, actively seeking to invest in and utilize large language model technologies currently in the initial adoption phase. Our company stands as a principal provider of such large language model capabilities....
Benefiting from the expanding user basis and engagement of our clients, which in turn drives the increase in model usage, API calls, and token counts. This has resulted in our generative AI revenue doubling in just the first half of the year. Indeed, the industry is still in its early stage, given that it has only been around for two years. Therefore, we are also hoping to maintain high growth and a rapid pace over the next five years or even longer. To achieve such sustained high growth and rapid acceleration, an incredible focus is essential. Thus, it is challenging to diversify like mega tech companies by branching into numerous businesses. We must concentrate our resources. Currently, we are primarily focusing on the field of generative AI to ensure that we can effectively address and serve the specific needs and pain points of our users.
In the market report recently presented, you can observe that, for instance, in computing services, we rank third in the Chinese market. The entities ahead of us in the first and second positions are prominent internet giant companies, and the ones following behind in fourth and fifth places are also major players in the internet sector. While these companies may surpass us in terms of the scale of their investments, we have nonetheless sustained our position within the top five and proudly stand as the only non-internet company among the top three. This accomplishment truly highlights our superior technological and services capabilities, enabling us to robustly support and keep pace with the vigorous growth of the market. We have onboarded numerous new clients who, much like international enterprises, are expanding their computational resources and large language model usage.
This trend is expected to persist for several years as large language model capabilities continue to evolve and killer applications start to emerge, prompting future and further industrial investment. Currently, we are in a phase of infrastructure development. Sequoia Capital's recent report predicted 2025 to be the year of the data center, a significant explosion of AI infrastructure. We are well-positioned in this growth phase, focusing our investment and enhancement in AI infrastructure and large models to ensure strong technical and product competitiveness, thereby driving healthy growth and profitability.
Both aspects are at the very core of our focus within generative AI. We are committed to investing and enhancing our capabilities in these areas to ensure that we can achieve formidable technical competitiveness and product competitiveness in the future. Only by doing so, can we drive healthy growth and maintain strong profitability.
Due to the time constraint, we will take the one last question, which come from Ryan Wang of DBS.
From Mr. Ryan Wang: "Thanks for taking my question. I am Ryan Wang from DBS. We have observed since Tesla's release of FSD last year, the end-to-end approach has gradually become a consensus in the autonomous driving industry. We have also clearly seen the SenseAuto of intelligent automotive sector achieving doubling growth in the first half of this year. Therefore, I would like to inquire about the company's current progress in commercializing end-to-end algorithm, and how the autonomous driving algorithms synergize with SenseTime's underlying large models, please?"
The question is being answered by the management. Thank you for your question. I will address this swiftly. Regarding our SenseAuto autonomous driving, it is known that SenseTime has been dedicated to this field for many years, always adhering to a pure visual technology path. We had mentioned earlier that we developed an extensive visual neural network to drive the recognition of the entire world. These efforts are aligned with the technological trajectory we have been pursuing.
Due to our consistent commitment to this path, we have continuously expanded our computational resources. It is worth noting that before the advent of ChatGPT, a certain foreign automaker was actively increasing its compute, namely Tesla. Similarly, Tesla has insisted on using an increasing amount of visual data to train more potent neural networks. This approach demands a substantial amount of data and computing power. Domestically, there is also a company that continuously expands its compute resources to train autonomous driving technologies.
SenseAuto has always been one of the major consumers of our computational power, consistently seeking to employ more resources to drive a continuous loop of data feedback. Currently, we also utilize a significant amount of synthetic data, hoping to trigger a ChatGPT moment in the field of autonomous driving. Of course, in the past two years, Tesla has indeed achieved this. Its new FSD version, based on end-to-end large language model, has significantly enhanced the upper limits of intelligent driving capabilities. Many people are already familiar with and looking forward to this product's experiences. However, in China, currently, no automaker possesses enough computing power to rival Tesla. Thus, many automakers have turned their attention to our scarce computational resources with over 50,000 GPUs to train their models.
Historically, we have also served these automakers with various in-vehicle technologies for mass production, providing various level two and level three traditional autonomous driving technologies. Now, the value of a one-piece end-to-end neural network is evident, and automakers wish to reach and integrate our UniAD's technological capabilities for mass production.
We have the opportunity, especially under the backdrop of cooperation with automakers, to drive domestic automakers to develop a native version of FSD technology at this critical juncture, when Tesla FSD enters the Chinese market. The significance of this matter is self-evident, as this technological path is of great importance and requires strategic planning. Traditionally, most automakers have adopted a two-stage approach, separating recognition and decision-making within an AI technology framework. However, this framework has reached its limit and can no longer keep pace with the evolution of FSD.
It is essential to move towards an evolution driven by ultra-large-scale computation and high-quality data towards a one-piece neural network. I believe this will drive better growth in our SenseAuto sector in the future. Our target for mass production is set for 2025. At this stage, our team is indeed very busy collaborating with various automakers on joint technology development to ensure we meet the mass production timeline in 2025.
And with this, we conclude the Q&A section of our earnings results today. We thank you for dialing in and joining our senior management to share our stellar results delivered and achieved in the first half. Good evening and good night!