Raising 'lobsters' drives up computing power demand! Where are the investment opportunities?
SenseTime 2025 Annual Earnings Presentation
[AI Key Points Summary]
Financial Performance
- Revenue growth of 32.9% in 2025, generative AI business revenue reached 3.6 billion RMB, a year-on-year increase of 51%
- Gross margin at 40.0%, slightly down from 42.9% in 2024 but still better than market guidance
- EBITDA and net profit losses narrowed by 85.0% and 58.6% respectively, with EBITDA turning positive for the first time in the second half of 2025
- Cash reserves increased to 14.2 billion RMB, bank credit lines of 10.6 billion RMB grew 101% year-on-year
Business Progress
- Launched the Kaiwu 3.0 world model, with inference speed 72 times that of NVIDIA cosmos 2.5
- Officially launched and open-sourced the next-generation native multimodal model architecture Leo, achieving convergence with only one-tenth of the data volume of equivalent performance models in the industry
- SenseTime Xiaohuanxiong selected as one of Frost & Sullivan's top ten most practical agents, serving thousands of enterprises and over 15 million individual users
- The large-scale device maintains the top share in China's native AI cloud market and received the MIIT software security excellence certification
Guidance for next quarter’s performance
- It is expected that the adjusted net profit level will continue to maintain a significant loss reduction trend in 2026
- Adjusted EBITDA has a high probability of turning positive for the full year
- A completely new model based on the Leo 2.0 architecture will be released in Q2 of 2026
- In 2026, the focus will be on increasing planning and promotion in domestic computing power
Opportunity
- Overseas business transitions from single-point breakthroughs to scaled replication; the first overseas domestic computing power cluster launched in the Middle East
- The first to validate a new scaling law under a native multimodal architecture unifying understanding and generation
- Collaborated with more than ten chip manufacturers including Huawei Ascend, Hygon, and Cambricon to release the SenseTime large-scale computing power cabin
- Achieved annual electricity cost savings of 7% through collaborative large-scale models, equivalent to reducing carbon emissions by 4,000 tons per 10,000 cards annually
Risk
- Facing severe volatility in the AI cloud market, there is uncertainty on both the supply chain side and market pricing
- The complex geopolitical environment in the Middle East could potentially impact overseas business development
- Under the challenge of process limitations, continuous investment in algorithm optimization and localization is required
[AI Conference Record]
Spokesperson
Looking at it, SenseTime's Kaiwu 5 model ranks first among independent third parties and OEMs in comprehensive capabilities. It achieves global leadership in multiple key technical indicators. Thanks to the joint optimization of the model and inference tools, Kaiwu 53.0’s inference speed reaches 72 times that of NVIDIA Cosmos 2.5, making it the first world model capable of driving a full-body entity on the edge.
So what makes this model so powerful? On the right, we have a video generated based on the Kaiwu 3.0 model. Simply put, it can deeply integrate with agent intelligence. The interactive videos it generates not only maintain scene coherence throughout but, more importantly, fully adhere to the laws of physical reality.
As everyone knows, the hardest part of embodied intelligence is enabling machines to understand physical laws. The emergence of the Kaiwu model marks a new milestone in capturing these physical laws, paving a much more efficient path for the training and implementation of embodied cooling solutions.
In exploring how to further push the limits of intelligence, by the second half of 2025, we gradually observed that the mainstream multimodal model architecture has hit significant limitations in its ability to further increase intelligence density.
Therefore, in the fourth quarter of 2025, we officially launched and open-sourced a new generation of native multimodal model architecture, which completely abandons the traditional design of visual encoders and backbone models stitched together, creating a unified language and vision end-to-end native multimodal architecture from the ground up.
This new architecture significantly improves model learning efficiency. We achieved SOTA performance with only one-tenth of the data volume and computing power compared to equivalent models in the industry. Moreover, Leo's potential doesn't stop there. In the second quarter of this year, we will release a brand-new model based on the second-generation Leo architecture.
We are the first in the industry to verify a new scaling law for unified understanding and generation within a native multimodal architecture. The graph displayed here shows the learning curve of independent methods in our new ideal unified understanding and generation architecture. We can see that its accuracy improves with a very steep slope, significantly surpassing the efficiency of the previous open-source benchmark for unified understanding and generation models, Vigo, which was pushed by ByteDance.
This highly efficient unification of understanding and generation, along with a multiple-fold leap in performance and cost-effectiveness, will open up more downstream application spaces for the entire AI ecosystem. Stay tuned.
Having discussed technology R&D, let's now look at the application logic of AI. Here, we reference an industry research report released by CICC showing that the scaled application of AI is evolving following a logic moving from high fault tolerance to low fault tolerance and from simplicity to complexity.
Currently, the fastest-growing commercial closed loops with the most mature business models are in the fields of office automation, marketing, education, finance, and video production. SenseTime’s layout in these areas predates the industry consensus by quite some time. Next, we’ll share examples of our key progress.
First, in the field of AI-powered office tools, SenseTime’s Xiaohuanxiong has been selected as one of Frost & Sullivan’s top ten most practical agents and is currently serving thousands of enterprises and over 15 million individual end-users. In terms of content creation, our SQL was the first in the domestic market to achieve single-episode output of three-minute continuous generation for a hundred episodes of short dramas while maintaining consistency across such long-form videos.
SQL is already serving over 300,000 creators with monthly active users exceeding 100,000. Meanwhile, we have also achieved success in AI marketing and educational scenarios. For example, RuYing digital humans cover over 2,000 e-commerce live streaming rooms, and Smart Education reaches more than 500 educational institutions.
After using SenseTime's RuYing and SenseTime’s educational products, their work efficiency has increased several times over. As model capabilities continue to advance, these mature scenarios will become the strongest drivers of our business growth.
In addition to productivity tools serving the B-end, SenseTime has also launched several hit applications in the consumer-grade market on the X-end. In the authoritative media QbitAI’s 2025 flagship AI Top 100 list, three out of four SenseTime products were selected.
As you can see, our Kapi Camera has topped the App Store charts in multiple overseas countries. In the AI Civilization Plus Life app, monthly new downloads have rapidly grown from fourth place last year to sixth place this year. Kapi Bookkeeping has also been rated as one of the top ten most popular agents domestically, achieving peak popularity right from launch.
This success is built upon our integrated capabilities from foundational models to application products, proving that SenseTime not only excels in hardcore technology but also accurately understands user needs to create the most popular AI-native applications. Looking ahead, we will continue to expand our C-end product portfolio, enabling SenseTime’s AI to truly permeate everyday life.
Next, I will hand over to Yang Fan to introduce the business behind the large-scale infrastructure. Thank you.
Yang Fan
The significant achievements and momentum in our model development and AI applications are inseparable from the robust large-scale infrastructure. Here, I want to emphasize that the value of this infrastructure is not just about providing computational resources; more importantly, it lies in the technical capabilities of platforms and services derived from a deep understanding of models, helping AI infrastructure continuously improve cost-effectiveness and stability across various fields.
Let me give an example of Light X to V, an AI framework open-sourced by SenseTime that is highly recognized in the field of video generation inference. Facing the challenge of process limitations, we focused on algorithm optimization and localization efforts. By deploying distillation and extreme quantization, we achieved performance on domestic hardware that surpasses mainstream overseas chips, offering better cost-performance for models and applications.
Currently, the number of downloads for open-source models based on Lite X Two V has exceeded ten million, consistently ranking among the top ten globally. Meanwhile, we are also committed to making domestic computing power not only usable but excellent. We have taken the lead in achieving heterogeneous mixed training and inference at the ten-thousand GPU-card level, supporting the major trend of supply chain development today.
To break through the energy efficiency bottleneck in AI development, with the support of CATL, we launched the industry's first large-scale compute-chain collaboration model. Its core is using AI task data to accurately predict electricity load, which directly translates into reduced costs through deep synergy.
After the system went online, annual electricity cost savings reached 7%, equivalent to reducing carbon emissions by four thousand tons per ten thousand units of computing power annually. This is roughly equivalent to the annual carbon emissions of three thousand Chinese households or the emissions generated by a compact car circling the Earth more than five hundred times.
Through such technological iterations last year, SenseTime collaborated with Huawei Ascend, Hygon, Cambricon, and more than ten other chip manufacturers to jointly release the SenseTime Large-Scale Infrastructure Computing Cap. Here, computing resources, platform tools, and industry models can be freely combined like products.
Particularly, we integrated the world’s first Huawei Ascend 910C liquid-cooled cluster and were the first to complete comprehensive adaptation for the Ascend 384-node system, demonstrating our end-to-end delivery capability from underlying infrastructure to AI platforms.
When this kind of large-scale operational experience deeply integrates with the domestic ecosystem, it becomes an extremely rare core competitive advantage. This scarcity is opening up global markets for us. In 2025, we launched the first overseas domestic computing power cluster in the Middle East.
It validates our engineering capabilities in complex heterogeneous environments as well as our asset-light overseas expansion model, achieving full commercial success. Through continuous technological iteration, the large-scale system has been able to continuously capture core market share and gain recognition from various research institutions.
According to the latest report by Frost & Sullivan, SenseTime's large-scale system firmly holds the largest share in China’s native AI cloud market. However, being the market leader is just a result; more importantly, our technology has achieved several highly prestigious 'firsts' in core standards within AI.
We are not only among the first enterprises to receive the MIIT software security Excellent-level certification, but also obtained the industry's first top-tier 5A Excellence-level certification in the CAICT computing power platform service capability test. In addition, in the IDC inference computing power services report released last August, the large-scale system received full marks across all key dimensions such as performance optimization and Thunder integration.
These authoritative endorsements strongly demonstrate that in the current computing power market, SenseTime’s large-scale system is synonymous with high performance and reliability, providing the most solid foundation for building trusted AI. The generated page of AI customer logos serves as even more direct proof.
The multi-modal model capabilities we discussed earlier, including industry empowerment and AI productivity, are not just our own assessments but have already been validated by numerous clients in real-world scenarios alongside us. Moving forward, we will continue to deepen our efforts in these industries, scale up operations, penetrate the markets thoroughly, and gradually turn these cases into replicable models for sustainable growth.
Next, allow me to introduce visual AI, which forms the core business foundation of SenseTime and represents our most critical competitive edge in the multi-modal era. By 2025, our main visual AI business segment, CV 2.0, is gaining increasing popularity among overseas customers.
Our existing customers continue to make repeat purchases while we accelerate our breakthroughs in traditional strongholds like Southeast Asia, Northeast Asia, and the Middle East. Customer loyalty is robust, with repurchase rates steadily rising. Meanwhile, over the past year, there has been a noticeable increase in attention towards the South American and European markets, with inquiries and pilot projects on the rise, indicating two things.
Firstly, after validation across multiple regions, the market adaptability of our products has been well proven. Secondly, the international influence of SenseTime's visual AI is expanding from regional to global, with overseas operations transitioning from single-point breakthroughs to scaled replication, showing strong growth momentum.
Here, I would like to report that in 2025, SenseTime will further deepen its 'One Plus X' strategic layout, leveraging the underlying capabilities of large models and spatial intelligence to build the most dynamic AI ecosystem matrix domestically. This 'mother-building empowering sub-build冲锋' model unleashed tremendous vitality in 2025.
For instance, SenseTime Healthcare has built the industry's most comprehensive product matrix, accelerating its overseas expansion with Chinese solutions. The large and small robots achieved the industry's first end-to-end autonomous spatial intelligence, realizing one brain with multiple forms, marking a tangible breakthrough in embodied intelligence.
As the world's first mass-produced robot to implement the open cloud ecosystem, Yuan Turnip has truly entered millions of households. In addition, products like endpoint chips and smart driving have completed independent financing and started independent operations, which not only stimulated the operational vitality of ecosystem enterprises but also allowed the group to fully capture the growth dividends of AI integration with the real economy through efficient collaboration.
Next, I will invite our CFO, Wang Zheng, to report on behalf of the finance department.
Wang Zheng
Hello everyone, I am very pleased to share with you the full-year performance of 2025. Under the guidance of SenseTime's 'One Plus X' strategy, we have focused on high-quality development of core businesses. During this reporting period, the company achieved accelerated growth in both revenue and gross profit. Through meticulous management and efficient cost control, we realized a simultaneous improvement in scale and efficiency.
This has significantly narrowed the company’s losses at both the EBITDA and net profit levels, with the rate of narrowing notably accelerating compared to 2024. All major financial indicators for the year exceeded the median analyst expectations.
SenseTime recorded 32.9% revenue growth in 2025, with generative AI revenue reaching RMB 3.6 billion in 2025, increasing by 51% from 2024. The proportion of generative AI revenue to total group revenue rose from 64% in 2024 to 72% in 2025.
It is fair to say that our focus on the transformation towards generative artificial intelligence has been quite successful. Supported by the strong performance of this business segment, the company has achieved continuous high revenue growth, reaching new historical highs in scale.
Visual AI business revenue steadily rebounded in 2025, growing by 3.4% annually. Over the past two years, the company has actively optimized the business structure of this segment, focusing on improving cash flow levels and revenue quality.
Excluding the impact of a single client in Northeast Asia, the overall visual AI business achieved an annual growth rate of 33.5% in the second half of 2025. As a leading player in the industry, we are confident that this business will maintain healthy growth in both domestic and international markets and fully unlock its profitability potential.
The gross margin for 2025 was 40.0%, slightly lower compared to 41.0% in 2024 and 42.9% in the previous year, but still better than the market guidance range provided during our last earnings report.
The increase in gross profit and the continued strong control of operating costs have led to a significant narrowing of the group's losses at both the EBITDA and net profit levels compared to 2024, with the narrowing reaching 85.0% and 58.6%, respectively.
It is worth noting that the group achieved its first positive EBITDA in the second half of 2025. Even for the few profit-related indicators that have not yet turned positive, such as net profit, adjusted net profit, and EBITDA, their total losses in the second half of 2025 narrowed significantly more than in the first half.
This series of progress has laid a solid foundation for the company's subsequent high-quality sustainable development. In terms of operating expenses, the company has continued to maintain strict and efficient overall control. Coupled with the successful financing of some innovative businesses, all three expenses reported during the period have achieved year-on-year decreases, with sales expenses having consecutively achieved double-digit annual reductions over several reporting periods.
Administrative expenses also saw a double-digit decline of 16.3% in 2025, primarily due to the continuous improvement in personnel efficiency and asset utilization efficiency, as well as the optimization of comprehensive costs related to workplace management.
The company has always placed great emphasis on R&D investment to continuously strengthen its core competitiveness within the industry. In 2025, we increased our computational investment related to R&D. However, driven by the rapid improvement in R&D personnel efficiency and the spin-off of certain innovative businesses, overall R&D expenses decreased by 8.6% year-on-year.
Overall, the company’s total operating expenses in 2025 decreased by 10.9% year-on-year and were lower than the absolute value of total operating expenses in 2023. Total operating expenses in the second half of 2025 also decreased by 13.6% sequentially from the first half and dropped by as much as 21.6% compared to the same period in 2024.
Given the significant increase in revenue and gross profit scale, it is no small feat that we have been able to control various operating costs so effectively, which also demonstrates the enormous potential for SenseTime’s long-term profitability.
Our overall operational capital flow efficiency improved significantly in 2025. On the left are the days turnover of key working capital metrics calculated using end-of-period values. Using this measure, it can be seen that the cash conversion cycle shortened by 43% at the end of 2025 compared to the end of 2024, or by 99 days.
The main driving factor here is still the trade receivables days, which saw a significant decrease of nearly 50 days over the past year. We have continued to focus on collecting accounts receivable, and as the company rapidly transitions towards deeper urban AI business, the quality of our revenue has also been improving. This is reflected in the fact that our total trade receivables collection reached a record high of 4.87 billion in 2025.
In the income statement, the net amount of impairment losses on financial assets in 2025 also fell sharply compared to 2024, from 780 million to 290 million, accounting for only 5.7% of revenue, far below the 20.7% revenue share in 2024.
After our capital expenditures contracted year-over-year for two consecutive years in 2023 and 2024, growth resumed in 2025, reaching 3.49 billion RMB. The capital expenditure was primarily focused on the construction of large-scale computing power infrastructure.
This accelerated investment is based on our assessment of explosive growth in the artificial intelligence industry, including related applications and computing power, as well as the continuous enhancement of the company's integrated capabilities, giving us greater confidence. These investments have laid a solid foundation for the company’s long-term healthy development.
We will also remain flexible, continuously balancing the strengths of both asset-heavy and asset-light operating models. By the end of 2025, the company's total cash reserves increased to 14.2 billion RMB. This cash definition includes structured deposits but does not include the 7.5 billion RMB balance of equity and bond investments shown on the right.
In addition to nurturing multiple successful X innovative businesses internally, SenseTime has also built a diversified ecosystem matrix externally over the years. The fair value of these investments reached nearly 6.2 billion RMB by the end of 2025, steadily increasing as invested companies go public and their valuations rise.
It is worth emphasizing that the company’s self-sustaining capabilities have significantly improved, with both the operating cash flow turning positive in the second half of 2025, making us one of the few AI companies committed to sustainable and healthy growth.
Additionally, as of the end of 2025, the company’s unused bank credit lines amounted to 10.6 billion RMB, a year-on-year increase of 101%. Meanwhile, our overall debt level slightly decreased by 3.3% by the end of 2025.
Overall, our financial position is robust, with ample funding reserves providing full support for long-term strategic planning and business development. That concludes the financial chapter; I'll now hand the time back to Xu Li. Thank you all.
Xu Li
Looking ahead to 2026, we see AI entering a phase of scaled explosive growth. SenseTime is well-prepared, with clear and defined points of focus moving forward. First, we will continue to adhere to the path of native multimodality, establishing new paradigms of high-definition AI through technology and setting new standards of intelligence, solidifying SenseTime's leading position at the forefront of AI.
Second, we will deepen our expertise in industrial intelligence, creating intersections in the agent-native era. I believe AI will evolve from an auxiliary tool into real productivity. In vertical sectors such as general office, education, marketing, and smart terminals, we aim to seize entry points to achieve dual explosions in traffic and commercial value.
Third, we will push the unit cost of intelligence to its extreme. Leveraging the synergy between computing power, models, and applications, we will continuously reduce inference costs, making unparalleled cost-performance a hard barrier for SenseTime while fully promoting domestic computing power from being merely usable to becoming highly efficient.
Fourth, to fully capitalize on the scaling dividends of visual AI, we need to grasp the profitability code of the CV 2.0 era, drive business expansion and scale-driven profits while maintaining a dual focus on domestic and international growth. Our goal is to establish our commercial solutions as the global benchmark in the AI industry.
Finally, we will continue to deliver on the 'One plus X' ecosystem dividend by further optimizing the collaboration model between our main platform and sub-platforms. Together with our ecosystem partners, we aim to seize the value-added opportunities brought by the vertical industry boom. By 2026, we will harness even stronger original capabilities to fulfill our commitments in the AGI era. Thank you all.
Host
Thank you, management, for sharing. Now let's move on to the Q&A session. Friends, please raise your hands to ask questions on the Skill Finance platform. The first question comes from Mr. Yang of CITIC Securities. Mr. Yang, please unmute yourself.
Mr. Yang
Thank you to the SenseTime management for giving me the opportunity to ask a question. I’d like to seek advice on one point: As we look at SenseTime, we are very much looking forward to the release of new models. My question to management is, how should we understand the technological and commercial relationships among large language models, multimodal models, and agents? Why is SenseTime so committed to advancing native multimodal technology?
Different manufacturers clearly have varying value propositions and technological approaches regarding this path. So, my follow-up question is, how much does this help us in breaking through the upper limits of AI intelligence?
If possible, could management also elaborate on the core competitive barriers of our neo architecture? Additionally, we previously mentioned achieving relatively strong performance even with smaller datasets—how does this specifically benefit us? This would help us better understand our overall business strategy, technical barriers, and the key competitive advantages that underpin them. Thank you.
Xu Li
Thank you, Mr. Yang. First, I believe that SenseTime fundamentally remains committed to our native multimodal approach because it represents a significant breakthrough in pushing the upper limits of intelligence and can drive breakthroughs in many AI applications. Of course, after open-sourcing our neo architecture at the end of last year, the next version will be released in Q2 this year. Allow me to provide an overview of this progress.
Spokesperson
Thank you for Mr. Yang’s question. To address the first question, there are two important reasons why we are deeply focused on multimodality. First, in terms of long-term development, pure language data has already peaked. The deep integration of language and vision will become a critical pathway for breaking through the upper limits of AI capabilities.
And as we demonstrated in the report, many complex business scenarios in this era of AI actually involve the comprehensive processing of multimodal information. Equipping intelligent agents with visual capabilities and deeply integrating these with overall language processing can significantly increase the success rate and efficiency of completing complex tasks while substantially reducing token consumption.
Returning to our Leo architecture, it is designed to break down the barriers of the mainstream splicing paradigm. At its core, fundamental multimodal-native modifications have been made to its embedding, position encoding, and attention mechanisms, achieving kernel-level unification of recognition and language. This greatly enhances learning and reasoning efficiency, allowing us to achieve search performance with only one-tenth of the data volume.
Through further enhancements and upgrades to the model's architecture at a more foundational level, its efficiency has increased two to three times in the new generation of Leo. Such extreme efficiency has built a very deep technological moat for our overall operations.
Finally, I would like to emphasize that by unifying visual and language expressions, we have created an excellent space that supports native mixed-media thinking, integrating logical reasoning and spatial intelligence. This allows for deeper exploration across broad business scenarios. The model we are about to launch is something everyone can look forward to.
Host
Alright, thank you for sharing. Okay, let's move on to the next question. Thank you, thank you. The next question comes from Mr. Wang of Central Company. Mr. Wang, you may unmute yourself.
Mr. Wang
Hello to all management. Thank you very much for giving me the opportunity to ask a question. First, I would like to congratulate the company on achieving such impressive operating results. We see that the company has continuously reduced losses, including successfully breaking even at the EBITDA level in the second half of the year. In our view, this is indeed a crucial milestone breakthrough in the company's operations.
I have several questions to ask management, which can be divided into two parts. The first part focuses on financials and overall operations. From a financial perspective, could management help us analyze the key quantifiable factors driving this round of continuous loss reduction and achieving breakeven in the second half?
Additionally, on the investment side, the company has always placed significant emphasis on R&D investment. In this process of pushing for loss reduction or optimizing the business structure, how do we balance the goals of R&D investment and loss reduction? And regarding the current trend of improved profitability, how should we view its sustainability in the coming year or in the longer term?
For the second part, I would like to specifically ask about the generative AI business segment, which maintained very high growth throughout the year. What is driving this growth—project completions, API calls, or subscription-based revenue? For these potentially different business models, does the company have any outlook on their respective gross margin trends moving forward?
Wang Zheng
Thank you, Mr. Wang, for this series of questions. I think I've noted them down and will answer them one by one. Your first question is about how to quantitatively explain our breakeven at the group level.
This situation is actually the result of many factors. For instance, driven by the AI sector in Shenzhen, we achieved accelerated growth in revenue and gross profit, which was a key condition. At the same time, we rigorously implemented cost control strategies. Everyone may recall that in the three expense categories of operating costs in 2025, we saw year-over-year declines, with the total annual decrease in OPEX reaching 11%.
So at the overall group level, we have taken many specific measures to reduce costs and improve efficiency. Just to give a few examples: the close integration of our large-scale installations and models has generated ultra-high cost-effectiveness. Additionally, we established talent hubs in places like Wuhan, which further helped us reduce labor costs.
Moreover, we continuously optimized rental costs for our domestic and international office spaces. Overall, we've been managing costs meticulously from all angles. In addition, our revenue structure has been constantly optimized. The provision for accounts receivable, which negatively impacts the income statement, was significantly reduced — from a negative impact of 780 million in 2024 to only 290 million in 2025, which is also a major contributing factor.
Lastly, the deepening of the One Plus S strategy led to some S-ecosystem companies successfully securing external financing and being removed from our consolidated financials, creating additional room for overall group profitability improvement. Therefore, it's the result of a combination of these factors.
Your second question is about how to balance R&D investment and loss reduction. We have always regarded R&D as a core competitive advantage of the company, right? In 2025, we further increased our investment in computing power for R&D. However, due to significant improvements in R&D efficiency and the removal of certain X businesses driven strongly by R&D, the total R&D costs actually decreased in 2025.
Therefore, we believe that improving efficiency and strengthening R&D outcomes are not contradictory. Regarding your third point, you asked whether the current profitability trend will have sufficient continuity in the future, right? I believe there is definitely continuity and great potential.
Overall, we expect that in 2026, at the adjusted net profit level, we will continue to see a significant trend of loss reduction. Moreover, for the full year, there is a high probability that the adjusted figure will turn positive.
Finally, you asked about the main drivers behind the high growth rate in generative sectors, whether it’s project-based or positioning-based. As of 2025, the main growth drivers were primarily cloud services related to public clouds, along with AI model services that integrate computing power into various vertical scenarios.
So these are all very high-quality revenue streams, right? In terms of private deployment, it is relatively close to what you mentioned as project-driven. By 2025, its share of our total generative AI revenue will have significantly decreased. Regarding the trend in gross margin, we've always believed that 2024 and 2025 will be relatively stable. In fact, the overall group's gross margin ultimately exceeded the guidance range we provided at the interim report. That’s all I’ll say for now. Thank you.
Host
Alright, thank you for the management's response. Moving on, the next question comes from Mr. Xu of CICC. Mr. Xu, please unmute yourself.
Mr. Xu
Thank you very much for giving me the opportunity to ask a question. I mainly want to inquire about how the recent breakout or explosive popularity of open core impacts the application and implementation of AI agents for our company’s business. Additionally, regarding specific products like the Little Raccoon family, how might they assist or create synergies with agents similar to Open Cloud?
Spokesperson
Thank you for Mr. Xu’s question; it is very pertinent. Everyone has observed that since the Lunar New Year, the surge in token consumption driven by Open Pro has become a major hallmark indicating that the entire industry has entered the AI era.
Open Pro can be viewed from a technical perspective as an open-source agent operating system. Its ecosystem position can be likened to iOS. It reshapes how intelligent agents run continuously in the background without interruption and connects seamlessly with instant messaging tools, providing a highly user-friendly platform for token circulation.
However, we would like to emphasize that tokens are merely carriers of information and knowledge, while the true value lies in the final service outcomes received by users. Therefore, if Open Core is to provide sustainable high-value services to users, it must integrate with various professional capabilities such as deep analytical report generation, marketing video production, and other specialized skills.
These skills can be analogized to apps designed for silicon-based intelligence screens. The deep integration of this system with specialized capabilities delivers complete value to users, which aligns perfectly with SenseTime’s systematic advantages.
Our native multimodal capabilities at the technical level cover areas including text and image comprehension, webpage and document analysis, multimodal memory, and more. These can serve as core skills to empower the entire Agent OS, fully expanding the boundaries of AI applications to deliver high-value results to users.
Moreover, it is worth mentioning that the breakthrough in the new native architecture paradigm of Leo we just discussed is also expected to provide a more efficient multimodal engine for future agents and enable their capabilities to extend from online spaces to offline physical spaces.
Of course, looking back, this new paradigm of open cloud also inspires us to integrate transformations in its existing products and industry solutions, helping us create an even better user experience. So, this is actually a two-way advancement, and its impact on SenseTime is very positive.
Xu Li
Yes, just to add one point: if our unified model generation can break through the upper limit, I think it might be like equipping such systems or operating systems with eyes and a comprehensive understanding of vision. I believe this capability will unlock many downstream applications. Thank you.
Host
Let's move on to the next question. The next question comes from Yang Ming, General Manager of Guotai Haitong. Mr. Yang, please unmute your microphone.
Yang Ming
Alright, thank you, Jessie, and thanks to Mr. Xu and everyone else. I am Yang Lin from Guotai Haitong. I would like to ask the management team: at present, what level is our computing power utilization maintained at? And how do we balance the ratio between internal R&D and external leasing?
Since recently, there has been a trend of price increases among major cloud providers. Will SenseTime follow suit with a price adjustment? And where does the current cost advantage of SenseTime's computing clusters lie? These questions are mainly about computing power. Thank you.
Yang Fan
Thank you, Mr. Yang. First of all, over the past year, our operational computing power has continued to expand, supporting nearly a million model R&D tasks throughout the year. Therefore, the overall utilization rate has remained at a very high level. Meanwhile, as our business scale expands, the proportion of our external services is also gradually increasing. Looking at last year as a whole, the usage by external clients accounted for about 40%, and it shows a continuously rising trend.
Of course, from our perspective, we highly value that we have now become key suppliers for companies including JD.com, Xiaomi, and other leading enterprises in various industries. We are committed to providing them with multiple services over the long term. This is what we focus on more in this field.
Your second question is actually about the issue of cost and pricing. We understand this matter as follows: the AI cloud market has experienced significant fluctuations in recent years, including both its supply chain and market prices. What we focus on more is the certainty within these fluctuations.
Our certainty lies in SenseTime's philosophy, where we hope to gain premium returns in the market by providing additional value. For instance, leveraging our leading support capability for domestic computing power allows us to help certain clients address their key needs when using domestic solutions.
Alternatively, by using our highly efficient inference framework, we can reduce the usage costs for customers when they utilize tokens. Or, through integrated software and service solutions tailored to specific industries and scenarios, we help particular types of customers achieve critical value. Overall, we aim to provide customers with the most considerate service content and pricing rather than simply acquiring and retaining customers through price alone.
Host
Alright, thank you. Let’s move on to the next question. The next question comes from Mr. Guo of CICC. Mr. Guo, please unmute yourself.
Mr. Guo
Okay, everyone, I have a question. Since computer vision has been our traditional advantage business, from a long-term strategic perspective, how does computer vision contribute to the group's large model strategy? And how should we understand the synergistic effects in relation to our traditional businesses?
The second question is regarding our large-scale infrastructure. It seems that it can now achieve the seamless integration of ten thousand domestic computing cards. I believe that going forward, particularly for industry clients, if AI applications develop rapidly, the demand for domestic computing power will be substantial.
I’m not sure about the current situation regarding client types using our domestic computing platform and how widely it is being applied. These are my two questions. Thank you.
Xu Li
Regarding what Mr. Wang mentioned about computer vision, from an industry position perspective, our traditional computer vision has always been at the forefront of the industry for nine or even ten years. This prolonged market leadership gradually leads to fewer competitors, which means higher profits. Thus, we see a positive iteration within this industry.
Now, speaking of CV and large models, I think there are several ways in which CV contributes to our commercialization efforts. First, CV helps us quickly capture market share. The customer base generally overlaps with that of large models since all CV customers have aspirations to upgrade to large models. In this process, our delivery capabilities and commercialization skills actually help us expand further.
We just looked at several B2B tracks, which often stem from our early customer accumulation and then quickly expand coverage. Later on, during the internationalization process, the progress of computer vision (CV) in the international market has been rapid. Large models have had slower progress due to overseas models including those from the US, but we can use CV to create a deep connection and provide comprehensive services.
So in this regard, it is actually a huge help for customer acquisition. On the technical side, as we mentioned earlier, we firmly believe that overall, visual capabilities can enhance the ceiling of text-based models. Therefore, CD also supports our understanding and generation within a unified framework as a core element.
Only by understanding visual content can we better construct the core training framework and model architecture. Similarly, native multimodal systems can be better applied to various agent systems we discussed earlier, such as core systems for dispatching various CV skills. I believe this synergy will exist long-term and be highly effective, forming a core competitiveness for SenseTime’s strategic direction.
Yang Fan
Your second question was about Wan Card Easy Purchase, and the types of clients Mixun targets. Thank you very much, Mr. Zhu. What we are seeing now is that initially, its market came from government and enterprise sectors, including some large state-owned enterprises and many domestic research institutions, where there were clear demands for domestic solutions, and the scale has been continuously growing over the past two years.
In recent months, there has been a clear new trend: more and more internet companies and early-stage tech startups are adopting a more open attitude towards domestic chips. We can observe this entire process accelerating. From this perspective, it's positive, allowing SenseTime to leverage the collaboration with domestic hardware manufacturers over the years, including exploratory work and accumulated knowledge and technologies. We hope to align well with the current domestic acceleration market trend and seize the opportunity effectively. Thank you.
Mr. Guo
Understood. Let me follow up with one more question. Is the main demand for our current domestic heterogeneous computing platform still primarily driven by training? Are we seeing any signs of this trend gaining momentum?
Yang Fan
Yes, in fact, my understanding is that the usage of domestic cards for inference tasks has become very widespread. In terms of cost-effectiveness, including if we consider our inference framework across video, SQL, and other areas, everything is completed using domestic solutions, even our world model relies on domestic inference. However, looking back step by step, it still mostly focuses on training rather than inference.
As for the direction, I think the localization of inference usage has already been put on a relatively extensive agenda, as Yang Fan mentioned earlier. This includes some in-depth research institutions and early clients. So, we see that most applications in our industry can actually use localized solutions for inference.
Xu Li
Let me add one point: there are indeed some trends within the industry. When performing large-scale inference, information separation sometimes occurs, right? Now, there is often integration of multi-tasking, so different parts of the inference process may involve different chips to achieve an optimal cost-performance balance. This area is actually a significant growth point currently.
Mr. Guo
Thank you, understood. Thank you, leaders. My question is as above.
Host
Alright, let's move to the next question from Zhou Yuan of GF Securities. Mr. Zhou, you may unmute yourself.
Zhou Yuan
Thank you, management, for giving me the opportunity to ask a question. I would like to inquire about the accounts receivable situation. On the 2025 financial statements, I noticed that the company has significantly improved its collection efficiency. However, the market still has concerns about the quality of past accounts receivable. Has the bad debt risk from previous years been largely resolved?
As the group shifts its focus towards deeper AI-driven businesses, the group's revenue has reached a new historical high. How do you view the quality of the company’s future accounts receivable? What measures have been taken around improving operating cash flow? Thank you.
Wang Zheng
Alright, let me answer Mr. Zhou’s question. Thank you for your inquiry. Regarding the bad debt risk, as you mentioned, it has indeed decreased significantly over the past few years. You can observe the trend of our aging accounts receivable for more clarity.
Thus, we can see that the proportion of AR outstanding for more than two years has decreased from approximately 80% at the end of 2024 to less than 60% by the end of 2025. Meanwhile, the share of shorter-term AR outstanding within 12 months increased from only 16% at the end of 2024 to 37% by the end of 2025.
From these two trends, it's clear the direction of optimization is very evident. On the other hand, historically, our balance sheet-level bad debt provision for AR has been relatively high, reaching 66% by the end of 2024 and dropping to 57% by the end of 2025. This also implies that future bad debt losses at the PML level should be relatively limited.
Additionally, as you mentioned, the bulk of our revenue comes from premium water, where the average revenue quality is generally higher. Another point is that, as you mentioned at the end, improvements in operating cash flow were noted. We shouldn't forget that our turnover days have improved from both inventory and accounts payable perspectives, which also helps the group further enhance its operating cash flow. By the second half of 2025, we generated positive operating cash flow for the first time. Thank you.
Host
Alright, thank you. Let’s move on to the next question. The next question is from Helen Fang of HSBC. Ms. Fang, please unmute yourself.
Helen Fang
Hello, thank you very much for giving me the opportunity to ask a question. I'd like to quickly inquire about the computing power issue because the market is currently very focused on the actual usability of domestic computing power. I would like to ask management, what is the current percentage of domestic chips used in SenseTime's own operations and external services?
Looking ahead to 2026, what are our plans for expanding computing power? What will be the ratio between self-built computing power and outsourced operations, and how do we ensure that our domestic computing power remains competitive with international benchmarks in terms of cost-effectiveness? Thank you.
Yang Fan
Thank you, Ms. Fang. Actually, this question overlaps somewhat with Mr. Guo's earlier question. First, regarding the feasibility of domestic computing power, significant progress has been made over the past few years. Much of the work we discussed last year, including adapting to domestic computing power like Model 10C super nodes and deploying robotics solutions, was based on specific commercial customer needs. Moreover, among different customers, their demands now cover a wide range of training scenarios, which reflects a definitive trend we're observing.
Secondly, the integration scale of domestic chips is increasing significantly. Previously, there might have been hundreds of petaflops (P), but now we are seeing more individual clients requiring thousands or even several thousand petaflops. Therefore, these two trends reflect well on the evolution and improvement of domestic capabilities.
As for SenseTime itself, our total computing power scale is now 40,400 petaflops (P), with domestic computing power exceeding 5,500 P, roughly in that proportion. We see that the increase in its share is a definitive trend. Moreover, SenseTime will leverage our technical accumulation in this field to continuously collaborate with domestic manufacturers on software co-optimization, thereby enhancing its cost-performance ratio and expanding the range of applicable models so that an increasing number of customers can accept and embrace the domestic system.
Regarding your second question about self-construction versus outsourced operation, we are still adhering to this dual-driven approach, meaning two legs walking simultaneously. Based on market conditions and rhythm needs, both legs will develop in sync. Additionally, in 2026, we will focus on increasing our planning and promotion of domestic computing power.
We hope that by exporting highly adaptable software solutions within an environment, we can help customers build end-to-end cost-performance advantages while providing domestic services that offer high cost-effectiveness, high reliability, and wide applicability.
Host
Alright, I see there are more questions from Mr. Liu. Let's move on to the next question. The next question comes from CLSA’s Chris Will Yuzon. Mr. Liu Xiaoming, you may unmute yourself.
Liu Xiaoming
Thank you to management for giving me the opportunity to ask a question. Also, congratulations to the company for achieving impressive results in the past five years. I would like to ask about the progress regarding large model exports. We see 2026 as a critical year for large model exports, and the company has launched the first domestic computing power cluster in Saudi Arabia.
So, I would like to ask management, in light of the complex geopolitical environment in the Middle East, how does the company plan to manage and control these potential risks? Additionally, is our export model currently inclined towards the output of lightweight asset software solutions or towards heavy asset infrastructure investment?
What is the target revenue contribution of overseas business to our total income over the next few years? These are my questions. Thank you.
Xu Li
Thank you, Mr. Liu. We actually have a rather unique positioning. AI B2B internationalization has always been our steadfast strategy. For overseas enterprise services, service quality and cost are particularly crucial. This is why we have worked with partners overseas to establish our domestic Tier 3 computing cluster, which holds significant importance. Overall, we adopt a lightweight asset approach but also use a combination of light and heavy assets for development.
Such investments carry two important strategic values. First, they promote the influence of China’s domestic ecosystem. Since we provide services here, if we can achieve joint optimization with the domestic ecosystem, we can use this co-optimization to reduce costs.
The second point is to establish long-term customer service stickiness, as it would be very difficult to achieve this level of customer retention without such infrastructure. Therefore, we also anticipate that overseas business will account for a larger share of total revenue in the future. Currently, our operations in the Middle East appear relatively stable and have not been affected by the temporary economic environment. Thank you.
Host
The next question comes from Amen Ho at DBS. You may unmute yourself, Amen.
Amen
Thank you, management, for giving me the opportunity to ask a question. I would like to inquire about SenseTime's thoughts and strategic positioning in the field of advanced artificial intelligence and physical AI. I understand that Sizeable Robotics is an enterprise incubated by SenseTime. Could you explain how its technology uses environmental data to enhance event modeling?
Additionally, regarding commercialization within high-value service scenarios, I would like to ask about the latest progress of Sizeable Robotics' implementation. Do you have guidance on whether there will be large-scale deployment by 2026? Thank you.
Wang Xiaogang
This is Wang Xiaogang, and I'll answer this question. Sizeable Robotics represents SenseTime's key strategic move from large models into the era of advanced AI, achieving the connection between the digital and physical worlds.
In March 2025, we were the first in the industry to introduce environmental data collection methods. This approach does not rely on operating real machines to gather data but instead records human behaviors during real-life production and living activities. As a result, the volume of data for advanced AI has rapidly grown from hundreds of thousands of hours to tens of millions of hours.
In December 2025, Sizeable Robotics unveiled the first combination of environmental data with world models, establishing a new paradigm for advanced AI research and development, leading us closer to what could be considered the 'ChatGPT moment' for advanced AI.
Therefore, in February 2026, we open-sourced the Enlightenment Advanced World Model 3.0. This model leads across all benchmarks for advanced AI, running 72 times faster than NVIDIA’s Cosmos 2.5. It also marks the first time in the industry that real-time driving and deployment on robot endpoints has been achieved, thanks to our real-machine data collection solution.
Sizeable Robotics has now formed strategic partnerships with multiple local governments to jointly build environmental data collection bases. The company’s autonomous navigation-enabled robotic dogs are being deployed at scale in various sectors, including smart cities and cultural tourism. Thank you.
Host
Due to time constraints, we will move on to the last question. The following question will be posed by Mr. Lu from Bocom International. Mr. Lu, please unmute your microphone.
General Lu
Thank you, management, for accepting my question. My question mainly concerns SenseTime's 'One Plus X' strategy. Under the current strategy, what do you consider to be the most core assets of the group, and what are the sources of long-term value?
Also, after these subsidiaries operate independently, what mechanisms will the company use to ensure that they remain high-quality contributors of data back to our large model capabilities, as well as key consumers of core computing power?
Additionally, a follow-up question: In SenseTime’s view, how does the current collaborative cluster model, which combines the parent company with its subsidiaries, compare to the delivery models of traditional standalone software companies? What is its main significance or differentiation? Thank you, that is my question.
Xu Li
Thank you. Actually, since the end of 2024, we have been implementing the 'One Plus X' strategy based on several considerations. First, AI development moves very quickly. Some highly complex, low fault-tolerance scenarios—like those mentioned earlier, such as AI-powered robotics—require more time. However, these areas hold immense strategic value, even though their commercial cycles may be longer than those in the intelligent enterprise application space.
As a company that has developed over ten years, the core value of both parties lies in continuous iteration between underlying AI model innovation and business closed-loop refinement. This requires two-way iteration to solidify product and service capabilities. Therefore, our 'One Plus X' strategy focuses on incubating industries with strategic value and future orientation into self-built operations to occupy industry high ground.
I believe that the servitization of AI will definitely be a crucial trend in the future. Our core competency at SenseTime lies in connecting and integrating models with services, ultimately empowering various intelligent applications such as smart sports through our golden models.
At the same time, regarding the construction of foundational models, infrastructure, and the skills previously mentioned, I believe that such national-level collaboration avoids reinventing the wheel and ensures a stable framework with better synergy.
This design also has a very positive impact on our financial reports, allowing us to focus more effectively and adapt our products and industry positioning more rapidly to changes in the times. Lastly, thank you, Mr. Lu, for your question. Thank you.
Host
Thank you once again for joining SenseTime's earnings release conference today. We look forward to the next release. Goodbye, good night, and thank you all.
More details:SENSETIME-W IR
Note: The above content is generated by an AI language model based on publicly available data and third-party automatic subtitles. The content does not represent any position of Futu, nor does it constitute any investment advice. Futu Group makes no express or implied guarantees or statements regarding the accuracy, timeliness, or completeness of the above content.
Risk Disclaimer: The above content only represents the author's view. It does not represent any position or investment advice of Futu. Futu makes no representation or warranty.Read more
Comments (2)
to post a comment
38
1