SENSETIME

was live · ·

2026 Interim Results Presentation

[AI Key Takeaways]

Financial Performance

- The Group achieved revenue of RMB 2.91 billion, a 23.4% year-on-year increase

- Generative AI revenue reached RMB 2.33 billion, up 28.2% year-on-year, accounting for nearly 80% of the Group's total revenue

- Gross profit reached RMB 1.21 billion, a 32.9% year-on-year increase, with gross margin rising to 41.4%

- Achieved IFRS-based net profit of RMB 620 million for the first time

Business Progress

- Released the SenseNova U1.5 model, achieving seamless integration of understanding, reasoning, generation, and editing capabilities

- Total operational computing power reached 48,000 GPUs, with average daily token consumption hitting 2.4 trillion

- Total users of XiaoHuanXiong grew more than fivefold year-on-year, while enterprise user growth exceeded sixfold

- Ranked No. 1 in China's computer vision market for ten consecutive years

Next Quarter Guidance

- Expected to achieve profitability at the adjusted EBITDA level for the full year 2026

- The proportion of recurring revenue to total revenue is expected to increase significantly in the second half compared to the first half

- Capital expenditure growth rate will exceed revenue growth rate in the current period

Opportunity

- Overseas business revenue increased by 127% year-on-year, with computing centers established in Hong Kong and Saudi Arabia

- Driving the continuous evolution of multimodal long-horizon agents in real-world tasks

- Reduced unit electricity costs by approximately 8% through computing-power and electricity coordination, achieving an annual carbon emission reduction of 24,000 tons.

- Established strategic partnerships with Kylin Software and others to expand market coverage.

[AI Conference Transcript]

Phil, Head of Capital Markets at SenseTime Group

First, let me introduce the management representatives attending this press conference: Dr. Xu Li, Chairman and CEO of SenseTime Group, hello everyone. Dr. Wang Xiaogang, Executive Vice President; Mr. Yang Fan, Executive Vice President and President of the Large Device Business Group; Dr. Lin Dahua, Executive Vice President and Chief Scientist, hello everyone. And Mr. Wang Zheng, Executive Vice President and Chief Financial Officer.

Before we begin, I will read the disclaimer. Today's discussion may contain forward-looking statements, which involve inherent risks and uncertainties that could cause actual results to differ from current expectations. For detailed information on these risks and uncertainties, please refer to the latest announcements filed by SenseTime Group with the Hong Kong Stock Exchange. Unless required by applicable law, SenseTime Group assumes no obligation to update any forward-looking statements.

Today's discussion also involves certain non-IFRS financial measures, primarily intended to assist in comparative analysis. Management will deliver their prepared remarks in Chinese, with simultaneous interpretation provided on the English conference line. In case of any discrepancy between the original speech and the translation, the Chinese original by management shall prevail.

Next, we invite Dr. Xu Li and other management representatives to speak separately, introducing SenseTime Group's operational and development status for the first half of 2026. Finally, we will invite the management team to participate in a Q&A session. Let us first welcome Dr. Xu Li.

Dr. Xu Li

Distinguished guests, dear investors, hello everyone. Thank you for attending SenseTime Group's 2026 interim results press conference. In the first half of this year, we have observed the AI industry accelerating its transition from model intelligence to action intelligence. Previously, the focus was largely on whether models could answer questions or generate content. Now, competition is shifting towards whether AI can understand and decompose complex objectives, invoke tools, coordinate multiple agents to continuously execute hundreds of steps, and ultimately deliver usable and verifiable results.

Against this trend, we delivered a solid performance in the first half of the year. The Group achieved revenue of RMB 2.91 billion, a year-on-year increase of 23.4%. Among this, generative AI revenue reached RMB 2.33 billion, up 28.2% year-on-year, accounting for nearly 80% of the Group's total revenue. Visual AI revenue was nearly RMB 500 million, representing a 13.9% year-on-year increase. Overseas business revenue grew by 127% year-on-year, becoming a significant growth driver.

More importantly, the quality of growth continues to improve. In the first half of the year, our gross profit reached RMB 1.21 billion, a year-on-year increase of 32.9%, with the gross margin rising to 41.4%. We have introduced a new disclosure metric: recurring revenue, which reflects income from active contracts with continuous renewal attributes. This amounted to RMB 1.14 billion, up 124.4% year-on-year, accounting for 39.3% of the Group's total revenue.

This indicates that our customer relationships are shifting from project delivery and product sales toward a model based on the continuous consumption of tokens and agent services. Our operational efficiency has also improved significantly; the adjusted net loss for the first half narrowed to RMB 390 million, a 67.3% reduction year-on-year. Meanwhile, the Group achieved an IFRS-net profit of RMB 620 million, marking SenseTime's first IFRS-profitable period since its listing.

These figures reflect a single underlying trend: SenseTime's technical infrastructure and products are forming a more complete industrial closed loop. Incidentally, all charts in our presentation were generated by our new 1.5 model. This diagram summarizes SenseTime's current core capability framework. We are committed to building a system comprising three key components: a unified model, a token factory, and an agent control system.

Previously, we emphasized the 'trinity' approach, integrating infrastructure, models, and applications. The transformation we are undertaking now highlights our core competitiveness. Within this 'three-in-one' framework, the unified model is responsible for continuously raising the ceiling of intelligence. Built on the native multimodal unified architecture of SenseNova, it progressively integrates understanding, generation, reasoning, and execution capabilities, enabling the model to process not only text but also images, video, spatial data, and various application interfaces.

The token factory is responsible for scaling up the production of intelligence. It goes beyond merely providing computing power; through joint optimization of models and infrastructure, it connects different chips, clusters, energy sources, and delivery nodes to continuously produce higher-quality, lower-cost tokens. The agent control system, or 'agent harness,' converts models and tokens into complete tasks. It integrates knowledge, tools, workflows, identity permissions, memory, evaluation, and security governance to support various front-end agent products.

Our long-term view is that token efficiency determines the cost of intelligence, while final task delivery determines customer value. Therefore, the AI business model will ultimately evolve beyond simply selling models or tokens, moving increasingly toward measurement based on tasks completed and results delivered. Next, we invite Dr. Lin Dahua to introduce progress in model and agent R&D.

Dr. Lin Dahua

I will now share SenseTime's thinking and latest developments in model R&D. Our technical roadmap has been consistent and represents a continuous process. We firmly believe that multimodality is an essential path toward achieving more comprehensive AGI, and as intelligence levels iterate and improve, the boundaries of multimodal capabilities continue to expand.

In the early first stage, combining image perception with language models achieved basic multimodal understanding. Currently, in the second stage, we have made significant progress. Through effective combined training of native multimodal capabilities and agent capabilities, we have developed multimodal long-horizon agents capable of efficiently processing mixed-modal information. While maintaining the competitiveness of foundational agent capabilities, we have significantly expanded their application boundaries through enhanced multimodal abilities.

Looking ahead, we will continue to drive multimodal AI capabilities into the physical world, enabling AI to understand environments, predict changes, and drive actions in complex physical spaces. These are not separate paths, but rather continuous expansions of capability within the same architecture. Multimodality allows models to see more comprehensively, long-horizon agent capabilities enable models to act more deeply and completely, while spatial intelligence and physical feedback allow models to begin understanding and acting within the real world.

Early this year, building on the first-generation Leo architecture, we further launched the Leo Unify architecture, which features native unified multimodal understanding. This architecture represents a significant innovation by abandoning the traditional separation of encoders and decoders, effectively processing both images and text within a unified representation space. This shifts multimodal capabilities from mere functional stitching to true native integration.

Based on this new architecture, we released SenseNova U1 in April, achieving unified understanding, reasoning, and generation within a single model for the first time. The recently released U1.5 further integrates understanding, reasoning, generation, and editing, fully validating the strong scalability of this architectural framework.

The community influence and open-source ecosystem of the SenseNova series are rapidly developing. Within less than four months of its open-source release, the U1 series, along with the related SenseNova Vision and SenseNova Skill projects, has accumulated over 11,000 stars on GitHub. The unified model architecture has demonstrated significant effectiveness in integrating visual perception capabilities, enhancing visual reasoning, and enabling controllable generation.

In terms of visual perception and spatial understanding, the unified model consolidates capabilities for visual understanding, geometric prediction, image segmentation, and 3D geometry. We have also achieved important breakthroughs in visual understanding; through effective mixed-image-text reasoning, U1.5 Lite scored 68.3 on the authoritative MMVP Pro visual reasoning benchmark.

Regarding controllable generation, unified understanding and generation reduce information loss caused by cross-modal conversion. Despite its lightweight size of only 8 billion parameters, U1.5 Lite can generate highly detailed native 4K images, accurately execute long-form creation instructions with multiple constraints, and precisely modify specified content while maintaining overall structural stability. This transitions image generation from one-off outputs to a controllable, iterative professional workflow.

Thanks to the advanced efficiency of the model architecture, U1.5 Lite achieves high performance with a small footprint, reaching levels comparable to Nano Banana Two in authoritative tests and delivering industry-leading cost-performance ratios. Building on the unified architecture, we have further integrated multimodality with long-horizon agents. Released in August, SenseNova 6.8 Flash Light can process documents, charts, web pages, videos, and application interfaces within a single task trajectory.

It can operate continuously for hours across hundreds of steps, orchestrating more than ten specialized agents to autonomously complete the entire process from information search and data analysis to report and presentation generation. In this process, multimodality is not an additional cost; instead, the effective combination of language and visual elements significantly enhances the efficiency of long-horizon tasks. In tasks involving information search and mixed-image-text data analysis, 6.8 Flash Light reduces token consumption by approximately 60% on average compared to pure-text LLMs. It has also achieved leading performance in various authoritative evaluations for enterprise office work and team collaboration.

In content creation, SenseNova also stands out. Our SenseNova U1 Pro connects requirement understanding, creative reasoning, image generation, and editing verification into a complete model workflow. Its significance lies not just in being a single-point generation model, but in enabling AI to deliver professional content outcomes. In the third-party Super Crew Image benchmark released today, U1 Pro took the top spot, surpassing GPT Image 2. On the same leaderboard, the lightweight open-source model U1.5 Lite also achieved impressive results, competing with several of the latest closed-source commercial models.

Additionally, we would like to preview that we will soon launch Flash, a long-horizon agent model with larger parameters. We believe it will deliver better performance in more complex tasks and multi-agent collaboration. Looking further ahead, our goal is to continuously explore the boundaries of physical intelligence. SenseTime's SenseNova SI 8B spatial intelligence model has already reached a leading level among open-source models across eight spatial intelligence benchmarks.

We have co-developed the Kaiwu 3.1 world model with ecosystem partners, including Daxiao Robotics. This model integrates understanding, generation, and prediction of the physical world using a unified architecture. It achieves an inference latency of 125 milliseconds at BF16 precision. Even with such low latency, it can adapt to different robotic bodies through heterogeneous multimodal compatibility, demonstrating strong scalability. This forward-looking exploration provides us with a crucial R&D closed loop: from understanding the world and simulating the future to driving actions and obtaining feedback, which is then used to iterate the model.

Next, we invite Yang Fan to introduce the token factory that underpins the R&D and scaled services of these models.

Mr. Yang Fan

Thank you. Dahua just explained how models continuously raise their intelligence ceiling. Now, I will discuss how SenseTime’s Large Computing Facility, as the industry’s leading native AI infrastructure service provider, produces this intelligence in a stable and efficient manner. With the continuous evolution of large model training, the rapid growth of inference demands, and agents truly entering real-world workflows, the value of AI infrastructure is no longer just about computing power capacity, but rather the ability to consistently, stably, and efficiently produce high-quality tokens.

Tokens are not homogeneous products; behind the same token, the model’s inference capability, accuracy, reliability, response speed, and cost can vary significantly. Therefore, SenseTime’s Large Computing Facility is proactively following industry trends, upgrading from efficient supply of computing resources to providing high-efficiency, low-cost, and high-quality token production. On one hand, we increase output and reduce costs through model adaptation, inference optimization, heterogeneous scheduling, and compute-network synergy. On the other hand, by enhancing our understanding of models and tasks, we improve the inference, comprehension, generation, and execution capabilities embedded in each token.

In the long run, the token factory will evolve from merely producing tokens to completing tasks, with efficiency metrics gradually extending from cost per token to quality, success rate, and cost per task. As part of our AI infrastructure segment, one of our key missions has always been to support SenseTime models in continuously pushing performance boundaries and rapidly transforming frontier models into stable services. Through continuous technological iteration and innovation, we ensure that SenseTime’s systematic support framework for model R&D and services remains at the industry’s leading edge.

For long-context training of agent models, we leverage advanced computing interconnects and memory characteristics to optimize sequence parallelism, topology-aware multi-dimensional parallel strategies, and compute-communication synergy, doubling the training speed compared to pre-optimization levels. In the post-training phase of reinforcement learning, we decouple training and sampling tasks, then utilize a mix of domestic and advanced computing power to form heterogeneous inference resources, accelerating sample generation and improving post-training efficiency.

In image and video generation, our self-developed LiteX2A inference framework combines deployment distillation with multi-level computational optimization, increasing generation efficiency sixfold. Additionally, relying on a unified R&D and deployment foundation, we have shortened the cycle for scaling model deployment to new types of domestic chip clusters to just one week, significantly enhancing the efficiency of bringing model technology iterations to market at scale.

While strengthening our own system capabilities, we place particular emphasis on collaboration between our self-developed models and upstream chip manufacturers. For instance, on the day SenseNova U1 was released, it achieved initial adaptation with over ten domestic chip manufacturers, ensuring that open-sourced models run immediately and are compatible with various domestic chips. These capabilities collectively shorten the distance from model innovation to commercialized services.

Meanwhile, we have persisted for years in driving the ecosystem construction and scaled commercial adoption of domestic computing power, moving it from mere operability to efficient commercial use. To achieve this goal, we developed a full-stack adaptation system covering models, frameworks, operators, toolchains, and hardware. Through low-level operator optimization, multi-card parallel tuning, separation of prefilling and decoding (PD separation), and heterogeneous inference, the Model FLOPS Utilization (MFU) of mainstream domestic chips can reach up to 2.5 times the baseline level set by original chip manufacturers.

The significance of this improvement lies not only in optimizing technical indicators but also in enhancing the cost-performance ratio of domestic intelligent hardware, safeguarding the pure market-driven profit margins of domestic chips, thereby supporting the endogenous scaled growth of the domestic chip service ecosystem. At the same time, we are replicating our accumulated construction and operational capabilities to more regions, including overseas. SenseTime collaborates with Hong Kong Science Park to build the largest domestic computing power center in Hong Kong, targeting a capacity of 40,000 cards by 2030. Additionally, we have launched our first overseas domestic computing power cluster in Saudi Arabia.

For us, this is not merely an expansion of computing power nodes, but a globalization of token production and delivery capabilities. Supported by robust performance guarantees, the scale of external services provided by our Token Factory is growing rapidly. As of today, our total operated computing power reaches 48,000 cards. As reported at last month's World Artificial Intelligence Conference, the daily average token service consumption of SenseTime’s large-scale AI infrastructure has reached 2.4 trillion, representing a year-on-year increase of approximately 22 times.

In addition to serving SenseTime’s Riri Xin large language model, our Token Factory also provides services to four external foundational model providers, covering highly complex scenarios such as AI for Science, video generation, embodied intelligence, world models, and city-level intelligence. Looking ahead, AI agents will bring longer reasoning chains, more frequent tool calls, and higher concurrency demands. Our goal is not simply to expand computing capacity, but to continuously increase the supply of high-quality tokens and steadily reduce the cost required to complete real-world tasks.

To achieve ultimate cost-effectiveness and represent future development trends, any Token Factory must drive down costs through deeper energy efficiency improvements and vigorously embrace green energy to ensure sustainable development. SenseTime’s large-scale AI infrastructure has integrated computing management, data center operations, and energy management systems, establishing a collaborative mechanism ranging from power load forecasting to computing task scheduling and O&M decision-making. Currently, through our computing-power-and-electricity coordination, the composite prediction accuracy for large models and agents has reached 96%.

The system can coordinate energy storage dispatch, server room maintenance, and computing task scheduling based on prediction results, continuously optimizing PUE (Power Usage Effectiveness) and electricity consumption structures. These measures collectively helped reduce our average unit electricity cost by approximately 8%. In the first half of this year, we saved over RMB 12 million in electricity costs, with annual carbon emissions reduction per 10,000 computing cards reaching 24,000 tons, marking a multi-fold increase compared to 2025 levels.

The value of computing-power-and-electricity coordination is also very direct: it not only reduces resource costs and carbon emissions but also provides a replicable operational methodology for large-scale Token Factories to continuously lower the unit cost of intelligence production. We will continue to promote pioneering exploration and commercial validation in related fields. Having introduced the models and the Token Factory, we still need a system to connect them to truly achieve task delivery.

What users see on the frontend are various products such as SenseTime’s Xiaohuanxiong office agent, Ruying marketing agent, and Cycle content creation agent. While they serve different users and scenarios, they are all supported by the same backend agent management system. This system is responsible for understanding user goals, decomposing tasks, invoking knowledge and tools, and orchestrating workflows. It also centrally manages identity permissions, long-term memory, execution evaluation, audit trails, and security governance.

Simply put, it connects models and the Token Factory on one end, and enterprise and individual real-world workflows on the other, translating what the model can do into what the system can reliably accomplish. Through the diversification of frontend agent products and the unification of backend capabilities, we can faster replicate the same task delivery capability across different products, industries, and user scenarios, continuously improving the quality, reliability, and delivery efficiency of task completion.

In the enterprise market, we have integrated unified agent capabilities into customer workflows through services such as public cloud APIs and enterprise-grade private deployments. In office scenarios, SenseTime’s Xiaohuanxiong already serves leading enterprises including Lenovo, Ping An Technology, the three major telecom operators, JD.com, and Kylin Software. For instance, through cooperation with Kylin Software, Xiaohuanxiong has been integrated into domestic operating systems, connecting OS entry points, AI application distribution, and enterprise delivery chains.

In JD.com’s supply chain business, we built an end-to-end intelligent service system around actual workflows through private deployment, validating our delivery capabilities in complex environments with high security requirements and critical business systems. Usage data shows that Xiaohuanxiong’s total user base grew more than fivefold year-on-year, enterprise users increased by over six times, and monthly active users on the cloud grew nearly tenfold. This growth stems not just from user experience and traffic, but largely from the product entering more real-world work scenarios.

In the content production sector, Cycle has advanced video creation from single-point generation to end-to-end delivery covering demand understanding, scriptwriting, and storyboard production. In the June Quantum AI Application Monthly Report, Cycle ranked first in average visit duration per user on the web. Currently, its daily video output reaches 10,000 minutes, with active users creating content for over an hour each day on average. The cumulative estimated views of Cycle-created content on short-video platforms have exceeded 1.5 billion, with ten short dramas individually surpassing 100 million views.

This indicates that Cycle is evolving from a standalone video generation tool into a specialized platform supporting continuous creation and scaled content production. Meanwhile, the beneficiaries of AI productivity are expanding from large organizations to smaller teams and OPCs (One-Person Companies). In the first half of the year, the Little Raccoon OPC Capability Challenge attracted over 650,000 participants, with daily active users peaking at more than 700,000. The first batch of the Cycle OPC Industrial Base has also attracted over twenty startup teams.

With the help of AI, an individual can orchestrate models, tools, and workflows to complete complex tasks that previously required professional team collaboration. This will become a significant new form of production and entrepreneurship in the AI era. In the individual user market, we have built the Capy product matrix centered around the concept of the 'Personal CXO.' As shown by the growth curve on the left, the cumulative user base of the Capy family has exceeded 45 million, maintaining rapid growth since the beginning of 2025.

Our goal is not simply to deploy a general-purpose chatbot across all scenarios, but to provide truly proactive and professional agents tailored to specific goals in personal life. For example, leveraging daily updates in multimodal capabilities, Capy Camera has launched an AI photography agent that proactively understands scenes and creative intent, helping users capture images and create content more easily. Capy Bookkeeping has been upgraded to a full-scenario financial agent, supporting income and expense insights, budget planning, and multi-account management.

We have also launched a new member, Capy Health, and observed that its retention rate has reached a leading level among similar products. We believe the value of personal agents should not be limited merely to faster execution; more importantly, they should help users achieve their goals and enjoy a better quality of life. In the long term, we aim for the Capy series to become personal life assistants that truly understand users and provide sustained companionship.

Visual AI remains a core business for SenseTime and serves as a mature entry point for system-level AI capabilities to penetrate industries and overseas markets. Our Visual AI business now covers more than 20 countries and regions, serving over 4,500 clients domestically and internationally. Furthermore, the sustainability of our Visual AI business is robust, with a revenue contribution rate from existing customers reaching 67%. This means that 67% of current period revenue comes from customers who purchased our products or services in the previous year, reflecting our long-term accumulation in customer base, industry understanding, and scaled delivery capabilities.

More importantly, our product and service ecosystem has become more comprehensive. As client needs extend from single-point applications to broader business processes, SenseTime continues to provide multimodal models, agents, and computing power services. These support analysis result filtering, content realization, digital content generation, and cybersecurity optimization, continuously enhancing the depth and breadth of our services. In terms of market position, according to IDC's China AI Software Market report, SenseTime has ranked first in China's computer vision market for ten consecutive years.

Notably, at the National Science and Technology Awards Conference held this July, we were awarded the Second Prize for National Scientific and Technological Progress. This recognition highlights our accumulated technical strength and industrial contributions in the field of visual intelligence. Finally, I would like to introduce SenseTime's perspective on the ecosystem and our forward-looking strategic layouts over the years in areas such as world models, embodied AI, intelligent terminals, optical computing, quantum computing, space computing, and AI infrastructure.

Our ecosystem enterprises include both companies incubated and nurtured internally under the 'One Plus X' strategy and upstream/downstream partners that form synergies through strategic investments. Our investment logic is not simply to chase hot trends, but to selectively layout and continuously cultivate directions with long-term technological scarcity, scenario value, and scaling potential, focusing on key supplies for the future AI industry.

These ecosystem enterprises provide SenseTime with real-world scenarios, industry feedback, productization capabilities, and commercialization channels. In return, SenseTime helps ecosystem partners shorten their R&D and implementation cycles through models, token factories, and engineering capabilities. As these ecosystem enterprises move from zero-to-one technology and product validation to one-to-hundred scaled applications, their industrial and commercial value is gradually being realized. This year, companies such as Daxiao Robotics and Hope SenseTime Medical completed new rounds of market-oriented financing, reflecting further recognition from capital markets and industrial partners regarding their technical routes and commercial progress.

In the first half of this year, some AI ecosystem companies in which we hold equity stakes successfully went public, leading to an increase in the fair value of related investments. However, compared to short-term financial returns, the more significant value of our ecosystem layout lies in enabling SenseTime to enter frontier scenarios such as embodied AI, intelligent terminals, and next-generation computing platforms earlier. This allows us to continuously obtain real-world task data and feedback, which in turn enhances the delivery capabilities of our unified multimodal models and systems. Next, I invite Wang Zheng to present the Group's financial performance for the first half of the year.

Mr. Wang Zheng

Thank you, Yang Fan. Hello everyone, I am pleased to share our performance results for the first half of 2026. Guided by our unified model architecture, token factory, and agent management system, SenseTime continues to focus on high-quality development in its core businesses. Our "1+X" strategy and the AI ecosystem we have built over the years are beginning to show increasingly positive financial effects.

During this reporting period, the company achieved healthy growth in revenue and gross profit, while operating costs under refined management continued to decline, resulting in simultaneous improvements in scale and efficiency. This helped us achieve our first net profit turnaround since our listing. In the first half of 2026, SenseTime recorded revenue of RMB 2.91 billion, representing a year-on-year growth of 23.4%. Among this, generative AI revenue reached RMB 2.33 billion, an increase of 28.2% compared to the same period last year, with its share of total group revenue rising from 77% in the first half of 2025 to 80% in the same period of 2026.

The rapid growth of our generative AI business has also driven a gradual improvement in the overall quality of the company's revenue. We disclosed recurring revenue at the group level for the first time, which reached RMB 1.14 billion in the first half of this year, a 124.4% increase from the same period in 2025, accounting for 39.3% of total group revenue. The largest source of recurring revenue is our generative AI business.

As our products and services become increasingly embedded in customers' workflows, assisting or even directly handling increasingly complex continuous tasks, we expect the contribution of recurring revenue to our overall business to further increase. Revenue from our visual AI business saw a steady rebound in the first half of 2026, with a 14% year-on-year growth, showing acceleration compared to the full-year growth rate of 2025. Our continuous and proactive optimization of the business structure in this segment has also enabled us to rank first in China's computer vision market share for the tenth consecutive time.

Our long-accumulated customer base, technology, and reputation have laid a solid foundation for the comprehensive integration of SenseTime's system-level AI capabilities into customers' business processes and for our all-around overseas expansion. Overseas revenue grew by 127% year-on-year in the first half of this year, significantly higher than the group's overall growth rate, demonstrating that the company is leveraging its leading visual AI capabilities, unified multimodal models, and localized delivery systems to expand its system-level AI capabilities to broader regions and customer groups.

SenseTime's gross profit for the first half of 2026 reached RMB 1.21 billion, with a year-on-year growth rate of 32.9%, marking a significant improvement in growth speed compared to previous years. The gross margin was 41.4%, an increase of 2.9 percentage points from 38.5% in the same period of 2025, surpassing our previous guidance range. This performance is the result of the company's continuous optimization of business structure and project quality, as well as ongoing improvements in delivery quality and efficiency through joint optimization of models and infrastructure, increased token production efficiency, product standardization, and reuse of underlying capabilities.

Healthy growth in gross profit from core business revenues, coupled with strict control over operating costs, and a significant increase in the fair value of both internally incubated enterprises and externally invested AI ecosystem companies under the "1+X" strategy, led to the group achieving positive adjusted and unadjusted EBITDA for the first time in the first half of 2026. We also recorded our first net profit under IFRS standards since our listing, amounting to RMB 620 million.

Even after excluding adjustment items such as increases in the fair value of financial assets and share-based compensation expenses, we found that the adjusted net loss for the first half of 2026 decreased by 67.3% compared to the same period in 2025, indicating a trend of accelerating loss reduction.

In terms of operating expenses, the company maintained strict and efficient control. Additionally, some X innovation businesses were deconsolidated following successful financing rounds. During the reporting period, sales, administrative, and R&D expenses all decreased year-on-year. Notably, among these three expense categories, those with larger amounts saw greater year-on-year declines. Total operating expenses for the first half of 2026 decreased by 14.8% year-on-year, a significantly larger reduction than in previous reporting periods. Achieving this amidst healthy growth in revenue and gross profit is no small feat, highlighting SenseTime's substantial potential for long-term profitability.

We have always placed a high priority on R&D investment to continuously strengthen the company's core competitiveness within the industry. Overall R&D expenses decreased by 17.1% year-on-year, primarily due to the deconsolidation of several innovative business units. Excluding the impact of deconsolidation, R&D expenses at the group level actually increased, particularly in computing power infrastructure.

Our overall working capital efficiency improved slightly in the first half of 2026 compared to the same period in 2025, with the cash conversion cycle (calculated using averages) shortening by 34 days compared to a year ago. It is worth noting that as the company continues to improve the overall quality of its revenue, including an increase in the proportion of recurring revenue, the impairment allowance ratio for trade receivables on the balance sheet decreased from 67.7% a year ago to 46.5% at the end of the reporting period. This relatively increased the accounts receivable balance and days sales outstanding (DSO) on the books.

The significant decrease in the impairment allowance ratio was reflected in the income statement, resulting in the net loss from impairment of financial assets turning positive for the first time, shifting from a loss of RMB 143 million in the first half of 2025 to a gain of RMB 35 million in the first half of this year. Our capital expenditures reached RMB 2.34 billion during the reporting period, mainly focused on the construction of large-scale computing power infrastructure.

The increase in capital expenditure is based on our judgment of the explosive growth in the AI industry and our confidence in the long-term, efficient monetization of the company's full-stack system-level AI capabilities. These investments will lay a solid foundation for the company's long-term healthy development. We will also maintain flexibility, continuously balancing the advantages of both asset-heavy and asset-light operating models.

As of the end of the first half of 2026, the company's total cash reserves reached RMB 13.8 billion. This definition of cash includes structured deposits but excludes the record-high balance of equity and bond investments totaling RMB 12.8 billion listed separately. In addition to incubating multiple successful 'X' innovative businesses internally, SenseTime has strategically built a diversified AI ecosystem through external investments over the years.

The fair value of these investments has approached RMB 11.5 billion by the end of the reporting period, driven by the IPOs and valuation increases of the invested companies. Furthermore, as of the end of the reporting period, the company's unused bank credit facilities amounted to RMB 9.9 billion, a year-on-year increase of 36%, while our overall debt level decreased significantly by 25.7% compared to a year ago. Overall, our financial position is robust with ample liquidity, providing sufficient support for long-term strategic layout and business development.

That concludes my remarks on the financial section. I will now hand over to Dr. Xu Li.

Dr. Xu Li

Thank you, Wang Zheng. To summarize our outlook for the next phase: future AI must not only have higher intelligence ceilings but also enter more scenarios at lower costs, collaborate to complete more complex tasks, and continuously learn and evolve through real-world outcomes. Therefore, centered around a unified model, a token factory, and an agent management system, we will focus on advancing the following three directions.

First, promote the continuous evolution of multimodal long-horizon agents in real-world tasks, with key breakthroughs in complex task decomposition, long-horizon execution, tool orchestration, dynamic error correction, long-term memory, and multi-agent collaboration. Data generated from tasks, execution results, and user feedback will continuously feed back into model training and workflow optimization, forming a closed loop for self-evolution and self-iteration of the model.

Second, we are driving the "token factory" to evolve from large-scale intelligent production toward task delivery. We will continue to promote the large-scale commercial adoption of domestic computing power, optimize heterogeneous inference, and enhance the synergy between computing and electricity, thereby significantly reducing the cost of high-quality tokens. As AI agents become capable of handling increasingly complex tasks, our metrics for efficiency will gradually shift from the cost per token to the quality, success rate, and cost of task completion.

Third, we are advancing multi-agent collaboration to integrate more deeply into individual and enterprise workflows. Multiple agents with specialized capabilities will share context, coordinate planning, cross-verify each other’s work, and jointly complete complex tasks. By leveraging diverse products and accumulated industry knowledge, we aim to elevate AI from a point solution tool to a system-level productivity platform that covers understanding, planning, execution, verification, and delivery.

In summary, SenseTime’s goal is to create a long-term flywheel connecting technological capabilities, token production, task delivery, and real-world feedback. This will continuously expand the audience and application boundaries of AI, fostering mutual reinforcement between customer value and business growth. Today, AI can answer questions; tomorrow, it will complete tasks. The transition from model intelligence to action intelligence represents a decisive turning point for the industry. We stand at the forefront of this commercial transformation. I believe the flywheel is already spinning, capabilities are scaling up, and the exciting journey has just begun.

That concludes my presentation. Thank you all.

Phil, Head of Capital Markets at SenseTime Group

Thank you, Dr. Xu Li, and thanks to the management team for their presentations. We will now move to the Q&A session. Investors on the phone line may press * followed by 1 to request to ask a question. Webcast participants may click the "Raise Hand" button on the screen. When asking a question, please state your name and affiliation, and try to limit yourself to one primary question to allow management to provide a comprehensive response. Next, we invite Ms. Stacy Wang from CICC to ask her question.

Wang Qianlei, CICC

Thank you, management, for giving me this opportunity to ask a question. I am Wang Qianlei, a computer sector analyst at CICC. I would like to seek clarification on the recurring revenue (RR) metric, which is being disclosed for the first time. Could you elaborate further? For instance, how does the RR figure disclosed by your company differ from the Annualized Recurring Revenue (ARR) typically cited by other model companies? What were the main considerations behind disclosing RR rather than ARR? Additionally, what are your expectations for future RR growth?

Mr. Wang Zheng

Thank you very much, Ms. Wang, for your question. Recurring Revenue, or RR, is defined as revenue derived from contracts effective during the reporting period that possess characteristics of continuous renewal. In the first half of this year, RR reached RMB 1.14 billion, representing a 124.4% year-on-year increase compared to the same period in 2025, and accounting for 39.3% of the Group's total revenue. The primary contributor to this figure is our generative AI business.

This is our first disclosure of RR (Recurring Revenue). Compared to the ARR you mentioned earlier, the underlying core logic and concepts are essentially consistent. However, RR is a more tangible and objective metric because it is calculated based on the actual business performance recorded in the first half of 2026. Therefore, it represents historical data for that period.

Since you mentioned ARR, we are pleased to announce that our ARR as of June this year has reached RMB 2.53 billion. By comparing this RMB 2.53 billion with the RMB 1.14 billion from the previous six months, you can see that our recurring revenue is growing very rapidly. We also expect the proportion of RR to total revenue to increase significantly in the second half of the year compared to the first half.

Dr. Xu Li

Let me add that ARR is relatively more susceptible to factors such as model release schedules and emerging hotspots. For instance, extrapolating monthly figures by multiplying by 12, or weekly figures by multiplying by 52, introduces volatility that fluctuates with operational rhythms. Therefore, disclosing RR may represent a more rigorous financial metric.

Wang Qianlei, CICC

Understood. Thank you to the management team for the clear explanation.

Phil, Head of Capital Markets at SenseTime Group

Thank you. Next, we invite Mr. Pan from CITIC Securities to ask his question.

Pan Rusheng, CITIC Securities

Hello Dr. Xu Li, and hello to all SenseTime executives. I am Pan Rusheng, a computer industry analyst at CITIC Securities. Thank you for giving me the opportunity to ask a question. I would like to inquire about SenseTime's key focus areas in the large model space going forward. We have noticed your Xiao Huanxiong office productivity product, the Capy series personal assistant products, and the Cycle video generation agent. How do you plan your future layout on the application side? For example, do you plan to venture into coding assistance? This is my main concern. Thank you, executives.

Dr. Xu Li

Thank you, Mr. Pan. First, our model R&D has remained consistent. We firmly believe that advancing large models through native multimodality will push the upper limits of AI intelligence. This includes long-horizon agents, as well as extensions into spatial and physical intelligence. SenseTime has been resolute on this path, and we believe it will soon reach a tipping point where quantitative growth leads to significant qualitative change at scale.

Regarding our applications, although they appear as front-end agents facing diverse multimodal scenarios—such as work environments, content generation, interaction, and personal assistants—they are not independent. They share a unified underlying model, a unified token factory, and a unified agent harness. They simply address different front-end tasks, allowing for substantial synergy across various high-frequency scenarios.

In terms of strategic choices, we entered the office scenario early. Many companies, both in the U.S. and China, have aggregated work-related scenarios in the first half of this year. This is because, as agent applications deepen, the target audience has shifted significantly: from professionals to the general public, and from back-office departments to front-line operations. This expansion has greatly increased the potential market size, making it an essential area for development.

For us, this shift represents a larger market opportunity, clarifying the commercialization logic. Secondly, we chose this path because work-focused agents align perfectly with SenseTime’s capabilities and market demands. For instance, in work scenarios, our multimodal capabilities—integrating understanding and generation—are highly compatible with most applications, providing a competitive advantage.

By the same token, Cycle’s agents in video and content generation fully reflect our strategic choice of multimodality. In summary, while we may have diverse agent scenarios, our commercial layout depends on combining our core capabilities—such as multimodality—with the ability to handle complex, long-horizon, verifiable tasks, and to achieve high-frequency usage and sustainable recurring revenue.

Regarding code, we launched our coding agent very early. Our 'Little Raccoon' (Xiao Huanxiong) coding product went online three years ago, making it one of the earliest offerings in this space. These capabilities have now been integrated into the unified entry point of the Little Raccoon platform. Coding models are a fundamental capability; when building agents, coding proficiency essentially extends to structured thinking and tool invocation.

We are currently more focused on how to further enhance the model’s logical and reasoning capabilities through other modalities, thereby elevating overall agent performance. A standalone coding product is not our current primary focus, given the intensifying competition in that specific segment. However, coding capabilities remain integral to our broader agent product offerings.

From a commercial perspective, we do not segment our business by individual products. Instead, we emphasize a unified agent control model (agent harness) and a unified token factory for iteration. This unification of underlying logic, models, tokens, and the agent harness simplifies front-end delivery and provides cost advantages.

Pan Rusheng, CITIC Securities

Alright, thank you.

Phil, Head of Capital Markets at SenseTime Group

Let's move on to the next question. We invite Mr. He from UBS.

He Xuhui, UBS

Thank you to the management team for the opportunity to ask a question. First, congratulations on the company's achievements in the first half of the year. I am He Xuhui from UBS, and joining me online are my colleagues Xiong Wei and Charles. I have a question regarding the company's strategy and products for the management team. We noted that in the earnings announcement, the company mentioned 'one model, one token factory, and one agent management system.' Could the management elaborate on our specific strategic approach when facing internet giants or large-model vendors like DeepSeek?

Specifically, will the future focus be more on foundational models, AI computing power, or industry-specific agent applications? In which areas do we see the most differentiation or the greatest potential to generate scalable revenue?

Dr. Xu Li

Thank you, Mr. He. First, regarding these 'three ones,' this does not imply a business model where we choose only one option. To some extent, the original model APIs can provide services. When converting to tokens, we refer to it as a 'token factory' because it involves not just a single model but a suite of models that deliver broader capabilities. We even serve leading vendors of the four foundational models mentioned earlier, helping them better refine the production efficiency and capacity of our token factory.

Our perspective is that as we take the next step, while the unit price of tokens may decline, the customer utility derived from a high-quality token-based agent management system will increase. Therefore, our models address the challenge of reaching peak performance, the token factory addresses cost and pricing issues, and the agent management system addresses task execution and, essentially, utility-based pricing. To some extent, this represents a continuous and evolving business model.

Regarding large tech companies or other model vendors, we believe we are among those who have developed a relatively complete framework across these three commercialization pillars, with strong synergy and coordination. Fundamentally, at any stage, as models mature, their commercial value gets compressed; similarly, as token factories mature, the commercial value of tokens gets compressed, leading to intensified competition.

Therefore, it is essential to extend these capabilities into an agent management system and deliver them to our customers. We observe significant shifts in the breadth of our customer base along this trajectory. Consequently, we aim for the integration of these 'three ones' to ultimately form a business model that better aligns with commercial value and captures the final utility value.

He Xuhui, UBS

Alright, thank you, Dr. Xu, and thank you to the management team.

Phil, Head of Capital Markets at SenseTime Group

We would now like to invite Mr. Yang from Guotai Junan Securities to ask a question.

Yang Lin, Guotai Junan Securities

Thank you, management, for the opportunity to ask questions. This is Yang Lin from Guotai Junan Securities. My question is: since the company achieved IFRS profitability for the first time, how much of this was driven by improvements in core operations, and how much came from investment income and gains from the disposal of subsidiaries? After excluding these items, what was the core operating loss? When do you expect to achieve sustained IFRS profitability based solely on core operations?

Could management also provide an overview of future capital expenditure plans? Specifically, what proportion of the current computing power capacity is allocated externally versus internally, and what percentage represents new assets? Apologies for the detailed questions. Thank you.

Mr. Wang Zheng

Thank you very much, Mr. Yang, for your series of questions. I will attempt to address them one by one. First, it is true that we achieved IFRS profitability for the first time in the first half of the year. Fundamentally, this is attributable to healthy growth in revenue and gross profit. As mentioned earlier, gross profit increased by 32.9%. Meanwhile, total operating expenses—specifically the sum of the three major expense categories—decreased by 14.8% compared to the same period last year, representing a significant decline.

These two factors are clearly directly related to our core business and constitute the primary drivers. You also noted that there were other factors contributing to the IFRS results. Following your suggestion, if we exclude certain non-recurring items, such as adjustments for changes in the fair value of financial assets, and make even more aggressive adjustments—for instance, excluding gains arising from the deconsolidation of subsidiaries—

We have indeed conducted internal calculations. After excluding certain items, we found that the reduction in losses for these specific performers in the first half of 2026 compared to the same period in 2025 was actually more significant. In absolute terms, the loss narrowed by approximately RMB 1.2 billion. However, if we use our commonly disclosed metric of adjusted net loss, the narrowing of the absolute loss value is only around RMB 800 million.

Therefore, using a more aggressive adjustment method actually reveals a larger absolute reduction in losses, with the percentage decrease being quite substantial, at approximately 60%. From any perspective, this represents a fundamental trend of rapid and significant loss reduction, driven essentially by gross profit growth in core businesses and efficient declines in operating costs.

We fully understand the concern from investors and analysts regarding the timing of profitability. Looking ahead, with sustained healthy growth in revenue and gross profit, coupled with effective cost control, we expect to achieve profitability at the adjusted EBITDA level for the full year of 2026. In terms of IFRS net loss, or specifically adjusted net loss, we anticipate a significant proportional reduction in losses for the full year of 2026 compared to 2025. This is my current assessment. Thank you.

Additionally, you previously raised some points about capital expenditure (CapEx), which I would like to discuss further. There is inherent uncertainty in CapEx, particularly because the supply side of the computing power market carries significant intrinsic uncertainty. Our consistent strategy of combining asset-light and asset-heavy models provides us with considerable flexibility, allowing ample room to adjust between purchasing and leasing based on prevailing supply-side conditions.

This flexibility is highly valuable, enabling us to better capitalize on favorable market conditions compared to some massive industry giants. Furthermore, the projected unit economic return on computing power CapEx clearly influences the intensity of CapEx growth. Based on current comprehensive market conditions, the return on investment for computing power CapEx appears very attractive.

If this favorable environment persists over the long term, we believe CapEx growth will definitely outpace revenue growth during this period due to the strong returns. However, precisely because high CapEx returns naturally drive a more substantial increase in future revenue, it is likely that in the subsequent stage, the CapEx growth rate will fall below the revenue growth rate.

Thus, from a longer-term perspective, the long-term growth rate of capital expenditure should gradually fall below the revenue growth rate. You also mentioned the ratio of external to internal computing power usage. To provide a rough estimate, our current split between external services and internal self-use is approximately 70:30. Regarding the ratio of asset-light to asset-heavy structures, it is roughly 50:50, maintaining a relatively flexible state.

Yang Lin, Guotai Junan Securities

Thank you, Mr. Yang. All right, thank you.

Phil, Head of Capital Markets, SenseTime Group

Next, I will invite Maggie from CITIC Securities to ask questions.

Maggie from CITIC Securities

Good evening to the management team. Thank you for taking my questions. I am Maggie, an analyst at CITIC Securities. In your earnings announcements, you have repeatedly mentioned SenseTime's integration of native multimodal capabilities with agent functionalities. Could you please elaborate on how these capabilities are integrated? Furthermore, based on these capabilities, which industries and application scenarios will you prioritize for deeper penetration? And how do you plan to achieve more concrete productization? Thank you.

Dr. Lin Dahua

Let me address this question, and thank you, Ms. Ye, for raising it. First, I would like to clarify one point: when we refer to long-horizon multimodal agents, we do not mean a simple patchwork where an agent merely calls upon a multimodal model. Instead, we aim to deeply integrate multimodality with long-horizon agent capabilities.

In fact, across the vast array of business scenarios we cover, users need to collect diverse forms of information to complete tasks, such as reviewing reports containing mixed text and images, navigating different software user interfaces, and analyzing images. This is a natural part of the workflow.

In this process, handling multimodal information and performing deductive reasoning are inevitable. Many current agents simply use tools to invoke multimodal perception and understanding capabilities. However, these tools often lack a complete context to accomplish the task, resulting in low efficiency in tool invocation.

If we were to combine an agent model with a multimodal model for text and images through simple concatenation, it would involve significant context transfer between different models, leading to decreased token efficiency. Our core approach to solving this problem is leveraging Leo's unified new architecture, which allows us to process image, text, and coding information within a shared context. This avoids any loss of context and eliminates the efficiency losses associated with transferring context across different models.

In real-world environments, this approach has demonstrated exceptional work efficiency. Beyond the efficiency gains mentioned earlier, we found that integrating multimodality into our agents enables more effective processing and comprehensive analysis of various modalities. This empowers the model to perform more complete tasks and enhances its ability to execute complex, multi-step, long-horizon assignments.

These advantages are fully reflected in scenarios ranging from content creation to complex enterprise office workflows. Looking ahead, the agent harness control system we have built on top of this foundation will help us more effectively unleash the capabilities of our native multimodal long-horizon agents, thereby better serving our customers.

Maggie from CITIC Securities

Okay, thank you.

Phil, Head of Capital Markets at SenseTime Group

Let us now invite Mr. Zhou from GF Securities to ask his question.

Mr. Zhou from GF Securities

Good day to the leadership team at SenseTime. First, congratulations on the company's outstanding performance in the first half of the year. I would like to ask about the company's competitive moats and its subsequent commercialization roadmap. Specifically, how does the company's model capability form a sustainable competitive barrier? We are aware that foundation models both domestically and internationally are undergoing rapid iteration, while the performance of open-source models continues to improve as their costs decline.

In this context, how can the company's Leo Unify native multimodal architecture and its long-horizon agent capabilities be converted into customer revenue? Regarding the competitive landscape, compared to domestic peers such as Alibaba, ByteDance, and DeepSeek, as well as international model providers, what are the company's differentiated advantages that are difficult to replicate? Are they rooted in the model itself, computing power, agent harnessing, or industry delivery capabilities?

Dr. Dahua Lin

Thank you for your question, Mr. Zhou. I believe this is a core issue within the industry. As the AI sector enters the era of agents, the key value of AI lies in its ability to solve user problems and deliver ultimate user value. This is critical to the sustainability of our AI business.

As you mentioned, model iteration is currently very rapid, with new versions released every two to three months, and declining model prices represent a long-term trend. Against this backdrop, it is difficult to establish a lasting barrier by relying solely on a single model or outstanding performance on a one-time benchmark.

From our perspective, the true barrier, as Mr. Xu mentioned earlier, is the closed-loop flywheel we have established. There are two important aspects worth noting here. First, the closed loop formed by our models, our agent systems, and user scenarios allows our multimodal capabilities to gather substantial feedback on end-user behavior and needs while serving them.

This feedback loop enables our entire model and agent architecture to develop self-evolving capabilities, leading to continuous improvements in both model performance and overall system capacity. As user adoption and product usage grow, we expect the evolution of our model and system capabilities to accelerate further.

On another note, as you mentioned, token prices are continuing to decline, making service cost a critical factor. Thanks to SenseTime, we possess both strong model innovation capabilities and a robust large-scale infrastructure system built up over many years. The deep synergy between these two aspects allows us to maintain highly competitive efficiency and cost structures for our model services.

This approach gives us a sustained advantage in cost control while also feeding back into the continuous upgrading of our large-scale infrastructure system. In summary, we believe our enduring competitiveness does not stem from any single aspect, but rather from the closed-loop integration and continuous evolution of our models, applications, and underlying infrastructure. As our user base expands, this evolutionary pace will accelerate, helping us build sustainable long-term barriers to entry.

Mr. Zhou, GF Securities

Thank you. That was very clear. I have no further questions.

Phil, Head of Capital Markets, SenseTime Group

Alright, thank you. Due to time constraints, let's move to the final question. We invite Mr. Li from Citi.

Li Yiming, Citibank

Thank you. Good day, management team. I am Li Yiming, an analyst at Citibank. Thank you for the opportunity to ask questions. I would like to inquire about the company's Token Factory strategy. Given the rapid expansion of major internet companies and public cloud providers, who are also laying out services in this area, could you clarify what specific services the Token Factory offers? Additionally, where do you see points of competition or cooperation with these internet giants and public cloud providers?

Additionally, regarding computing power investment, does the company have specific plans? Will it rely primarily on leasing, or will it continue to build its own infrastructure? Are there any quantitative targets or guidance for this? Those are my two questions. Thank you.

Mr. Yang Fan

Let me address this question. Thank you, Mr. Li. Regarding the concept of a "token factory," there is a general consensus that it refers to providing token production capabilities. At SenseTime, we leverage our infrastructure software capabilities to efficiently and effectively produce tokens by integrating our proprietary日日新 (Ri Ri Xin) models, high-impact open-source models available in the market, and specialized models from select partners. We then provide these services to third parties in the industry.

Compared to major internet tech giants, we believe we have several distinct competitive advantages. First, the infrastructure of today's tech giants primarily serves their own core businesses, whether in e-commerce, social media, content, or other sectors. In contrast, our positioning is more externally focused, serving both small and medium-sized AI application developers and enterprise clients. Indeed, data indicates that SenseTime is currently the largest native AI infrastructure service provider in the market, with a clear mandate to serve third parties.

Second, we observe that neither downstream models nor upstream chips—particularly domestic chips—are currently "all-rounders"; each tends to have its own specific strengths. A key characteristic of our approach is our high degree of openness regarding GPU hardware and models. This allows us to offer customers tailored, cost-effective combinations of solutions that best fit their specific application needs.

Third, rather than simply offering standardized computing power or token products, we focus on addressing the key demands of different customer segments. We combine the production capabilities of our "token factory" with our service expertise to deliver solution-oriented outcomes that help customers solve real-world problems.

Finally, as previously mentioned, we have accumulated significant innovations in areas such as compute-network synergy and green energy utilization. These advancements help us achieve a differentiated cost advantage. To answer your question about computing scale, our current operational computing capacity is approximately 48,000 GPUs.

However, our current focus is less on the sheer scale of computing power and more on how we can leverage various chip combinations to produce high-quality tokens, continuously reduce token production costs, and ultimately optimize task completion. In this regard, as noted earlier, our daily average token service volume in July increased by more than twenty-fold compared to a year ago. We are currently providing services to four external foundation model providers.

Therefore, we anticipate that future investments in computing power will continue to grow at a robust pace, driven by clear customer demand and token consumption. Our strategic focus will be on domestic computing resources, nodes in Hong Kong and overseas, and green energy initiatives. Thank you.

Li Yiming, Citi

Okay, thank you.

Phil, Head of Capital Markets, SenseTime Group

Thank you, everyone. Due to time constraints, we will conclude today's Q&A session here. If you did not have the opportunity to ask questions, please feel free to follow up with our Capital Markets team. Once again, thank you to the management team, and thank you to all investors, analysts, and media friends for attending today's earnings conference call. This concludes today's meeting. Thank you all, thank you. We appreciate your participation. Goodbye.

More details:SENSETIME-W IR

Disclaimer: The above content is generated by an AI language model based on public data and third-party automatic subtitles. The above content does not represent any position of Futu and does not constitute any investment advice. Futu Group makes no express or implied warranties or representations regarding the accuracy, timeliness, or completeness of the above content.

37K Views