Dissecting SenseTime’s Financial Report: Over 5 Billion Yuan Collected in One Year—Who Are the Buyers of AI Large Models?

Ask AI · How does SenseTime’s AI business achieve a spectacular turnaround from loss to profit?

“By 2026, the global artificial intelligence industry will enter the true ‘finals’ stage.” On the evening of March 24, at the SenseTime Technology (0020.HK) 2025 annual earnings release, Chairman and CEO Xu Li stated.

The scope of this “finals” has crossed borders and set sail overseas. Facing this global competition, SenseTime holds a highly convincing chip. Currently, the capital market’s standards for evaluating AI companies have shifted from “how flashy the technology is” to “how healthy the books are” and “whether the business model can be replicated across regions.” In the major test of commercialization, SenseTime has delivered a report card of “returning to high growth” and “breaking out overseas.”

Financial data shows that in 2025, SenseTime Group achieved total revenue of over 5 billion yuan, a year-on-year increase of 33%, marking the fastest growth in nearly three years; losses narrowed significantly by 58.6% year-on-year, and EBITDA shrank by 85%. Among them, customers in Southeast Asia, Northeast Asia, the Middle East, and other regions continued to repurchase visual AI products and services, while more clients in South America, Europe, and elsewhere began to pay attention to this business. SenseTime’s visual AI has formed a replicable and scalable “Chinese-style overseas expansion plan.” Additionally, SenseTime’s consumer applications repeatedly topped app charts in multiple countries, successfully exporting “Chinese-style blockbusters” overseas.

What further drew market attention was the reversal of SenseTime’s core profit indicators and cash flow. In the second half of 2025, SenseTime’s EBITDA reached 380 million yuan, achieving its first positive figure since listing; simultaneously, the group’s operating cash flow also turned positive for the first time in the second half. The cash cycle (CCC) was sharply shortened from 228 days at the end of 2024 to 129 days.

In financial statements, turning EBITDA positive and achieving positive operating cash flow are often regarded as milestone signs that a tech company has crossed the investment phase and possesses self-sustaining profitability. Moving from the technical investment stage to explosive growth, from domestic deep cultivation to global deployment, what growth logic and strategic restructuring underlie SenseTime’s impressive results?

Generative AI Takes the Lead: Accelerating Commercialization of Large Models

Dissecting SenseTime’s revenue exceeding 5B yuan, it’s clear that generative AI is no longer just a concept but the main engine driving growth.

In 2025, SenseTime’s generative AI business revenue reached 3.63 billion yuan, a substantial increase of 51% year-on-year. More importantly, this business now accounts for over 70% of the group’s total revenue, establishing an absolute pillar position. Behind this explosive growth is the continuous surge in market demand for training, fine-tuning, and inference of large models, as well as SenseTime’s “Daily New” large models accelerating deployment across thousands of industries.

Unlike the industry’s previous focus on “ranking lists” and “showing parameters” two years ago, SenseTime’s actions in 2025 are more grounded. The core logic boils down to four words: scene monetization. SenseTime explores the closed-loop value of intelligent agents across various strategic tracks by considering task complexity and fault tolerance.

In the broad office domain, SenseTime’s Little Raccoon family has demonstrated remarkable penetration. Currently, this product has served over 15 million individual users and thousands of enterprise clients, with monthly active users increasing sevenfold over the year. By autonomously creating high-quality, editable PPT documents, its enterprise data analysis product has improved data accuracy to 95%, and shortened business analysis cycles by 90%. For enterprises, this is a tangible cost reduction and efficiency boost tool; for SenseTime, it is a highly sticky recurring revenue source.

In the intelligent marketing track, SenseTime has made AI no longer just a “brain” for chat companionship, but also grown “hands” capable of operating backend systems and executing tasks. For example, in strategic cooperation with ValueBuy Technology, SenseTime’s AI technology helped increase store operation efficiency by 20 times and live broadcast efficiency by 6 times. In film and content creation, SenseTime’s first integrated creative AI agent Seko has enabled the “one-person crew” concept, attracting over 300k creators within months of launch.

Besides building scalable business models on the B-end (enterprise side), SenseTime’s exploration on the C-end (consumer side) has also shown bright results. The industry’s first multimodal large model-based AI photo assistant “Kapi Camera” topped app store charts in multiple countries; “Kapi Ledger,” as a young people’s AI financial assistant, maintains a T+1 retention rate of 70%, a top industry level. With millions of users, these native consumer applications demonstrate the explosive potential of AI-driven consumer-grade products.

As SenseTime co-founder and Chief Scientist Lin Dahua pointed out, the entire industry has entered the Agentic AI era. Deeply connecting the visual and language capabilities of intelligent agents is greatly improving success rates and efficiency in complex tasks.

Visual AI and Computing Power Infrastructure: CV2.0 Achieves Profitability First, Building a “Horsepower” Fleet

While generative AI surges forward, SenseTime’s original business—visual AI—is quietly transforming.

Financial reports show that in 2025, revenue from SenseTime’s visual AI business was 1.08 billion yuan, maintaining a steady trend. Notably, this traditional advantage has undergone a complete upgrade. Official disclosures reveal that CV2.0 achieved profitability for the first time during the reporting period and has maintained positive cash flow for two consecutive years, becoming a solid pillar driving revenue growth and cash flow improvement.

SenseTime has built a “fusion of general and specialized, integrated training” multimodal algorithm production system. Large models handle understanding complex long-tail scenarios and perform progressive reasoning, while small models efficiently execute specific tasks. This approach effectively solves the pain points of traditional AI deployment—balancing high computational costs with recognition accuracy. Projects deploying related capabilities soared from nearly 0% in 2024 to about 60% in 2025. As the nine-year leader in China’s computer vision market share, SenseTime is transforming accumulated B-end experience into scalable profits with a more agile model.

Supporting this is SenseTime’s massive computing infrastructure—SenseCore.

In 2025, SenseCore achieved a deep leap from technical strength to industry closed-loop. To date, its total computing capacity has increased to 40.4k PetaFLOPS. It has established a complete path from hardware to top-layer applications, from software stacks to model adaptation, supporting nearly one million model development tasks throughout the year.

Facing the high energy costs of computing power, SenseTime’s refined operational capabilities are emerging. With support from CATL, it has developed the world’s first AI intelligent system integrating “compute management, IDC operation, and energy storage.” This system can accurately predict power loads via large models and dynamically generate optimal energy dispatch strategies, potentially saving 7% on electricity bills and reducing over 4,000 tons of carbon annually per 10,000 PetaFLOPS.

Additionally, in the highly watched domestic compute adaptation and overseas expansion, SenseTime is leading the way. Yang Fan, Vice President of SenseTime and President of the SenseCore division, revealed that in 2025, the “SenseCore Compute Mall,” jointly launched with more than a dozen domestic chip manufacturers, is accelerating the domestic system from “usable” to “highly usable.” It also transforms leading domestic AI computing experience into global competitiveness, with the first overseas Chinese compute cluster launched in Saudi Arabia. By exporting highly adaptable SenseTime software stacks, it demonstrates excellent delivery in heterogeneous environments and a light-asset overseas model.

Overseas Expansion and Profits: Exporting System Capabilities to Draw the “Smile Curve” of AI Going Global

As the underlying computing power and application ecosystem mature, SenseTime’s overseas strategy made a qualitative leap in 2025. Unlike the past where it mainly sold products and sought clients, SenseTime is now exporting a replicable “Chinese-style system capability” that includes “applications, computing power, and talent.”

On the path of internationalization, SenseTime employs a dialectical approach of “B-end rooted, C-end breakthrough.” Besides the global success of “Kapi Camera,” its deep cultivation of overseas B-end markets has also borne fruit: in Indonesia, SenseTime partnered with the largest private medical group Siloam to deploy the SenseCare chest X-ray intelligent analysis solution. By ensuring algorithm inclusiveness, local diagnostic standards, and language adaptation, this became China’s first successful overseas AI imaging solution; in Saudi Arabia, SenseTime not only provides intelligent computing centers but also co-founded the AI Innovation Center with King Saud University, empowering local education and research; in South America, SenseTime’s Visual AI “Ark” platform has attracted over 150 multinational and local enterprise executives to China, exploring “Chinese-style innovation” solutions for urban governance.

While the industry is still expanding overseas clients, SenseTime has already established a complete international business cycle. Behind this are three strategic choices: not “de-Chinese-ifying,” but “localizing Chinese capabilities”; not “scaling first, then profit,” but “recurring purchases first”; not “standardization to dominate,” but “standard base plus flexible adaptation.”

Leveraging the “Deep-Hong Kong synergy” rule adaptation capability, SenseTime has successfully overcome integration barriers under different sovereignty requirements, precisely positioning its overseas business along the high-value “smile curve” of “compute infrastructure” and “application solutions,” drawing a high-value growth trajectory.

“1+X” Strategy Enters High-Plateau and Deep Waters: Native Multimodal Positioning in the “Second Half”

The improvement in capital efficiency, self-sustaining core business, and solid foundation of the infrastructure provide confidence for SenseTime’s longer-term strategic layout. Since the 2024 launch of the “1+X” strategic restructuring, the ecological momentum of this strategy is fully unleashed in 2025.

In this system, “1” represents SenseTime Group’s core businesses: generative AI and visual AI; “X” is the ecosystem of incubated companies. Financial reports show that X’s innovative business revenue in 2025 was about 300 million yuan, a slight decline year-on-year. But this is not business contraction; rather, mature sectors like autonomous driving have successfully completed independent financing and separated from consolidated financials.

By building an efficient “mother ship (group) providing the base, sub-ships (ecosystem) competing in tracks” collaboration system, SenseTime has achieved a leap from single business growth to ecosystem value multiplication. Over the past year, progress in vertical fields like the embodied intelligence robot Daxiao, the household robot Yuanluobo, and SenseTime Medical has been promising, gaining high recognition from internet giants and top-tier venture capital. This model grants innovation businesses great flexibility and explosive potential within vertical markets.

At the technological crossroads, SenseTime firmly believes in “native multimodal large models.”

“Pure language data has peaked; the deep integration of language and vision will be the key path to突破 AI’s upper limit,” Lin Dahua said. In December 2025, SenseTime released and open-sourced the new native multimodal model architecture NEO. Unlike mainstream “encoder-connector-LLM core” splicing architectures, NEO achieves true fusion of text and image information at the core level, enabling understanding of complex physical world analysis and decision-making. Even more impressive is its efficiency: it can reach top performance with only one-tenth of the data and compute of industry-standard models.

The benefits of forward-looking technology are about to be realized. SenseTime announced that in Q2 2026, it will launch a new foundational model based on the second-generation NEO architecture, pioneering the industry’s understanding and generation under the native multimodal “scaling law,” expected to achieve multiple-fold improvements in performance and cost-effectiveness.

“With ongoing operational strategy optimization, SenseTime is fully prepared,” Xu Li outlined the future blueprint at the earnings conference: defining new heights of AI with the native multimodal route, creating the entry point for the Agent era, optimizing unit AI costs to promote domestic innovation, scaling up visual AI business, and continuously releasing the ecological value of “1+X.”

Enduring the cyclical pains, through strategic and structural reshaping, SenseTime proved in 2025 that with over 5 billion yuan in revenue and key profit turning positive, its business model is closed and resilient. As the AI industry shifts from “model competition” to “deployment and profitability,” this tech giant is entering its own golden era with a lighter, more agile posture.

Reporting by: South China Morning Post Finance Reporter Yan Zhaoxin