Chinese artificial intelligence company SenseTime Group said in Hong Kong on August 26 that it recorded its first profit under International Financial Reporting Standards since listing, supported by growth in generative AI and valuation gains from investments in its AI ecosystem. Revenue for the first half of 2026 increased 23.4% year on year to RMB2.91 billion ($405 million), while gross profit rose 32.9% to RMB1.21 billion.
Its gross margin increased by 2.9 percentage points to 41.4%. Generative AI generated RMB2.33 billion, accounting for nearly 80% of total revenue, while overseas revenue increased 127% from a year earlier. SenseTime also disclosed recurring revenue for the first time, reporting RMB1.14 billion, or 39.3% of group revenue.
The company recorded an IFRS net profit of RMB620 million, compared with losses in previous reporting periods. However, its non-IFRS adjusted net result remained a loss of RMB390 million, although the deficit narrowed by 67.3% year on year. SenseTime said fair-value gains from higher valuations of some AI-related investments contributed to its IFRS profit.
The difference between the two figures is significant. The IFRS result marks a financial turning point, but the adjusted loss shows that SenseTime’s core operations have not yet reached full profitability. Its next phase will therefore depend on whether higher-margin generative AI services and recurring contracts can convert revenue growth into sustainable operating earnings.
Generative AI becomes the central business
SenseTime has reorganised its commercial AI strategy around three components: foundation models, computing infrastructure described as a “Token Factory,” and an agent platform connecting models with enterprise tools and workflows. The structure is intended to move the company beyond model development and computing provision towards AI systems capable of completing defined tasks.
During the first half, SenseTime released models covering multimodal reasoning, visual perception, content generation, spatial intelligence and robotics. These included the SenseNova U1 series, SenseNova Vision, SenseNova 6.8-Flash-Lite and SenseNova SI-8B, as well as the Kairos 3.1 world model developed with Ace Robotics. Claims that some models matched or exceeded products from Google, OpenAI and DeepSeek on selected benchmarks were based on results reported by the company.
SenseTime is also expanding SenseCore, its computing infrastructure, into a platform for producing and delivering AI inference at scale. In July, the system processed an average of 2.4 trillion tokens per day, approximately 22 times the volume recorded a year earlier. Total deployed computing power had reached 48,000 petaflops by the reporting date.
The company said infrastructure and model optimisation had tripled training speed for long-context agent models and increased image and video generation efficiency sixfold. It has also adapted the SenseNova U1 model to more than 10 Chinese chip platforms and provides computing services to four external foundation-model developers.
These figures point to a broader shift in the AI market. As models become more widely available, competitive advantage increasingly depends on inference cost, system utilisation and the ability to connect models with revenue-generating workflows. SenseTime’s emphasis on token production reflects an attempt to control both the intelligence layer and the infrastructure used to deliver it.
Energy management is part of that strategy. SenseTime said its compute-energy management agent achieved 96% load-forecasting accuracy and reduced electricity expenses by more than RMB12 million during the period. It also estimated that the system lowered annual carbon emissions by 24,000 tonnes for every 10,000 petaflops of computing power.
Agents expand into enterprise workflows
SenseTime reported that its AI agents served more than 1,000 corporate customers across over 20 industries, including education, government, logistics, banking and insurance. Customers of its office agent Raccoon included Lenovo, Ping An Technology, JD.com and China’s three major telecommunications operators.
The company said total users increased more than fivefold year on year, while enterprise users rose more than sixfold. Monthly active users on its cloud platform increased nearly tenfold, and its consumer-focused Kapi agents surpassed 45 million cumulative users across health, finance and content-creation services.
Its Seko video-generation agent was producing 10,000 minutes of video per day. SenseTime estimated that content made through the platform had accumulated more than 1.5 billion views on short-video services, although audience reach does not by itself indicate the revenue generated by the product.
Computer vision, SenseTime’s original core business, generated RMB500 million, equivalent to about 17% of total revenue. The division served more than 4,500 customers across over 20 countries and regions, with repeat customers accounting for 67% of its revenue.
The results show that SenseTime is becoming less dependent on conventional computer-vision projects and more exposed to generative AI, inference infrastructure and agent-based services. The rising proportion of recurring revenue could improve earnings visibility, while overseas growth may reduce reliance on China’s domestic market.
The central test will be whether this changing revenue mix produces sustained operating profitability without relying on investment valuation gains. SenseTime plans to develop pricing based on completed tasks rather than token volume, linking revenue more directly to business outcomes. If implemented at scale, that approach would represent a shift from selling AI usage to charging for measurable work performed by AI systems.
