For years the question around China's AI giants was simple: when do they stop burning cash? For SenseTime (商汤), the answer arrived in its latest earnings — and the reason is not a single product but a shift in how the company packages intelligence.
Key takeaways
- SenseTime (商汤, 0020.HK) reported its first-ever half-year profit under IFRS: net profit of RMB 617.3 million (≈ US$86 million) for the first half of 2026.
- Revenue reached RMB 2.91 billion (≈ US$405 million), up 23.4% year on year, with generative AI contributing RMB 2.33 billion (≈ US$324 million) — 79.9% of the total.
- The company's "one model system, one Token Factory, one Agent Harness" framework reframes its business around recurring token services and task-completing agents rather than one-off project delivery.
- SenseCore, its AI cloud, delivered about 2.4 trillion tokens per day in July 2026, roughly 22 times the year-ago level, on total deployed compute of 48,000 petaflops.
- Overseas revenue jumped 127% year on year, and open-sourced SenseNova models drew more than 11,000 GitHub stars within four months.
From computer vision to a token economy
SenseTime made its name in computer vision — faces, cameras, cities. That business is still there, but it is no longer the story. In the first half of 2026, generative AI accounted for nearly four-fifths of revenue, and the company said recurring revenue — income from ongoing contracts rather than single projects — reached RMB 1.14 billion (≈ US$159 million), or 39.3% of the total, up from 21.6% a year earlier. Recurring revenue more than doubled, growing 124.4%.
The numbers matter because they describe a company trying to escape the project-by-project services trap that has weighed on Chinese AI earnings. Gross profit rose 32.9% to RMB 1.21 billion, and gross margin improved to 41.4%. On a non-IFRS basis the company still posted an adjusted net loss of RMB 386 million, but that loss narrowed 67.3% year on year — the gap between the headline profit and the adjusted loss largely reflecting fair-value gains on invested companies.
The "Token Factory" thesis
SenseTime's chairman and CEO, Dr. Xu Li (徐立), summarized the strategy as "one model system, one Token Factory, one Agent Harness." The middle piece is the most concrete. SenseCore, the company's AI cloud, is being repositioned from a training platform into what it calls a production-grade "Token Factory" — infrastructure that turns raw compute into standardized, billable intelligence at scale.
The scale claim is specific: in July 2026, SenseCore served an average of 2.4 trillion tokens per day, about 22 times the volume a year earlier. Total deployed compute reached 48,000 petaflops. The factory metaphor extends to economics — by jointly optimizing models and infrastructure, SenseTime says it tripled training speed for long-context agent models, raised image and video generation efficiency sixfold, and cut the cycle from model validation to large-scale deployment to one week.
The third pillar, the "Agent Harness," is where the company expects value to migrate from raw token pricing toward paying for completed tasks. SenseTime's agent products reached more than 1,000 enterprise customers across over 20 industries, and its office agent "Xiaohuanxiong" (小浣熊) entered firms including Lenovo and the three major telecom operators, with cloud monthly active users up close to tenfold.
Open models as a funnel
SenseTime has also been pushing its SenseNova (日日新) models into the open-source world. The U1 series, built on a native unified architecture the company calls NEO-Unify that merges understanding and generation in a single model, was open-sourced in April 2026. Within four months the three related projects collected more than 11,000 stars on GitHub. A lighter office-focused model, SenseNova 6.8-Flash-Lite, and a spatial-intelligence model, SenseNova SI-8B, followed, alongside a world model, Kairos 3.1, co-developed with Ace Robotics (大晓机器人).
Open release here is not charity. It is a funnel: developers who adopt the weights are more likely to consume SenseTime's token services and cloud, where the company actually earns. The generative-AI revenue share nearing 80% suggests the funnel is already doing real work.
Honest limitations
The profit figure is IFRS-based and includes fair-value gains on invested companies under SenseTime's "1+X" strategy; on the operating level the company remains loss-making on a non-IFRS adjusted basis. Revenue, margin and token-volume figures come from the company's interim report and its official results announcement, corroborated by China Securities and Xinhua Finance; they have not been independently audited beyond the filing. The "22x" token-growth and "48,000 petaflops" numbers are company disclosures and should be treated as stated, not benchmarked. Competitive benchmark claims that SenseNova models "matched or outperformed" Gemini, GPT and DeepSeek are the company's own and are not independently confirmed here. I have not run the models myself.
What readers can do now
- If you buy or evaluate enterprise AI in China, treat SenseTime's recurring-revenue share as a signal of how much of its business is now subscription-like rather than project-based.
- If you build agents, test the open-sourced SenseNova U1 series on GitHub and compare its token economics against closed APIs for your workload.
- Watch the July 2026 token-volume and compute figures in future filings — sustained growth there is the real test of whether the "Token Factory" thesis converts to durable profit.