SenseTime posts first-ever profit as generative AI hits 80% of revenue
The day's clearest signal that China's AI companies can actually turn a profit lands from Hong Kong, where long-unprofitable SenseTime just posted its first-ever half-year profit — and it's not a cost story, it's a generative-AI revenue one.
SenseTime posted its first profit since its 2021 IPO, reporting ¥620 million in net income for the first half of 2026 under IFRS. Revenue climbed 23.4% year over year to ¥2.91 billion, gross profit hit ¥1.21 billion at a 41.4% margin, and the company disclosed recurring revenue of ¥1.14 billion for the first time — 39.3% of the total. The headline figure is the mix: generative AI contributed ¥2.33 billion, close to 80% of all revenue, while overseas business surged 127% year over year, far outpacing the group's overall growth.
The profitability matters less for the number itself than for what it implies about the pivot. SenseTime frames its strategy as "a set of models, a Token factory, and one intelligent agent control system" — in plain terms, it stopped selling one-off vision projects and moved to metered model-token usage plus agentic sales, the same recurring-revenue playbook Western AI labs are chasing. That it reached this milestone while generative AI is still nearly eight-tenths of revenue suggests the model-rental economics underpinning the current AI boom are real enough to sustain a company, not just a valuation. Recent context points the same way: rivals and fast-followers are all converging on selling tokens and agents instead of bespoke deployments. SenseTime's maiden profit is early proof that model as a service, not project consulting, is what makes the money in China too.
Google and Microsoft are proposing a web standard called WebMCP that would let a website declare exactly what an AI agent can do on it — registering structured tools an agent can call directly, instead of forcing the agent to reverse-engineer a page by reading pixels and guessing at buttons. It's live today as a Chrome origin trial, and the pitch is that a site keeps a stable interface for machines that survives redesigns, runs inside the user's own logged-in session, and can require a human confirmation for consequential actions like checkout. The underlying idea extends MCP from the server to the page itself, and early demos show an agent chaining read-only tools freely while stopping for approval before anything real happens. It's early — Chrome-first and subject to change — but it's a concrete step toward a web that is built to be operated by agents cleanly rather than scraped.
Volvo is rolling out connected hazard alerts across three of its electric vehicles — the EX90, ES90, and EX60 — so one car can warn another about large animals, pedestrians and cyclists, roadwork, or accidents ahead. The system uses sensor and camera data from a fleet of about a million connected Volvos: when a vehicle spots a hazard, it logs the location to Volvo's servers, which then push an alert to other Volvos traveling along the same road. Unlike crowdsourced apps like Google Maps or Waze, it relies on the automaker's own fleet, and Volvo says it shares that data with traffic-management centers and other vehicle brands. It's a modest consumer feature, but it doubles as a working example of vehicle-to-vehicle AI coordination in production — cars teaching each other what their cameras just saw.
If a site declared its own agent tools instead of being scraped, would you trust a checkout that runs in your own logged-in session? Tell us in the comments.
Sources: GMT Eight · 上观新闻 (Shanghai Observer) · The BambooWorks · WebMCP (Chrome docs) · WebMCP explainer — Sreenath M Menon · The Verge
