Dobot doubles H1 revenue as its robot-AI orders stack up
Cobots are quietly becoming an AI business. Dobot (Yuejiang Technology, HKEX 02432) reported first-half revenue of RMB 320 million, up 106.6% year over year — and the number that matters most is a single line item: embodied-intelligence robot sales jumped more than twenty-fold to over RMB 45 million, roughly 14% of everything the company sold.
The Shenzhen cobot maker's earnings read like proof that Physical AI is turning from demo reels into invoices. The company says its embodied-AI unit has signed 231 customers and holds a backlog of more than RMB 60 million in orders, including deployments for Valeo, Leapmotor auto parts supplier chains, and a European cosmetics group; nearly 100 of those customers are industrial manufacturers. The commercial proof point getting the most attention: an Atom humanoid running unassisted at a Shenzhen cinema, working 14-hour shifts and selling over 1,000 cups of popcorn a day. Dobot attributes the jump to its in-house world action model, DobotWAM, which it claims tops benchmarks like LIBERO (99.25% average completion) ahead of Nvidia's GR00T — vendor numbers, worth taking with a grain of salt until independently rerun.
The catch sits lower in the income statement: growth is being bought with red ink. The net loss attributable to shareholders widened about 159% year over year to roughly RMB 106 million as R&D spending surged 148% to about RMB 100 million — the company is spending on embodied AI faster than the new business pays for itself. That is the trade almost every player in this space is making right now; Dobot is just one of the first Chinese robot makers large enough to show it in audited numbers rather than fundraising announcements. The strategic bet is that its "one brain, many bodies" architecture — one model driving arms, humanoids, wheeled and quadruped platforms — turns a decade of motion-control data from real factory floors into the scarce asset everyone else is missing.
We've tracked this capital flood all month — physical-AI companies raised $47.4 billion in the first half alone, per Physical AI raises $47.4B in H1, topping 2022-2024. What makes Dobot different is that the money is arriving as customer orders, not term sheets. When a robotics P&L starts showing an AI line item growing 20x, the question stops being "does anyone want this" and becomes "who gets to unit economics first."
China's AI systems now burn nearly 175 trillion tokens a day
China's National Data Administration says daily token consumption across the country's AI models hit nearly 175 trillion in June — roughly 125,000 tokens per person per day, counting every citizen. The figure, highlighted by CCTV Finance, cements China's position as the world's largest inference market and marks another step in a vertigo-inducing curve: daily usage was around 140 trillion in March, up more than 40% from end-2025, and officials say volume has grown over a thousand-fold in two years.
The scale explains both the confidence and the anxiety in China's AI industry. On one hand, token flow has become a headline metric in its own right — state media now tracks weekly model-usage rankings the way markets track trading volume, and "token economy" has entered official policy language. On the other, every query costs real compute: the same CCTV report notes model providers are still losing money on usage because inference costs scale with demand, even as chip iterations and cheaper compute keep pushing unit costs down. The bet is that whoever aggregates the most users wins once inference gets cheap enough — which is why DeepSeek, ByteDance and Alibaba are locked in a price war for call volume they currently lose money on.
The counterpoint worth keeping in mind: usage leadership isn't capability leadership. Analysts have repeatedly argued that raw call volume reflects China's population-scale consumer apps and subsidized pricing more than frontier-model quality. But as a demand signal for chips, power and data-center build-out, 175 trillion tokens a day speaks for itself.
What to watch: whether the next official reading keeps compounding at this pace, or plateaus once subsidy-driven pricing normalizes.
If your own AI usage is any guide, how far can this curve climb before inference costs stop falling? Tell us in the comments.
