Zhiyuan Robot has spun out its data business for standalone financing, with China Telecom taking over as lead investor — the third round in six months.
On August 17, the one-stop physical AI data service platform Mifeng Technology announced completion of a new funding round worth several hundred million RMB. Though only founded in February 2026, the company has closed three rounds in half a year: a seed + angel round led by Sequoia China, an angel+ round led by Guofang Venture Capital in June, and now this round led by China Telecom with Zhangjiang Group co-investing. Sequoia and Yuanqi Innovation over-subscribed as existing shareholders.
The money came fast, but the story behind it deserves a closer look. Mifeng did not appear out of nowhere — it is the data business entity carved out of Zhiyuan Robot's late-2025 "one splits into four" strategy. Zhiyuan split four business lines — dexterous hands, data, leasing, and quadruped robots — into four companies: Lingjiedian, Mifeng Technology, Qingtianzu, and Zhiyuan Kutuo. Qichacha records show Zhiyuan Innovation (Shanghai) Technology Co., Ltd. remains Mifeng's largest shareholder.
Why split out data? Mifeng Chairman and CEO Yao Maqing put it bluntly: "The team judged that what will truly constrain industry development and form moats in the future is data — embodied intelligence models themselves will not have strong moats." The numbers make this stark: the training corpus for the LLM GPT-5 amounts to roughly 10 billion hours, while robots cannot learn directly from the internet, and the entire industry's aggregate high-quality embodied data is only around 500,000 hours — a mismatch of more than an order of magnitude.
Mifeng's business model is built on this "data desert," laid out across three ends:
Collection — Its self-developed MEgo series body-free collection terminals have entered mass production and delivery. The hardware does not depend on a robot body; lightweight portable devices reach real scenarios in factories, logistics, supermarkets, homes, hotels, restaurants, security, elder care, and tourism, synchronously capturing macro environmental data and micro hand-manipulation data. The product line covers full categories including real-robot teleoperation and body-free data collection, forming a layered data supply pyramid.
Governance — The MEgo Engine data governance platform connects preprocessing, spatial reconstruction, multimodal annotation, and quality evaluation into a complete automated loop. Self-developed VIO high-precision pose capture algorithms, combined with human-to-robot skill transfer, long-horizon motion auto-segmentation, and atomic-operation intelligent annotation, improve traditional manual data processing efficiency by more than 10x.
Evaluation — Facing closed-loop iteration, Mifeng builds evaluation systems and deploys a data flywheel, linking task execution, data return, case mining, and data updates into an automated loop. The company says that in some application scenario tests, each data loop iteration pushes task success rates up 5% to 10%.
Mifeng plans to reach tens of millions of hours of data capacity within the year. It has established innovation centers in over 20 Chinese cities and laid out more than 5 overseas nodes; its client roster includes embodied foundation models such as Ant Lingbo's LingBot-VLA 2.0.
China Telecom leading this round carries significance beyond the money itself. Horizontally, all three major telecom operators have now placed bets on embodied data: China Telecom jointly invested in Dimo Robotics with Inovance and set up a 2-billion-RMB AI industry fund; China Mobile's chain-leader fund also invested in Dimo; China Unicom built its own embodied intelligence pilot-testing base and accumulated an 80TB dataset. Operators view embodied data as the next-generation infrastructure entry point after "cloud-network + computing power," making the positioning value of platform-type data suppliers evident.
This path overlaps with Yuanli Wuxian's "Ego data route + AtomBrain" direction, but Mifeng differentiates by not betting on a single robot body: MEgo body-free terminals can be deployed independently, making the collection path lighter and cheaper — closer to "the China Mobile of the data era." By splitting data from an embodied brain's built-in capability into a platform business, Mifeng takes the To B service provider route rather than the To B OEM route.
Data has been underestimated repeatedly over the past 12 months. Model weights can be fine-tuned, distilled, or re-architected, but physical-world data can only be collected on site. When China Telecom is willing to pay for "tens of millions of hours of data capacity," it means the decisive battlefield of the embodied AI race is sinking from the model layer to the data layer.
Next validation points: whether Mifeng can truly exceed tens of millions of hours of data capacity within the year; whether MEgo terminals' ROI in real factories and retail can pay back for operators; and whether the other three Zhiyuan spin-offs (Lingjiedian dexterous hands, Qingtianzu leasing, Zhiyuan Kutuo quadrupeds) each complete independent financing. Only when all four land does the Zhiyuan ecosystem fully run through.