Alibaba-Backed Dexmal Targets $3 Billion Valuation in Embodied AI Funding Push

Two days before the weekend, a quiet filing of sorts rippled through China’s robotics scene: Dexmal, a Beijing-based startup building embodied artificial intelligence for robots, is in talks to raise fresh funding at a valuation of about 20 billion yuan — roughly $3 billion — according to a Bloomberg report published Friday, August 21. The talks are ongoing, company founder Tang Wenbin confirmed to the outlet, and previous rounds were led by e-commerce and AI giant Alibaba Group.

On the surface, it is another eye-popping number in a year that has already produced a $21 billion Etched and a record-setting Anthropic IPO. But Dexmal’s bid matters for a different reason: it is one of the clearest signals yet that China’s embodied-AI sector — the discipline of giving large models arms, legs, and grippers — is transitioning from demo videos to real industrial data, and that investors are pricing that transition accordingly.

What Dexmal Actually Builds

Dexmal is not a humanoid-robot manufacturer in the Unitree mold. The company develops end-to-end embodied-intelligence software and hardware, anchored by a foundation model designed to control physical robots across tasks. Its portfolio, laid out in materials from its June merger with warehouse-automation firm ATOMIX, spans four layers:

• DM0 — billed as the world’s first “embodied-native” foundation model, trained on a hybrid dataset combining robotic multi-perception data, autonomous-driving data, and internet-scale data. The company claims sub-millimeter operation precision and a No. 1 global ranking in real-robot evaluation.
• Dexbotic — an open-source general-purpose embodied-AI framework, described as “the PyTorch of the embodied AI era” and the second such framework globally. It claims adoption by dozens of academic institutions including Tsinghua, Peking University, and Princeton.
• RoboChallenge — a real-robot evaluation platform co-launched with Hugging Face, claiming more than 80,000 cumulative real-robot tests and use in CVPR- and ICRA-tier competitions.
• Ferrata — unveiled June 15 as the world’s first three-level sorting system for warehousing and logistics, with DM0 as its orchestration brain coordinating heterogeneous robot fleets.

Why a Warehouse Company Merged With a Model Company

The ATOMIX merger, completed in June with strategic financing, is the strategic heart of the story. ATOMIX is an AI-native flexible warehousing automation provider headquartered in Singapore, with a 4-way pallet shuttle product that ranks No. 2 in global sales and No. 1 in mega-scale deployments of more than 50 shuttles per system. Its client list reads like a retail who’s-who: Uniqlo, Coca-Cola, Yum China, Mixue Bingcheng, and Lenovo, across more than 20 countries.

The crucial number is 600,000 — the items ATOMIX ships daily. That volume makes it, by its own description, “the largest and most authentic logistics-picking data source in the embodied AI landscape.”

Dexmal’s thesis is explicit: just as coding was the watershed atomic task for language models — whoever cracked it first unlocked general capabilities — picking will be the atomic task of embodied AI. It offers large-scale real-world data, verifiable success signals (an item is either picked correctly or it isn’t), and cross-task transferability. Merging with ATOMIX gives the model company a perpetual data flywheel; the merger materials call it “the first time in the embodied AI arena that models and scenarios have truly grown together.”

The financing that accompanied the merger underscores how seriously China’s AI establishment takes this. Zhipu AI, StepFun, and SenseTime — three of the country’s leading foundation-model players — all invested, alongside Alibaba, industry partners Huaqin and SAIC Hengxu, and a long roster of funds including Jinpu Capital, Oriza Puhua, Suzhou Fund, and CMB International.

The China Context: A $26 Billion Push

Dexmal’s $3 billion ask does not exist in a vacuum. As PYMNTS noted in its coverage, China has more than 140 companies in the humanoid robotics space producing robots at scale for factories, hotels, and offices. The government has stated an ambition to lead embodied AI within five years, and Beijing alone has set up investment funds exceeding $26 billion for the industry. State-owned enterprises are deploying robots in museums, at events, and — in some cases — directing traffic.

Against that backdrop, a $3 billion valuation for a company that already holds a top-ranked foundation model, an open-source framework with thousands of developers, and a live industrial data pipeline starts to look less like bubble pricing and more like consolidation logic. The company is also moving fast on cadence: it has teased a next-generation “foundation model 5,” its first universal robot, and new application infrastructure for imminent global release.

Western counterparts are raising too — Travis Kalanick’s Atoms pulled in $1.7 billion in July for industrial physical AI, and the UK’s Humanoid closed a $152 million Series A to become “Europe’s first pure-play humanoid robotics unicorn.” But the Dexmal round is structurally different: rather than betting on a future humanoid workforce, it monetizes robots already running in production warehouses today.

The Skeptic’s Read

Three risks deserve mention. First, the valuation is a target in an ongoing round — terms can and do move. Second, “No. 1 in real-robot evaluation” and “largest data source” are company claims; independent benchmarking of embodied models remains immature. Third, the picking-as-atomic-task thesis is unproven — warehouse manipulation is narrower than general dexterity, and success there does not guarantee transfer to homes or hospitals.

Still, the direction of travel is clear. The first half of embodied AI was about demos; the second half is about data. Dexmal, with Alibaba’s backing and a warehouse network feeding its models around the clock, is placing one of the sector’s most credible bets on that transition — and investors may be about to stamp it at $3 billion.