China's Embodied AI Funding Surge: Humanoid Robot Commercialization Faces Three Critical Tests
0xBroomberg
In the first five months of 2026, China's embodied-AI sector closed over 230 deals totalling more than RMB 57 billion — a pace rivalling early-stage EV funding; yet cheaper hardware is only the entry ticket, and model reliability plus ecosystem control are the real thresholds that will decide survivors.
How fast is the money pouring in?
Five months: 230-plus deals, RMB 57 billion. Morgan Stanley projects the global humanoid-robot market will reach $7.5 trillion by 2050, with China accounting for 30% of adoption.
Deal sizes are equally striking: Spirit AI closed four rounds in three months, raising roughly RMB 5 billion at a valuation above RMB 20 billion; Astribot closed three rounds in three months, topping RMB 1 billion raised at a valuation past RMB 10 billion.
This means → capital is flooding in at an "EV-era" pace, betting on the next trillion-dollar hardware platform.
What signal does the IPO pipeline send?
Unitree Robotics, valued at RMB 42 billion, has cleared A-share listing review — described as the fastest A-share IPO on record. Deep Robotics (valued at RMB 13.9 billion) targets the STAR Market; Leju Robotics (RMB 4.3 billion) targets ChiNext.
Per 36Kr, nearly 50 robotics companies are queuing for a Hong Kong listing, making up over 10% of total IPO applications.
This means → primary-market capital is no longer sufficient; companies are rushing to lock valuations into public markets — a confidence signal and a bubble warning in one.
How far has hardware cost fallen?
From 2023 to 2026, Chinese-made harmonic reducers — the core gear assemblies inside robot joints — dropped from RMB 3,000–5,000 to RMB 1,500–2,000 per unit. Servo motors fell over 40%; six-axis force sensors dropped from above RMB 5,000 to RMB 1,500–2,000. Overall core-component costs fell more than 50%, saving RMB 100,000–200,000 per unit.
Unitree builds its own motors, reducers, and dexterous hands: the Go2 Air quadruped sells below RMB 10,000; the G1 humanoid starts at RMB 85,000. Compare that with Boston Dynamics' Spot at roughly RMB 500,000 and Figure 02 at roughly RMB 420,000.
In plain terms = China's supply chain has made the robot's "body" cheap — so cheap that over 70% of Tesla Optimus parts reportedly come from Chinese suppliers. But a cheap body is not the same as a product that can do useful work.
Where is the real commercialisation bottleneck?
Taiwan supply-chain executives flag systemic hurdles in structural stability, heat dissipation, durability, and gait-control precision — limited cooling space in joints and insufficient motion accuracy degrade perception and execution performance.
Return-on-investment timelines for home, retail, and general-service robots remain long; buyers are cautious.
This reflects a core contradiction: the supply chain is ready to "mass-produce," but downstream use cases are not ready to "mass-buy."
Is AI model capability catching up?
Spirit AI's Spirit v1.6 embodied foundation model ranked first on RoboArena as of June 3, surpassing NVIDIA's Cosmos3 and Physical Intelligence's Pi0.5. CEO Han Fengtao said the latest funding will be split roughly equally among model training, data reserves, and talent — targeting one million hours of real interaction data in 2026 and a 95% task-success rate after one hour of fine-tuning.
Alibaba's Amap has launched embodied-robot training data and maps; ByteDance's Seed unit has designated embodied AI a core direction.
Yet a clear gap remains between benchmark scores and industrial reliability — Spirit AI's collaborations with CATL, JD.com, and Bosch are still at pilot stage, with no large-scale repeat orders. This means → leading on a leaderboard does not mean customers are ready to pay.
What does the ecosystem-control battle point to?
Unitree and NVIDIA jointly launched the H2 Plus humanoid, running on the Jetson Thor compute platform and Isaac GR00T framework. Unitree marketing director Huang Jiawei called the core upgrade "compute performance"; founder Wang Xingxing called it a starting point for validating real robot skills.
This reflects a deeper division-of-labour question: data ownership, software control layers, chip-supply dependency, and brand positioning — if these four boundaries are not clarified, maintenance, upgrades, and trust all become liabilities. Unitree's IPO plans to raise RMB 4.2 billion, with nearly half earmarked for intelligent-robot model development — a pivot from hardware seller to integrated hardware-software platform.
Per Omdia 2025 data, China humanoid-robot shipments: Unitree 5,500 units, AgiBot 5,168 units, UBTECH 1,000 units. In plain terms = from "can move" to "can work," the entire industry still needs to clear its most critical proof point.
Content is for reference only, not financial advice.