Executive Summary
In the first half of 2026, China's embodied-intelligence sector closed 322 financing rounds worth more than RMB 90 billion (roughly USD 12.5 billion at an approximate rate of 7.2), with a new company crossing the $1 billion "unicorn" valuation line roughly once every ten days. The figures come from the "2026 China Embodied Intelligence Industry Insight" report presented by Zhang Yijia, Founder and CEO of ChinaVenture (Jazzyear), at the MaQiao Embodied Intelligence Industrial Ecosystem Forum during WAIC 2026. But behind the capital inflow, a growing number of industry figures at WAIC spoke openly about bubble dynamics, unforgiving customer return-on-investment demands, and the gap between building a robot and selling one. This article draws on first-hand speaker notes from more than a dozen WAIC 2026 forums to lay out the financing scale, unit economics (ROI, payback periods), industry stratification, and the power, data, and reliability bottlenecks that stand between demos and deployment.
Industry Context
Embodied intelligence was one of the clearest focal points of this year's WAIC. According to Wang Bo, Chief Content Officer of Jazzyear, speaking at the MaQiao forum, 242 robotics, embodied-intelligence and smart-hardware exhibitors — 23.6% of all exhibitors — appeared at the conference. Wang Tianmiao, Honorary Director of the Beihang University Robotics Institute, estimated more than 300 humanoid robot models were on display across the show floor.
The financing figures were equally dense. Alex Zhou, Managing Partner at Qiming Venture Partners, told the Qiming Venture Partners Entrepreneurship and Investment Forum that embodied intelligence has become the "second major consensus" formed in capital markets over the past 12 months (the first being AI agents), and that visiting China's embodied-intelligence robotics industry is now the top priority for Fortune 500 companies and top-tier funds touring the country. Roughly 350 companies have emerged in this track in China over the past two to three years, he said, with 10 already valued above RMB 20 billion and roughly 20 more around RMB 10 billion.
Wang Tianmiao's macro figures at the MaQiao forum add further texture: of the more than RMB 90 billion raised in H1, about RMB 30 billion flowed into large foundation models. The market broadly expects the embodied-intelligence sector to reach a scale of more than RMB 10 trillion by 2035, with an industry-wide consensus pointing to a "peak moment" between 2028 and 2030, and a "Physical Turing Test" — an assessment framework spanning cross-task, cross-environment and cross-disturbance usability and safety — expected to gain academic and industrial consensus around 2030.
Technology
Two engineering threads — data and compute — underpin this industrialization narrative, but several researchers at WAIC 2026 flagged unresolved constraints beneath the surface.
At the HKUST International AI Industry Innovation Ecosystem Forum, Guo Yandong, Founder and CEO of AgiBot Rival firm Zhiping Fang (AI2Robotics), offered what he called a "blunt viewpoint": a major factor preventing embodied large models from reaching their "GPT moment" may be power supply. He explained that the training power required for embodied models could be many times that of language models, and that the way forward is a brain-inspired computing architecture that sharply cuts training energy consumption and enables "learn while working" deployment at the edge. He also noted that robots commonly have more than 30 joint degrees of freedom, compared with a car's strictly two, which makes embodied control substantially more complex than autonomous driving.
At the same forum, Wang Jun, Professor of Computer Science at University College London, raised a pointed challenge to the currently popular reinforcement-learning narrative: embodied-AI companies "aren't actually doing reinforcement learning on real robots" — most rely on distillation instead, and "it can theoretically be proven that distillation methods cannot generalize." Compressing physical-world feedback into a few iterations to train a 70-billion-plus-parameter model, he argued, is "almost Mission Impossible," and the field needs entirely new algorithms to address sample complexity.
Data cost is the other major thread. Qiming's Zhou put the comparison plainly: collecting a unit of embodied-AI data in China costs roughly one-seventh of doing so in the United States. Zhang Wei, Founder of LimX Dynamics, told the Guotai Haitong AI Investment & Financing Forum that the technical paradigm has already converged on pure data-driven approaches, and what remains is "lowering the data cost of completing a task"; he assessed the embodied "brain system" as roughly at a "GPT-3 stage" today. Wang Tianmiao's figures show that simulation-transfer training data needs to scale from today's 300,000–500,000 hours toward nearly 10 million hours, at a unit cost of roughly RMB 300–500 per hour.
Engineering Analysis
If financing data captures capital's enthusiasm, reliability data captures engineering reality.
Shao Tianlan, CEO and Founder of Mech-Mind Robotics, laid out a product-maturity model at the WAIC Intelligent Trends Forum (industrial AI forum): most embodied-AI demos remain at maturity levels 0–3, and a product must reach level 8 to be a real revenue-generating business, level 9 to be profitable. "Don't believe anything you see online about robots," he said. "Offline demonstrations are far more convincing." He also pointed out that global annual shipments of industrial-plus-collaborative robots total only about 500,000 units, against roughly 1 billion manufacturing and logistics jobs worldwide — a gap he attributes not to the "embodied intelligence" concept itself, but to the high cost, long cycle time and lack of flexibility of traditional custom-integration automation.
Cao Yunan, Chairman and CEO of Elite Robotics, was even more direct on the same panel. Mapping physical AI's paradigm evolution onto autonomous-driving's L-level classification, he admitted the industry today is "only at L2, not even L3." He also described a "negative flywheel" at work industry-wide: insufficient physical data leads to poor model generalization, which leads to frequent execution failures, while the 99% success rates companies routinely claim have no independent verification — "Who evaluates that? Is there any genuine third-party fair standard? No."
That assessment was corroborated by benchmark data from Tang Wenbin, Founder and CEO of Yuanli Lingji (a Megvii co-founder), presented at the Qiming forum via the Robot Challenge evaluation platform: even single-task models achieve real-world success rates of only 64%, far below the 99% figures vendors advertise, and one well-funded company had its results disqualified after being caught gaming the evaluation platform. Tang's conclusion: "It's actually very hard for us to say that it's actually a ChatGPT moment. I'm sorry, we haven't got this moment yet."
Commercial Progress
Despite these engineering gaps, WAIC 2026 produced a number of deployments with concrete commercial figures attached.
Zhang Jianzheng, Chairman and CEO of Saizhi Intelligent, disclosed that his company is now a direct supplier to Apple, with hundreds of robots deployed in the final-cover inspection step of iPhone assembly — a task performed by roughly 350,000 workers globally, implying a potential robot demand of about 200,000 units if that workforce were compressed to 100,000. His unit-economics figure: a worker previously needed 2.5 hours to complete an electrical-cabinet inspection task at a 90% success rate; his company's six-arm robot now completes it in six minutes at 99.9%. The case was included in last year's World Economic Forum Global Lighthouse Factory whitepaper.
Li Mingyang, Chairman and CEO of JAKA Robotics, offered the sharpest customer-side numbers at the industry's flagship panel discussion: industrial customers generally expect payback on embodied robots within 24 months, with the most aggressive demanding 10 to 12 months — "very brutal numbers," he said, calling ROI "forever the core" of the buying decision.
Supply-chain costs are falling fast as well. Guo Renjie, Founder and CEO of Origin Robotics (Loona parent Livox Tech's consumer arm), said joint-module prices have fallen from several thousand yuan to RMB 200–300 over the past year, making sub-RMB-10,000 home robots possible; he added that household deployment requires both non-invasiveness (safe for children, the elderly and pets) and full autonomy (self-localization, mapping, obstacle avoidance and self-charging). You Wei, Chairman of Efort and Qizhi Robotics, cautioned from the manufacturing side that his company already ships 20,000 industrial robots a year with more than 100,000 delivered cumulatively — "mass production is not hard" — arguing the real bottleneck lies in "mass sales" and reliably scaling complex skills.
Chip and compute economics were equally concrete. Yu Kai, Founder and CEO of Horizon Robotics, told the AI-Enabled Ocean Forum that its subsidiary D-Robotics already supplies compute chips to more than 20 domestic embodied/humanoid robot companies, making it one of two dominant chip suppliers in the space alongside NVIDIA; Horizon's ADAS solutions have shipped roughly 14 million units cumulatively, the top share in China. Payback data from the industrial side was similarly tangible: Liu Zhen, President of the Digital Building Materials Research Institute, told the industrial AI forum that his industrial foundation model's payback period at customer sites is "just a few months, sometimes just a few weeks"; Han Jingrun, Senior Vice President at iSoftStone, offered a reference point — a manufacturing enterprise with roughly RMB 10 billion in output typically invests about RMB 80 million in first-phase AI infrastructure.
Market Perspective
Investment institutions are stratified in their read on this boom. Qiming's Zhou argues China holds structural advantages: a full-size humanoid robot over 1.7 meters tall has roughly 1,200 components, of which all but about 10 are sourced from the Yangtze River Delta and Pearl River Delta industrial clusters. But he also stressed that the industry-wide scaling law remains unvalidated, with training-data scale still the bottleneck; the top roughly 20 companies in China and the US are positioned to reach "million-hour" scale effective data this year.
Wang Bin, Partner at Oriental Fortune Capital, gave an even sharper China-US comparison at the Guotai Haitong forum: the US has fewer than 10 humanoid robot companies, while China has 100–200. He cautioned, though, that true physical AGI — top-tier mobility such as free climbing — remains far off; this generation of physical AI mainly addresses basic manual labor, and true generality still requires data accumulation at the tens-of-millions-of-hours scale.
Wang Tianmiao's framing leaned toward stratification within the industry: "The bubble exists. But if you are one of the leaders, the chain master of a vertical, there is no bubble for you — because you are the one stirring the wave, riding on it." He divided the investment landscape into three types of "unicorns" — full-stack ecosystem leaders, undervalued vertical leaders (in high-fault-tolerance-cost fields such as polishing, precision machining, mining, semiconductor processing and surgery), and providers of ecosystem tools, chips and models — plus four categories of "early-positioned hidden champions" in materials, edge chips, dexterous hands, and force control. His advice to founders: treat cash flow as a matter of survival instinct, shelve IPO ambitions, and focus on the shortest, smallest path to real cash flow.
Gao Hongbin, Dean of the Digital Economy Research Institute at Shanghai University of Finance and Economics, framed embodied intelligence macro-economically at the Intelligent Economy Development Ecosystem Forum, describing it as a second "battlefield" alongside open-source large-language models, and citing Unitree and AgiBot as examples of Chinese companies with a particular edge in this track. University President Liu Yuanchun, on the same panel, grouped robotics with AI, semiconductor chips and innovative drugs as the "emerging three" now driving China's export growth and industrial upgrading — a deliberate echo of the earlier "new three" of EVs, lithium batteries and solar panels.
Challenges
The hard commercialization constraints raised repeatedly at WAIC 2026 fall into four categories.
The first is a reliability gap. Elite Robotics' Cao described the "negative flywheel" (insufficient data → poor generalization → frequent failures → customer reluctance → even less data), while Tang Wenbin's benchmark-gaming disclosure underscores the lack of a credible third-party success-rate evaluation regime. You Wei of Efort framed this as a "skill iceberg": today's demos are mostly pick-and-place skills above the waterline, while real factory floors demand multi-robot interaction and tight process coupling below it — "unless the skills beneath the iceberg are solved, the industry bottleneck is clear."
The second is the absence of cross-embodiment data reuse. Zhang Yingying, Deputy Director of the Technology Innovation Department at China Logistics Group, said bluntly at the logistics forum that no company has yet achieved trainable, reusable data across different robot embodiments; robots today "can dance, can perform some retail-style services," but remain far from actually "entering the factory to work."
The third is a physical constraint on power and compute architecture. Guo Yandong's power-bottleneck thesis, echoed by Xing Xun, Strategic Advisor at TTVision at the same forum — "it is hard to imagine that future embodied systems would carry a hot, power-hungry central domain controller… that simply violates physical principles" — points to energy consumption and edge compute as engineering problems that must be solved before embodied AI can achieve language-model-scale emergence.
The fourth is customer-side economic pressure. JAKA's Li Mingyang's disclosure of a 10-to-12-month best-case payback demand shows the commercialization bar for embodied robots is being pulled up to match mature industrial and collaborative arms — "customers' payback expectations for embodied robots are equated with six-axis industrial or collaborative arms, with no special tolerance." Storytelling and funding rounds alone cannot pass a customer's financial review; the numbers have to pencil out.
RobotToday Analysis
Further reading: For the full conceptual taxonomy behind this piece — the Physical AI macro paradigm, Embodied AI methodology, and the VLM–World Model–VLA pipeline — see RobotToday's companion framework piece, Physical AI Landscape: From Digital Intelligence to the Embodied Physical World.
Technically, the embodied-intelligence sector on display at WAIC 2026 sits at a pivotal transition — from proving robots can move to proving they can work reliably and profitably. The power-consumption, sample-complexity and cross-embodiment data-reuse constraints flagged respectively by Guo Yandong, Wang Jun and Zhang Yingying will likely shape the industry's real growth rate over the next two to three years more than any single model-architecture debate.
Commercially, the more than RMB 90 billion raised in H1 2026 and the "one unicorn every ten days" pace already signal a track richly priced by capital. But JAKA's 10-to-12-month payback ceiling, Elite Robotics' self-assessment of "not even L3," and Tang Wenbin's 64%-not-99% success-rate data together sketch a widening gap between capital's expectations and engineering reality. Wang Tianmiao's and You Wei's shared refrain — mass production is not hard, mass sales is — is the single judgment from this wave of industrialization that both engineering and business readers should hold onto.
The remaining challenges are clear: no credible third-party reliability standard, cross-embodiment data still siloed, unresolved physical constraints on power and edge compute, and shrinking tolerance on customer payback timelines. The 2028–2030 "peak moment" that several speakers projected is not baseless, but Wang Tianmiao's own observation — that the bubble arises "because technological iterations are nonlinear" — suggests that consolidation at the top and a shakeout in the middle are the more likely story over the next one to two years, rather than a rising tide that lifts every participant.
Quotes and figures in this article are drawn from conference public remarks at WAIC 2026 and have not been individually verified with the speakers or their organizations. Where a speaker's name could not be confirmed from the source transcript, this is noted in the text. Corrections are welcome.
WAIC Review 1/4: Embodied AI's Industrialization Wave: Financing Data, Unit Economics, and a Widening Split
WAIC Review 2/4: World Models vs. VLA: Robotics' Next 'ChatGPT Moment'
Leave a comment