China's Embodied-AI Data Collection Business: ¥500/Hour Teleop Data, 100+ Training Centers, and a "Buy the Robot, Buy Back the Data" Debate

Why data became the bottleneck

Large language models got lucky: centuries of text and billions of web pages gave them roughly 300 trillion tokens of ready-made training material. Robots cannot use any of it. A robot needs multimodal interaction data — task instructions, environmental observations, joint states, force and touch feedback, and execution results — recorded in sequence, in the physical world, through the machine's own body. That data was never systematically captured before, and it cannot be scraped.

The shortfall is enormous. Guojin Securities estimates a mature embodied foundation model needs tens of millions of hours of quality data (and up to hundreds of millions for general intelligence), while the industry currently has only "hundreds of thousands to one million hours" — a supply gap of about 90%. A CAICT-linked estimate puts it even more concretely: to train 1.1 trillion tokens entirely from real-robot collection would require about 21,200 robots working for a full year to produce million-hour-class effective data. Zheshang Securities calculates that as of early 2026, globally available compliant real physical-interaction data is 20× short of what one general embodied model needs. And the payoff is measurable: overseas research cited by CAICT found that roughly doubling pretraining data scale lifted task success rates by about 1.5×.

The four data paths — and what each costs

The 2026 Embodied Data Development Report (China Information Industry Association, CESI Institute 5, Data Element Society) maps the industry's four main collection routes:

Source: Embodied Data Development Report 2026 via The Paper (澎湃新闻); CAICT Embodied Intelligence Training Grounds Research Report (2026) via Jiemian. Per-hour figures are market quotes for commissioned collection.

Industry analysts frame the structural problem as "three absences": no unified operating system ("pre-OS era"), non-standardized hardware and software (one embodiment, one dataset; cross-embodiment adaptation costs 3–5×), and unconverged technical routes (VLA vs world models vs VLM+control), producing data that is "non-standard, expensive, hard to reuse."

Who is building the collection industry

By June 2026 China had 70+ operating training grounds with 46 more built or planned across 18 provincial regions (CAICT); by August, media tallies counted 106 built and 84 operating. Three clusters lead: the Yangtze Delta (Zhejiang alone has 14), Beijing-Tianjin-Hebei and the Pearl River Delta. The Yangtze Delta cluster — Shanghai's Qilin training ground, Zhiyuan's data factory, Galaxea's Suzhou center, Hangzhou's humanoid pilot base and Hefei's pre-training ground — covers over 70,000 m².

• The "Guodi" national network (10 sites): the Beijing Humanoid Robot Innovation Center leads, anchored by Shanghai's Qilin main training ground with 9 branch grounds (Beijing, Henan, Shenzhen, Hebei, Chongqing, Jilin, Yulin in Shaanxi, Changshu and Wujiang in Jiangsu). Beijing's site runs a ~6,000 m² "training school" with mock homes, supermarkets, industrial assembly, chemistry labs and eldercare wards plus a 200 m² optical motion-capture hall.
• Zhiyuan / AgiBot (9 sites): Pudong-Lingang, Chengdu Pidu, Wuxi, Nanchang, Jiaxing, Chongqing, Zhuhai and Guian. Its Shanghai Pudong factory runs 200+ granular scenario tasks across 3,000+ object categories and open-sourced the AgiBotWorld real-robot datasets. With the Shanghai Academy, live data flywheeling lifted long-horizon sub-step scores ~50% versus supervised fine-tuning.
• UBTECH (5 sites): industrial-grade centers in Liuzhou and Zigong run 500+ robots continuously collecting joint torque and tactile data (~400+ effective hours per robot per year); UBTECH says its real-embodiment dataset totals ~11 million entries, 80%+ from industrial scenarios.
• Paxini (5 sites), Leju (5), Luming (3): Paxini's Tianjin "super factory" spans 12,000 m² with annual output targeted in the hundreds of millions of entries, and in May 2026 it completed a compliant export of embodied data. Twelve more companies — JD, Realman, Galaxea, Xinghai, Lens Technology, Zoomlion, Qunhe, Zero-Order, Qupai, Kaipuyun, UQI and more — each run one site.

Policy has been pushing alongside capital: eight ministries' "AI + Manufacturing" action plan calls for humanoid pilot bases and training grounds, MIIT and SASAC run a "real-scenario training" initiative, and the National Data Administration convened a work session on October 28 stressing "need-driven supply." China has issued 40+ local policy documents on training grounds between 2024 and June 2026, and the National Data Group was formally established on October 20.

Seven business models, one dominant one

Source: Embodied Data Development Report 2026 via The Paper. Custom collection remains the main revenue model; most suppliers sell one-off projects rather than reusable products.

Who sells? The report divides the field into data suppliers (Paxini, Lightwheel AI 光轮智能, Luming 鹿明, Lingchu 灵初, Mifei, Noyiterobot), platform traders (JD Cloud's embodied-data trading platform, Baidu Smart Cloud's "data supermarket" — commission models still unformed) and public-service operators (Shijingshan Beijing training center, the national innovation center). On the demand side, leading robot makers mostly self-collect; free open datasets dampen paid demand from model companies; and downstream scene customers are still at trial stage, unsure how to open business data or verify ROI. On the funding side, embodied-data pure-plays are drawing capital — Jianzhi Robot (简智机器人, Gen DAS rigs / Gen Matrix governance / Gen ADP production line) announced a $100M Series A+ on November 2, 2026.

The controversy: "buy the robot, buy back the data"

The fastest-growing model has also become the most contested. A typical arrangement: a local state-owned platform funds and builds a data center, buys robot hardware from a company, and the company then buys the data back — for its own model training or resale. On September 9, Mech-Mind founder Shao Tianlan publicly criticized "well-known, highly valued" embodied-AI companies for using such centers plus related-party deals with local governments, investors and suppliers to manufacture revenue ahead of IPOs.

Industry responses show a spectrum:

• Defense: Lingyu Intelligent (灵御智能) co-founder & CEO Jin Ge — whose own customers are ~50% data centers, ~30% JD 7FRESH-style real scenes and ~20% research/model firms, with centers+scenes contributing 70–80% of order revenue — calls the loop "order agriculture for data": state platforms hold the heavy assets, companies repurchase at below-market prices, and two-shift operation pays back a center in 8–9 months. He distinguishes it from "brushing orders" (fake revenue on useless data).
• Hard line: Songyan Dynamics founder Jiang Zheyuan says his company took zero data-center orders this year: "If they only buy the robot for collection, that's a real order. If they require you to buy data back, that's not real."
• Investor scrutiny: Shuinmu Capital chairman Tang Jingcao says state data demand has a genuine core (model training, policy incubation), but the key is operations — utilization, external data buyers. Due diligence now checks device uptime, power draw and maintenance records, traces who actually pays, and flags customers related to investors at unfair prices. "Worst case, we exclude that revenue from core valuation."
• Operating standard: the Shijingshan training center says sustainable collection needs three things — real industrial pull with data flowing back into real scenarios, high-dimensional (4D world-model) data assets rather than 2D trajectory clips, and a closed train/validate loop including failure samples.

Using Shao's L0–L10 robot-readiness ladder, Jin Ge argues long-term-value revenue sits at L8–L10 (real terminal business), while much of today's industry is at L0–L3 demo stage — "the market is pricing mature-stage valuations on pre-L3 reality." Multiple insiders told Securities Daily-affiliated media that centers won't disappear — post-training still needs real-robot data — but they are receding from "protagonist of the valuation story" to "stepping stone of an early industry."

What happens next

Analysts expect the collection segment to keep compounding even as revenue scrutiny tightens. Zheshang Securities models the midstream data-collection market reaching ¥174.1B by 2030 (2026–30 CAGR ~203%), with upstream hardware at ¥15.3B; Guojin Securities argues collection equipment and services are the industry's clearest, most cash-stable business, and recommends suppliers with orders/capacity plus firms that platformize processing and extend into evaluation, post-training and deployment. For reference, the global benchmark Scale AI is valued at $29B.

China's answer to the data famine is still being built — and argued over. The OECD of this market, if one emerges, will be whoever can make the data flywheel real: robots in real scenes, data back into models, better robots into more scenes. National Data Administration adviser Ouyang Rihui predicts the industry will converge on high-quality interaction data from small samples, a virtual-real hybrid paradigm, and business models bound to scenes that recycle data continuously. For now, the "data infrastructure war" is quiet, provincial, and very large — and it will shape how fast Chinese robots actually learn to work.

Frequently asked questions

What is an embodied-AI data collection center?

A facility where operators teleoperate robots (via gamepads, VR or master arms) through mock homes, shops, factories or labs, recording vision, joint angles and force feedback frame by frame to train embodied models. China had 70+ operating centers by June 2026 and 106 built by August.

How much does embodied data cost?

Teleoperated real-robot data: ¥500–1,000/hour, dropping to ~¥275/hour of effective data at scale; portable/no-robot collection ¥100–200/hour; synthetic data ~1/100 the cost of real data. Standard dataset licenses run ¥10k–500k per set, industry-exclusive up to ¥5M.

Who are the main collection players?

The national Guodi network (10 grounds, Shanghai Qilin anchor), Zhiyuan/AgiBot (9), UBTECH and Paxini and Leju (5 each), Luming (3), and 12 companies with one each including JD, Realman, Galaxea, Xinghai, Lens and Zoomlion.

Is the 'buy robot, buy back data' model real?

State platforms do fund centers and buy hardware, and some companies repurchase data — critics call part of this manufactured revenue; defenders say it is legitimate if data is genuinely used and externally valued. Investors now scrub utilization, payments and related-party relationships, and may exclude such revenue from valuations.

Why is real robot data scarce?

Robots need physical multimodal interaction data that was never systematically recorded; analysts estimate a 90% gap versus mature-model needs, with one CAICT-linked estimate of 21,200 robot-years for 1.1 trillion tokens.

Disclaimer: This article is independent editorial analysis based on CAICT's training-grounds report, the Embodied Data Development Report 2026, Zheshang/Guojin research notes and Chinese press coverage (Securities Daily / 科创板日报, The Paper, Jiemian, Xinhua, 36Kr). Market-size forecasts are broker estimates, not guarantees. Nothing here is investment advice.