China’s humanoid robots backed by ‘circular’ deals

Purchases from Beijing-backed training hubs lead investors to question level of real demand

Financial Times UK

25 Aug 2026

WILLIAM LANGLEY

China’s humanoid-robot makers are generating much of their revenue from selling

machines to government-backed training centres — which then collect and sell

training data back to the robot makers, raising concerns about actual demand in an

industry that Beijing is keen to promote.

The widely adopted model, reminiscent of Nvidia’s “circular financing” of artificial

intelligence data centres, has fuelled China’s so-called embodied AI industry. Valuations of start-ups such as AgiBot and Hong Kong-listed UBTech have soared on

expectations that humanoids represent the future of AI.

Investors are beginning to question, though, whether government-driven purchases

can lead to real commercial demand.

“There is a broad consensus among early-stage investors that humanoid robotics is

approaching the peak of the hype cycle,” said a senior investor at a Beijing-based

venture capital fund who asked not to be named. “We’re looking for opportunities

to sell down some of our holdings to other investors and secure an exit.”

As part of Beijing’s plan to develop China’s humanoid-robot industry, it has encouraged local governments to build large-scale training centres, where humans “teach”

robots how to perform physical tasks through a remotecontrolled process called

teleoperation.

teleoperation.

This has encouraged the proliferation of robotics start-ups. Nearly 370 have been

established in the past two years, and more than 50 of them have listed or are preparing to go public. The most closely watched, Unitree, surged more than 600 per

cent after listing on Shanghai’s tech-focused Star Market on Wednesday last week,

giving it a $50bn valuation.

The centres, often co-funded by local governments and robot makers, buy the

machines, generate training data and then sell the data to the robot makers to

improve their technology.

“This model is spreading rapidly because it reduces the cost of building facilities,

buying equipment and organising teleoperation teams,” said Poe Zhao, an independent China tech analyst and founder of Hello China Tech.

“But it also blurs the distinction between independent demand and demand created

within a policy-supported ecosystem.”

Only a small share of the data was sold to non-robotics groups such as carmakers

for assembly-line use, training centre staff said. “This model can’t last,” said another

investor. “If they fail to prove their robots can be deployed at scale on factory

floors, investors will begin to reassess valuations.”

More than 90 training centres had been established or were being built across

China by June, according to consultancy Interact Analysis. Leading centres said they

generated more than 10mn data points a year.

Prices vary, but one seller told the FT that training data for a five-minute robot

dance could cost as much as Rmb1mn ($148,000).

The training centre build-out is reshaping industry forecasts. Morgan Stanley raised

its estimate for China’s humanoid-robot shipments in 2026 to 50,000 units, from

28,000 in June, citing stronger than expected purchases by local governments and

commercial users. Supporters of the model argued that it would help to build up

the country’s robotics industry and supply chain, noting the examples of electric

vehicles and solar panels — sectors that China now dominates after government

purchases drove initial demand.

For local governments, the model helps attract investment, talent and supply chains

to areas where income from land sales has declined. Some centres hire university

students as robot trainers and offer paid tours for children and teenagers during

school holidays.

For the robot makers and their suppliers, the centres are a source of revenue amid

limited commercial demand. Shenzhen-based Leju Robot said training centres

accounted for 45 per cent of sales of its Kuavo humanoid last year, making them its

largest revenue source.

UBTech disclosed Rmb140mn ($21mn) of orders from governmentbacked training

centres last year. Although still lossmaking, the company said that robot deliveries

accounted for 41 per cent of its Rmb2bn revenue last year and expected government orders to drive further growth this year.

Almost three-quarters of Unitree’s humanoid revenue in the first nine months of

last year came from users in the education and research sectors, including universities. Analysts said that a relatively small proportion of shipments went to data

collection centres.

Analysts said that the close ties between local governments and robot makers made

it difficult to distinguish genuine demand from policy-driven purchases.

“Companies like ours need revenue, not necessarily profits,” said an algorithm

engineer at a Beijing-based company that sells software to robot makers. “The robot

training centres can tell their superiors they’ve bought the equipment and robots,

they’ve built data collection facilities and they’ve sold data.” He added: “Both sides

get what they need, and both sides have something to show.”

At Beijing’s largest robot training centre, where more than 100 Kuavo robots have

been deployed, Leju owns almost 38 per cent of the operating company, according

to records.

UBTech declined to comment. Leju and Unitree did not respond to requests for comment.

Another open question is whether the data generated justifies the investment.

Marco Wang, analyst at Interact Analysis, said training centre data would not be

“100 per cent useful” because the robots were not deployed in real-world settings.

“The real-world application or real-world manufacturing line, realworld warehouse

is always different from your scenario,” said Wang.

Analysts at Goldman Sachs said that scarce high-quality, real-world data remained

the biggest hurdle to widespread adoption.

A senior manager at a training centre in northern China said each robot maker’s

data could be used only by that company, raising concerns about compatibility. She

added that on average, only two or three hours of data from an eight-hour training

shift were usable.

“China often accepts duplication and failed projects in the early stages of a strategic

industry,” said Hello China Tech’s Zhao. “The expectation is that technical learning,

stronger supply chains and a handful of globally competitive companies may justify

losses elsewhere.