China asked humanoid robot IPOs to prove the humans are buying

01

WHAT ACTUALLY CHANGED

Reuters reported that Chinese regulators have used informal window guidance to hold back some humanoid-robot IPOs and raise the approval bar. One source described an effective temporary freeze, while another explicitly said there was no formal ban and called it a sector-specific slowdown. The second description is the safer public conclusion.

The reported shift followed Unitree Robotics' volatile Shanghai debut. Reuters said the stock rose more than fivefold when trading began a month earlier and later fell 55 percent from its peak. A price reversal can concentrate regulatory attention, but it cannot by itself establish the quality of the underlying company, product or revenue.

At least half a dozen Chinese humanoid-robot companies are preparing to go public, according to Reuters. The report named Deep Robotics, X Square Robot and AGIBOT among them. Those companies did not answer Reuters' questions about whether their plans had been slowed, so the reporting does not support assigning the same regulatory status to every candidate.

The review is concentrating on revenue from local-government-backed robot data centers and joint ventures. One investor source told Reuters that local governments could provide 80 to 90 percent of the initial investment in some joint ventures. Those arrangements can create paid orders and useful infrastructure while still leaving open whether an independent commercial buyer would choose the same product on the same terms.

The same source estimated that valuations at some robot companies could fall 60 to 70 percent if revenue associated with data centers were removed. That is an investor estimate, not an audited adjustment, a regulator calculation or a finding that all such revenue lacks economic value.

Mech-Mind Robotics chief executive Shao Tianlan alleged in a WeChat post that some highly valued embodied-AI companies were using data centers, related-party deals and other unsustainable arrangements while racing toward IPOs. Reuters reported the allegation and said Shao declined to elaborate. No named company, transaction ledger or independent substantiation accompanied the public claim.

The policy direction has not reversed. Reuters said Beijing continues to promote embodied intelligence as a strategic emerging industry. The reported intervention is an attempt to separate capital-market readiness from national support for the technology, not evidence that China has abandoned humanoid robotics.

Market expectations are becoming more selective. Reuters quoted investors and a banker asking about deployments, order volumes, useful factory work and valuation instead of treating a stage demonstration as a proxy for commercialization. That shift is basic due diligence arriving after an unusually loud opening act.

Commercial context remains small compared with conventional robotics. A separate Reuters report, drawing on the International Federation of Robotics, said about 7,000 humanoid robots were sold globally for industrial and professional-service use in 2025, while about 542,000 conventional industrial robots were installed in 2024. The periods and categories differ, so this is scale context rather than a direct growth comparison.

Unitree's official store shows that real products are available at very different price points, from an R1 listed from $4,900 to an H1 listed at $90,000. A public product page proves that a machine is offered for sale. It does not disclose unit shipments, discounts, buyer mix, gross margin, factory use, uptime, support cost or repeat demand.

02

WHY THIS MATTERS

A robot company can have impressive engineering and weak revenue at the same time. Hardware demonstrations answer whether a machine can perform a task under selected conditions. Commercial evidence asks whether customers will pay for that task repeatedly after installation, integration, safety work, maintenance, downtime and human supervision are counted.

State support is not fake demand by definition. Governments often fund shared infrastructure, early research, workforce training and first deployments because a new market would otherwise develop slowly. The problem begins when promotional purchases are presented as proof of a self-sustaining customer market without clearly separating the two.

Buyer independence changes the meaning of a sale. Revenue from a founder-controlled affiliate, investor, joint venture or public development project may be valid accounting revenue. It can still say less about market demand than a purchase by an unrelated customer that compared alternatives, paid from its own budget and returned for more.

Data centers create an especially tricky loop. Robots need demonstrations and operating data to improve. A publicly supported center can buy robots, employ operators and generate that data, which makes the product better. But the center is simultaneously customer, training facility and policy instrument. Readers need those roles separated before they infer broad adoption.

An order is only the beginning of industrial value. A credible deployment record includes delivery, acceptance testing, productive hours, task success, human interventions, safety incidents, downtime, maintenance, energy, tooling, integration and cost per successful unit of work. The factory floor is much less flattering than the product launch, which is exactly why it is useful.

Repeat orders are stronger than pilots because they force a customer to compare promised value with lived experience. Even then, concentration matters. Ten repeat orders from one subsidized group are not the same market signal as ten unrelated customers in different industries buying under ordinary budgets.

Public investors receive less control than private backers and depend more heavily on formal disclosure. If a company's headline growth depends on a small number of policy-linked transactions, the prospectus should identify the customers, relationships, payment terms, subsidies, receivables and concentration clearly enough for outsiders to test the story.

A listing review can improve the sector if it rewards measurable deployment instead of theatrical novelty. It can also slow useful companies if informal expectations are inconsistent or opaque. The answer is not permanent ambiguity. Regulators should publish the evidence they expect and apply it across comparable issuers.

The global robotics race is watching China because its supply chains, manufacturing base, public investment and willingness to run pilots can compress development time. Better disclosure would make Chinese experience more useful abroad by showing which deployments survived ordinary work rather than only which machines looked advanced on camera.

Workers matter inside the revenue test. A machine can generate sales while shifting risk, surveillance or unstable work onto data collectors, teleoperators, maintenance crews and factory staff. Commercial quality should include who keeps the robot functioning and whether the promised productivity depends on hidden human labor.

The new scrutiny does not prove a bubble has burst. It marks a change in the burden of proof. Companies that can show independent buyers, paid productive work, reliable operation and repeat orders should benefit from a market that no longer prices every humanoid story as though it were the same story.

03

WHERE IT COULD HELP

• Separate product revenue, research contracts, data-center purchases, joint-venture sales, subsidies, grants, services and related-party transactions in every financial table
• Identify the ultimate buyer, its ownership relationship, its source of funds and whether it selected the vendor through a competitive process
• Report units ordered, delivered, accepted, operating, returned and retired instead of using one shipment number for every stage
• Publish productive hours, task completion, human interventions, downtime, repair time, safety incidents and customer-authored outcomes by use case
• Show the full customer economics, including robot price, integration, tooling, software, training, maintenance, energy, insurance, supervision and facility changes
• Distinguish paid pilots from production deployments and disclose the criteria that convert a trial into a commercial rollout
• Track repeat orders, renewal rates, backlog cancellations, payment collection and concentration among independent customers
• Disclose related parties, investor-linked customers, joint ventures, local-government participation and any side agreement that changes the economics of a sale
• Explain what a data-collection center buys, who owns the data, how workers are paid, which robots use the data and whether the center has demand beyond the funded program
• Reconcile bookings, recognized revenue, cash receipts, accounts receivable and deferred obligations so a large contract does not masquerade as collected demand
• Give regulators a published sector checklist with consistent evidence requirements, review clocks and an appeal route instead of leaving the standard entirely informal
• Let workers and customers report unsafe deployment, hidden teleoperation, falsified uptime or coerced purchases through protected channels
• Price a company from filed financials and operating evidence when available, not from robot videos, social-media attention or a first-day share move