China’s Humanoid Makers Took 97% of Shipments, Bloomberg Reports

China’s humanoid robot makers captured more than 97% of global shipments during the first half of 2026, according to a new Bloomberg report. Worldwide volume reportedly exceeded 19,100 units, more than triple the comparable 2025 total.

Those numbers make China the clear early leader in manufacturing humanoid machines at scale. They do not establish that the machines can perform valuable work without close supervision.

That distinction defines the contest with American companies such as Tesla and Figure AI. Chinese manufacturers are shipping hardware while prominent US rivals remain focused on intelligence, autonomy, and ambitious production plans.

The rsshub bloomberg feed surfaced a striking headline, but the underlying development is larger than one market-share figure. China has created a supply chain that can turn experimental robots into repeatable products faster than competitors.

The result also carries an important warning. Shipments can include research units, demonstration machines, government-backed deployments, and robots assigned to narrowly controlled tasks.

A shipment is not proof of productive operation. It says little about reliability, utilization, customer retention, or the labor hours a robot can replace.

The next stage will test whether China’s manufacturing lead becomes a durable commercial advantage. It will also show whether US companies can convert stronger claims about artificial intelligence into dependable machines.

What the 97% Shipment Share Actually Measures

China’s lead is a manufacturing fact, but its commercial meaning remains unsettled.

The Bloomberg report says Chinese vendors commanded more than 97% of worldwide humanoid shipments in 2026’s first half. Total volume reportedly reached about 19,100 units.

That compares with approximately 5,100 units in the same period one year earlier. The market therefore expanded by more than three times while remaining overwhelmingly concentrated among Chinese suppliers.

The increase follows a strong 2025 for AGIBOT, Unitree Robotics, and UBTech Robotics. Omdia previously placed those companies in its first tier for general-purpose embodied intelligent robot shipments.

Embodied intelligence describes AI operating through a physical machine that senses, moves, and responds to its environment. It connects software decisions with motors, joints, cameras, and hands.

AGIBOT and Unitree each shipped more than 5,000 units during 2025, according to Omdia figures reported by the Associated Press. UBTech also cleared 1,000 units.

Tesla reportedly shipped fewer than 500 general-purpose embodied robots during that year. Figure AI and other American developers also remained far behind China’s leading companies by unit volume.

This gap matters because production creates feedback. Every manufactured robot can reveal component failures, assembly bottlenecks, maintenance needs, and weaknesses in real operating environments.

Higher volume can also support specialized suppliers. Manufacturers gain reasons to improve actuators, reducers, sensors, batteries, hands, and control hardware for a growing customer base.

China already possesses dense supplier networks developed through automobiles, consumer electronics, drones, batteries, and conventional industrial automation. Humanoid developers can adapt this infrastructure instead of building every component category from scratch.

The reported 97% share is consistent with other market measurements. ABI Research says China currently represents 97% of the installed humanoid base, although definitions and coverage can vary between research firms.

Its humanoid forecast projects the worldwide installed base rising from 24,000 machines in 2025 to 248,000 by 2030. ABI expects China’s share to fall as other regions increase deployment.

That expected decline would not necessarily indicate Chinese weakness. China could keep expanding rapidly while North America, Europe, and other Asian markets grow from much smaller bases.

Shipment leadership also differs from technological leadership. A vendor can ship many affordable research platforms without solving general-purpose manipulation, autonomous planning, or safe operation around people.

The rsshub bloomberg headline therefore captures one dimension of the race. It measures how quickly companies place machines into customers’ hands, not how independently those machines perform after arrival.

For now, that dimension favors China by an extraordinary margin. The more difficult question is how the country built such a wide advantage before a proven mass market emerged.

China Built a Hardware Flywheel Before Demand Was Clear

China’s advantage comes from manufacturing density, policy support, and rapid iteration, not one decisive robot invention.

The country had more than 140 humanoid manufacturers and over 330 models in 2025, according to official figures cited by the Associated Press. That crowded field creates both competition and duplication.

Many developers cluster around established electronics and automotive centers. These regions offer factories, engineering talent, component suppliers, logistics networks, and customers willing to test automation.

This structure shortens the path between a prototype and a production run. A company can revise a joint, source a different motor, or change a battery design without rebuilding its entire supply network.

The country’s electric vehicle industry offers a useful precedent. Dense competition, local suppliers, manufacturing experience, and sustained investment reduced costs while accelerating product cycles.

Humanoid robots draw from many of the same capabilities. They need battery systems, electric motors, power electronics, thermal management, cameras, processors, and production quality controls.

Government policy provides another source of momentum. China’s 2026 to 2030 development plan identifies humanoid robots among the technologies targeted for faster development and wider application.

Local governments and state-linked organizations can support factories, research centers, training facilities, and early deployments. These programs give manufacturers customers before private demand becomes predictable.

That support does not make every order artificial. Research institutes, manufacturers, power facilities, logistics operators, and public venues have legitimate reasons to test embodied AI.

However, policy-backed demand can blur the market signal. A robot purchased for research or demonstration does not validate the economics of continuous commercial operation.

Chinese companies are also pursuing varied use cases. Some machines sort parcels, guide visitors, inspect facilities, serve drinks, perform on stage, or collect training data.

These applications help companies develop specific capabilities. They also let vendors ship products before achieving the general autonomy suggested by the human-shaped design.

A humanoid can be valuable without performing every human task. Factories may need one machine that moves through human-designed spaces and completes a limited sequence reliably.

That narrower target is easier than creating a household assistant. Homes contain clutter, children, pets, stairs, fragile objects, and an almost unlimited range of unexpected situations.

Industrial environments can be redesigned around robots. Companies can mark routes, standardize containers, control lighting, reduce obstacles, and separate machines from workers.

China’s scale lets vendors test these narrower deployments across more machines. Field experience then informs the next hardware revision, creating a manufacturing and data flywheel.

Every deployment can generate recordings of movements, failures, interventions, and successful task completion. Developers can use that information to improve perception and control models.

The flywheel is not automatic. Data must be relevant, accurately labeled, legally collected, and diverse enough to improve behavior outside one installation.

Still, shipping creates opportunities that laboratory work alone cannot provide. A robot that never leaves a demonstration area cannot reveal the full cost of maintenance and customer support.

The rsshub bloomberg shipment figures show that Chinese manufacturers have entered this feedback cycle at greater scale. Their lead pressures American rivals to move from demonstrations toward repeatable production.

China’s Robot Volume Puts Tesla and Figure Under Pressure

The central contest is Chinese manufacturing scale against the US bet on advanced autonomy.

Tesla has given Optimus a central place in its long-term strategy. Elon Musk has repeatedly described humanoid robots as a potentially enormous business linked to the company’s AI capabilities.

Yet Tesla’s shipment volume remains small compared with AGIBOT and Unitree. That contrast turns production execution into an immediate test of the Optimus narrative.

Tesla does possess relevant advantages. It manufactures complex electromechanical products, develops computer-vision systems, designs AI hardware, and collects real-world data from a large vehicle fleet.

Cars and humanoids are not interchangeable, however. A robot must balance, manipulate unfamiliar objects, recover from errors, and operate safely near people.

Hands are particularly difficult. Human hands combine strength, precision, sensitivity, and adaptability within a compact structure that engineers struggle to reproduce economically.

Tesla is trying to connect its American AI development with manufacturing expertise in China. Wang Hao, Tesla’s vice president and China president, called the company’s Shanghai operations a key to solving mass production.

The Shanghai production plan remains short on operational detail. Tesla has not publicly explained exactly how the facility will support the robot business.

The company has also prepared to convert space at its Fremont factory from older vehicle production to Optimus manufacturing. The move demonstrates commitment but does not guarantee volume or reliability.

Figure AI represents a different American approach. It concentrates on general-purpose robot intelligence and partnerships for industrial deployment rather than leveraging an existing consumer manufacturing business.

American developers often emphasize foundation models, vision-language-action systems, and autonomous reasoning. A vision-language-action model converts visual observations and language instructions into physical actions.

This software could become a decisive advantage if it lets one robot learn many tasks with limited custom programming. Flexible behavior would make the humanoid form more economically useful.

Chinese manufacturers are not ignoring AI. Their machines also use multimodal models, imitation learning, teleoperation data, simulation, and reinforcement learning.

The difference is one of present emphasis. Chinese companies have already established a commanding shipment base, while leading US companies promise intelligence that should justify future scale.

That creates a reversal in the familiar technology narrative. American firms lead many global discussions about generative AI, but Chinese firms dominate the physical shipment count.

The pressure extends beyond Tesla and Figure. Boston Dynamics, Apptronik, Agility Robotics, and other developers must decide how much hardware to manufacture before demand becomes predictable.

Moving slowly preserves capital and avoids inventory built around immature designs. It also reduces access to production learning and real-world data.

Moving quickly produces more feedback but increases exposure to component failures, weak demand, and expensive support obligations. Early hardware can become obsolete after a major design revision.

China’s current position suggests its companies have accepted more of that manufacturing risk. Some have done so with help from public investment and state-linked customers.

US companies are betting that better intelligence will eventually matter more than early shipment volume. That thesis remains credible because autonomy drives the value of general-purpose machines.

However, software leadership cannot substitute indefinitely for production. An advanced model has limited commercial impact if the robot carrying it cannot be made, maintained, or delivered economically.

The 97% figure therefore changes the burden of proof. Chinese manufacturers must show that shipped machines provide value, while American rivals must show that their software advantage reaches customers.

The Shipment Boom Still Has a Buyer Problem

The largest uncertainty is not whether China can build humanoids, but whether enough customers can use them productively.

An AP demand review found a market filled with orders, demonstrations, and serious doubts about practical adoption. Those doubts qualify every shipment estimate.

Samm Sacks, a senior fellow at New America, told the AP that most humanoids remain performative rather than functional. She pointed to their weakness in messy, unpredictable environments.

Venture investor Chibo Tang similarly argued that limited use cases constrain demand and mass production. Both assessments focus attention on what happens after a machine reaches a customer.

A robot can appear capable during a prepared demonstration. The operator controls the setting, selects the task, removes difficult objects, and rehearses the sequence.

Commercial work offers fewer protections. A box arrives damaged, a tool shifts position, a worker blocks the route, or a network connection fails.

Reliability becomes more important than a dramatic single performance. Businesses need to know how often the robot completes its task and how quickly it recovers.

They also need the total cost of operation. That includes installation, integration, supervision, energy, maintenance, replacement parts, software, safety controls, and downtime.

A low acquisition cost cannot compensate for frequent human intervention. A machine that needs constant remote control may simply move labor to another location.

Teleoperation, where a person remotely controls the robot, remains useful for training and exception handling. It becomes a problem when vendors present supervised work as autonomous performance.

The reported shipment data does not disclose intervention rates. It also does not show how many robots perform daily work instead of research, entertainment, or data collection.

One real-world trial illustrates the limitation. A Beijing customer tested a robot that organized shoes, folded clothes, and changed garbage bags while accompanied by a human cleaner.

She found the machine impressive but inefficient and difficult to move inside a small home. That gap between novelty and utility remains central to consumer adoption.

Factories offer a better near-term environment, yet humanoids still face competition from conventional automation. Fixed robotic arms already complete repetitive tasks with high speed and reliability.

Autonomous mobile robots can carry materials without needing legs or humanlike hands. Specialized machines usually outperform general-purpose designs at a defined task.

Humanoids make the strongest case when customers need one machine to navigate spaces built for people and perform changing jobs. That flexibility must outweigh their added mechanical complexity.

Safety adds another constraint. A full-size robot working near people carries weight, force, batteries, moving joints, and software that can make incorrect decisions.

Developers need reliable emergency stops, collision detection, force limits, access controls, and predictable fallback behavior. Regulators and insurers will also demand evidence.

Cybersecurity will become more important as deployments expand. A networked machine with cameras, microphones, mobility, and physical strength creates risks beyond those of a conventional computer.

China itself has acknowledged the possibility of excess. Its national economic planning agency previously warned about bubble risks as many companies pursued similar humanoid projects.

The official bubble warning highlighted concerns about commercialization and repetitive product development. Government support does not eliminate pressure to find sustainable customers.

Shipment definitions require scrutiny as well. Research firms may classify educational platforms, service robots, industrial prototypes, and general-purpose humanoids differently.

A delivered unit might enter a customer trial without generating recurring value. Another could spend most of its time collecting data for future models.

These machines still contribute to learning. However, readers should not interpret 19,100 shipments as 19,100 autonomous workers replacing full-time employees.

The rsshub bloomberg story establishes an early industrial lead, not the completion of the humanoid transition. Commercial proof requires operating results that shipment counts cannot provide.

Why Manufacturing Scale Can Become a Software Advantage

Hardware volume matters most when deployed machines generate useful data and force better engineering decisions.

Humanoid development joins mechanical engineering with machine learning. Progress in one area can stall when the other fails.

A sophisticated model cannot compensate for an overheating actuator or a damaged cable. Better joints cannot make a robot understand an unfamiliar instruction.

Large production runs expose physical weaknesses quickly. Repeated failures reveal which parts need redesign, which tolerances matter, and which repairs consume excessive time.

Field deployment can also improve software. Cameras and sensors capture variations that simulation and curated laboratory tests often miss.

A warehouse floor changes throughout the day. Lighting shifts, packages differ, people cross paths, and equipment moves between expected positions.

Models trained on varied observations can become more adaptable. Yet developers must prevent the data pipeline from filling with low-quality or repetitive examples.

Human intervention data can be particularly useful. When an operator takes control, that moment identifies a situation where the autonomous system lacked confidence or chose poorly.

Teams can review these cases, label the correct action, and retrain the model. Over time, successful learning should reduce the intervention rate.

China’s shipment lead gives its manufacturers more potential data sources. It does not guarantee that every company has the infrastructure or talent to use them well.

Software platforms must synchronize recordings, protect sensitive information, compare model versions, and reproduce failures. They must also separate improvements from changes that create new risks.

This work resembles the knowledge challenge faced by other engineering organizations. Teams need searchable records connecting observations, decisions, tests, and design revisions.

A structured engineering knowledge base can help technical teams preserve those connections. The principle matters even when the underlying product is a physical robot.

The main advantage comes from closing the loop faster. A company ships hardware, observes failures, updates the design, retrains software, and validates the next release.

Chinese consumer electronics and electric vehicle companies already operate rapid product cycles. Humanoid vendors can recruit engineers familiar with those production disciplines.

Vertical integration can strengthen the loop. A company controlling more components can coordinate mechanical, electrical, and software changes without waiting for several outside suppliers.

It can also create blind spots. Internal components may remain in use because switching carries organizational costs, even when an external alternative performs better.

The crowded Chinese market provides another corrective force. Vendors can compare their machines against many local competitors and hire workers with relevant production experience.

Consolidation is likely as weaker companies exhaust funding or fail to secure customers. The surviving firms could inherit talent, supplier relationships, and intellectual property.

US companies retain important strengths in AI research, semiconductor design, cloud computing, and venture funding. They may also develop models that transfer between different robot bodies.

If intelligence becomes portable, Chinese hardware makers could adopt strong third-party models. Conversely, American software companies could license intelligence to manufacturers with better production systems.

That possibility complicates the national rivalry. The eventual market may not divide neatly between complete Chinese and American robots.

Supply-chain controls and security rules could push it in that direction, however. Governments may restrict robot components, advanced chips, training data, or deployment in sensitive facilities.

Such limits would encourage separate technology stacks. They could also slow learning by reducing access to the best hardware, software, and markets.

The decisive metric is therefore not shipment share alone. It is how efficiently each shipped unit improves the next generation while delivering enough present value to retain customers.

China currently has more machines entering that loop. The US still has an opportunity to build a stronger intelligence layer, but time spent outside production carries a growing cost.

Three Signals Will Show Whether China’s Lead Is Durable

The next evidence must connect shipment growth with dependable work, repeat customers, and a credible response from US rivals.

The first signal is the quality of Chinese deployments through the end of 2026. Watch for disclosed utilization, intervention rates, task success, downtime, and repeat orders.

A manufacturer that reports only cumulative shipments provides an incomplete picture. Customers renewing deployments or expanding from trials would offer much stronger commercial evidence.

Industrial and logistics sites deserve the closest attention. Their structured environments provide the most realistic near-term path to productive humanoid work.

A successful deployment should run across full shifts and handle normal variation. It should also recover safely when objects, routes, or work sequences change.

Evidence of repeat purchases would strengthen the claim that China’s manufacturing lead has become a market advantage. Rising inventories or government-dominated orders would weaken it.

The second signal is Tesla’s production execution. The company must translate Optimus demonstrations and factory plans into machines operating consistently at useful tasks.

Its Shanghai operations are especially important because they connect Tesla with China’s manufacturing base. Concrete details about component sourcing and production responsibilities would clarify that role.

Watch for robots performing sustained work inside Tesla facilities rather than isolated demonstrations. Performance should include disclosed human intervention and failure rates.

Tesla does not need to match China’s aggregate volume immediately. It does need to prove that its intelligence-focused approach produces measurable operational value.

A credible production ramp would weaken the idea that China’s head start is structurally unassailable. Further delays would widen the gap in manufacturing experience and deployment data.

Figure AI and other American developers provide related evidence. Their customer trials can show whether advanced models overcome the production advantage held by Chinese suppliers.

The third signal is the industry’s treatment of shipment data. Research firms and manufacturers need clearer distinctions between prototypes, research platforms, commercial deployments, and autonomous production machines.

Common definitions would make comparisons more useful. They would also reduce the temptation to treat every delivered robot as an equivalent unit of economic output.

Look for analysts to publish installed-base data alongside utilization and application categories. Shipment figures should identify geographic destination, customer type, and operational purpose where possible.

ABI Research expects China’s share of installed humanoids to decline as other markets ramp up. That would be healthy evidence of a widening global market, not necessarily a Chinese retreat.

The strongest scenario for China combines continued unit growth with repeat commercial purchases and improving autonomy. That pattern would validate both its manufacturing strategy and its data flywheel.

The weaker scenario combines rising shipments with limited utilization, heavy supervision, and concentrated public-sector demand. That would suggest production has moved ahead of the market.

For the United States, the strongest response is not another polished demonstration. It is a robot that customers can deploy, maintain, and trust across repeated work cycles.

The rsshub bloomberg headline offers a clear snapshot of the current race. China dominates the count, while the economic value behind that count remains open.

Readers should now look beyond dancing, boxing, and carefully staged manipulation. The important evidence will come from ordinary work performed repeatedly when no camera crew is present.

That evidence will determine whether humanoids become a major automation platform or remain expensive machines searching for enough suitable jobs. It will also decide whether China’s 97% share marks a temporary head start or the foundation of lasting industrial leadership.