1. China embodied AI funding: the first-half numbers
Companies building embodied artificial intelligence in China raised about 93.5 billion yuan, roughly $13 billion, across 322 financing rounds in the first half of 2026, according to figures compiled from disclosed deals. The total is about five times the amount raised in the first half of 2025. Investors have since grown more selective, and the money that is still flowing is moving away from robot hardware and toward the software that lets machines perceive and act.
The shift marks a change in how the sector is being judged. Chinese manufacturers have shown that they can assemble capable robot bodies, and the open question is now whether those machines can do useful work at a price customers will pay.
The headline number also shows how quickly the field has attracted attention. Deal counts of this size suggest that the sector was not driven by a handful of very large rounds alone but by a broad base of companies, from component suppliers to full-system makers, each pursuing a different part of the problem. A wide base of that kind produces many competing approaches, and it also means that many of the participants are unlikely to survive a tougher funding climate.
2. What embodied AI means
Embodied AI refers to artificial intelligence systems that operate through a physical form, such as a humanoid robot, a wheeled platform or a robotic arm. Unlike a chatbot that processes text on a server, an embodied system must read its surroundings through cameras and other sensors, decide what to do and then move through a world that is cluttered, changeable and unforgiving of mistakes.
The field is often described in two layers. The body consists of the mechanical frame, the motors, the joints and the sensors. The brain is the software, usually built on large models, that interprets sensor data, plans a sequence of actions and adapts when something unexpected happens. Progress in the body can be bought with manufacturing skill and supply chains. Progress in the brain depends on data, computing power and research talent, and it tends to be harder to copy.
Training the software side presents its own difficulties. A language model can learn from text gathered across the internet, but there is no comparable body of recorded physical experience for a robot to study. Developers therefore rely on a mix of demonstrations by human operators, simulated environments and data collected from machines working in the field. Each of these sources has limits, and gathering enough varied examples is among the most expensive parts of the effort.
3. A funding surge, then a cooling
The first-half figure reflects an exceptional burst of enthusiasm. A five-fold increase in a single year, spread over 322 rounds, indicates that capital was available to a wide range of start-ups, many of them at an early stage. Investors were drawn by the prospect that humanoid machines could eventually fill labor gaps in factories, warehouses and homes.
The pace did not hold through the second half. Funding totals for the third quarter came in well below the average for the first six months, and investors began to question valuations and the ability of humanoid makers to sell products in meaningful volume. The change in mood is typical of the way a new technology moves from a stage in which almost any plausible team can raise money to a stage in which backers ask for evidence of customers and revenue.
Counts of deals and totals differ between data providers because they define embodied AI differently, so the figures cited here should be read as indicative of scale and direction rather than as exact measures.
Valuation scrutiny has a practical meaning for a young company. A start-up that raised money at a high price in the first half now has to show growth strong enough to justify that price when it seeks its next round. Companies that fall short may have to accept lower valuations, merge with others or cut spending, and this process of sorting is a normal feature of a maturing market, not necessarily a sign that the underlying technology has failed.
4. Why investors are moving from bodies to brains
The reasoning behind the pivot is straightforward. If many companies can build a robot that walks, lifts and balances, the hardware itself becomes less of a differentiator. What separates a product that can be deployed from one that is a demonstration is the ability to handle varied tasks without being reprogrammed for each one.
That ability comes from the AI brain. A robot with a strong model can learn a new task from a limited number of examples, recognize objects it has not seen before and recover when a step fails. For a buyer, this reduces the cost of adapting a machine to a new setting, which is often the largest barrier to adoption. It is also the area where the practical value of a robot is most likely to be decided, and so it is where a growing share of the new money is being directed.
The emphasis, in the words of the industry commentary, is on practical applications over technological spectacle. A humanoid that performs a flashy routine on a stage proves little about whether it can complete a shift of repetitive work in a warehouse.
There is also a question of economics. A humanoid robot has to cost less over its working life than the labor it replaces or supplements, after counting purchase price, maintenance, energy and the time needed to set it up. Better software improves that calculation by raising the share of working hours in which the machine is productive, and by lowering the engineering effort required at each new site.
For investors, this changes the questions asked during due diligence. Rather than asking how many units a company can build, they ask how many tasks its model can perform reliably, how quickly it can be adapted and how many paying customers have placed repeat orders. These are harder metrics to inflate, and they favor companies with real deployments.
5. Unitree and AgiBot as test cases
Two companies are treated as the main indicators of whether embodied AI is becoming a real business. Unitree Robotics and AgiBot both reported notable revenue growth during 2026, which supports the argument that there is paying demand for the machines, at least among early customers such as research institutions, educators and industrial pilots.
Their results matter beyond their own fortunes. Sales that grow steadily give other companies a benchmark and give investors a basis for valuing firms that have not yet reached commercial scale. If revenue at the leaders stalls, the case for the wider sector weakens, and a reassessment of valuations across the field would be likely.
Revenue growth at the two leaders should be interpreted with care. Early customers often buy small numbers of machines for research, demonstration or limited pilots, and such orders do not always lead to large recurring contracts. The more meaningful signal will be whether buyers return to place larger orders after seeing how the machines perform in everyday conditions.
6. The August showcase in Beijing
The change in emphasis was visible at the World Robot Conference held in Beijing in August 2026. Shenzhen-based X Square Robot demonstrated humanoid machines performing household chores, built around its WALL-B model, which pairs AI software with hardware aimed at logistics and home use.
Household tasks are a demanding test, since homes are unstructured and every one is arranged differently. Showing a robot carrying out such work illustrates the claim that software intelligence, more than mechanical design, is the factor that will decide which products are adopted.
7. How China compares with its rivals
Estimates put China at between 80% and 97% of global humanoid robot shipments across 2025 and 2026, a wide range that reflects the difficulty of tracking a young market. By any of those measures, Chinese firms lead in the number of machines delivered.
The competitive picture is more nuanced when software is considered. United States companies have concentrated largely on foundational AI models, while Chinese firms bring manufacturing depth and access to large volumes of operational data. Those strengths could help China challenge American rivals as the two approaches converge, since a model improves with real-world use and large-scale manufacturing makes real-world use possible.
Another factor is the scale of the domestic market. China has a large manufacturing base in which robots are already common in the form of fixed industrial arms, and that base offers many sites where more flexible machines could be tried. Access to those sites gives developers a steady supply of examples of real work, which feeds back into better software.
8. What to watch next
The coming quarters will show whether the cooling in funding is a pause or a reset. Signals worth following include whether the largest makers keep growing revenue, whether start-ups that raised money in the first half can turn prototypes into recurring sales, and whether investors reward companies that emphasize software over those that emphasize hardware.
For the wider technology industry, the sector is also a gauge of how quickly artificial intelligence can leave the screen and enter physical work. A market that moves from funding everything to scrutinizing business models is, by most accounts, entering a more mature phase, and the next set of results from the leading companies will indicate how that phase is going.
A final point concerns timing. Hardware iterations take months, while software can be updated continuously, so a robot already installed at a customer's site can improve after purchase if its model is upgraded. Companies that can deliver such improvements steadily may build loyalty among their customers in a way that pure hardware sellers cannot.