The hardest problem in robotics is no longer teaching a machine to walk. It is teaching it to think. A global contest for humanoid robot brains is now pulling in tens of billions of dollars, and whoever wins it could own the next industrial platform.
By Blumefield | November 2, 2026
The Money Has Moved From Legs to Minds
For most of the past three years, the humanoid robot story was a story about hardware. Companies competed to show machines that could backflip, fold laundry, or stroll across a stage without toppling over. That era is ending. The frontier has shifted to software, and specifically to the humanoid robot brains that decide what a machine should do once it can physically do almost anything. The spectacle of motion has given way to a quieter and far more valuable question of cognition.
The reason is commercial. A robot that dances is a demo. A robot that can insert a wire into the same connector ten thousand times without failing is a product. Factory and logistics customers do not reward improvisation, and they run long validation cycles before they trust a machine on a line. Getting from impressive to dependable is almost entirely a problem of control software and learned behavior, which is why capital has started chasing humanoid robot brains rather than better actuators.
That shift is now showing up starkly in the funding data, and nowhere more dramatically than in China.
China's Thirteen Billion Dollar Bet
Chinese embodied-AI companies raised 93.5 billion yuan, roughly 13 billion dollars, across 322 deals in the first half of 2026, a severalfold jump over the same period a year earlier. The money is increasingly flowing not to the firms that build the most elegant limbs but to those writing the intelligence that runs them. Software-focused startups such as X Square, Galaxea, and Spirit AI each now carry valuations of around 3 billion dollars, remarkable sums for companies whose core product is essentially a model.
There are early signs the bet can pay off on the factory floor. Battery giant CATL has reported that Spirit AI machines achieved more than 99 percent success on wire-insertion tasks at one of its plants, the kind of reliability figure that turns a pilot into a purchase order. Numbers like that are what separate a research project from an industrial supplier, and they are exactly what investors are paying to see repeated across thousands of sites.
A second gold rush is forming one layer down, in the data these models learn from. On November 2, Beijing-based GenRobot AI disclosed a 100 million dollar round backed by an unusually heavy roster of strategic investors, including SAIC Motor, Baidu, Shoucheng Holdings, and Galaxy General Robotics, alongside financial backers such as GL Ventures and CITIC Goldstone. GenRobot does not build robots. It builds the recording hardware and the software platform, called Matrix, that captures human demonstrations and turns them into clean, labeled training data. In the robot economy, that is the picks-and-shovels business, and the appeal of humanoid robot brains has made it suddenly lucrative.
Why the Brain Is the Hard Part
The promise of humanoid robot brains is also the source of their difficulty. The goal across the industry is a single model that can run many different robot designs and generalize to tasks it was never explicitly programmed for, the physical-world equivalent of a large language model. But language models had the entire internet to learn from. Robots have almost nothing comparable. There is no vast, ready-made archive of a machine picking up a specific bolt from a specific bin at a specific angle.
Closing that gap is the central engineering fight, and companies are attacking it from several directions. Some hire large human workforces to teleoperate robots and generate demonstrations by hand. Others build enormous video libraries of people performing tasks, or lean on physics simulation to manufacture synthetic experience at scale. Each approach is expensive and imperfect, which is why a data-infrastructure specialist like GenRobot can raise nine figures without ever shipping a robot of its own.
The hard truth underneath the hype is that dexterity does not transfer cleanly. A model that masters one gripper, one lighting condition, or one warehouse layout often stumbles when any of those variables change. Building humanoid robot brains that stay reliable across the messy variety of the real world remains unsolved, and the firms that crack it first will have built something closer to an operating system than to a gadget.
Silicon Valley's Counterweight
China's funding surge has a mirror image in the United States, where the bet is more concentrated and the valuations are even larger. Physical Intelligence, the startup backed by Jeff Bezos and focused squarely on general-purpose robot foundation models, is in talks to raise about 1 billion dollars at a valuation near 20 billion dollars, roughly double the figure it carried only months earlier. Rivals such as Skild AI and Generalist are chasing the same prize from different angles. The American thesis is that whoever builds the best humanoid robot brains will not need to build the best robots, because every hardware maker will license the mind.
The lines between the two ecosystems are already blurring in uncomfortable ways. According to industry reporting, Meta has built a humanoid prototype in partnership with the Chinese robotics firm Fourier, and is working on a second version. For a company that has publicly framed its robotics ambitions around American technological leadership, leaning on Chinese hardware to move fast underlines how global the supply chain for robot minds and bodies has become, even as politics pulls in the opposite direction.
That tension is the backdrop to everything happening in the sector. Readers following the broader buildout can trace the money and the models through Blumefield, which has covered the physical-AI race from its earliest venture rounds.
The Chokepoints Ahead
Two obstacles could slow the whole enterprise. The first is silicon. Training and running advanced robot intelligence demands the same high-end accelerators that power frontier language models, and tightening United States export controls have left Chinese developers short of the most capable Nvidia chips. Compute scarcity does not stop the research, but it raises its cost and widens the gap between those with guaranteed chip supply and those without.
The second is the public market. Regulators and investors have grown wary of the valuations attached to companies whose revenue is still modest relative to the capital they have absorbed. A tighter listing environment, driven partly by concern about inflated numbers, could strand some of the 3 billion dollar startups now multiplying across the sector. Money raised is not the same as a business proven, and a 13 billion dollar funding half-year does not by itself make a profitable industry.
Still, the direction of travel is unmistakable. The humanoid robot race has become a contest over cognition, not choreography, and the capital flooding into humanoid robot brains on both sides of the Pacific reflects a shared conviction that the machine which can think its way through physical work will become one of the defining platforms of the decade. The bodies, it turns out, were always the easy part. For more on the companies building that future, follow Blumefield and its continuing coverage of the robot-brain race.