Embodied AI drew 322 funding rounds worth 93.4 billion yuan in the first half of 2026, five times the total from the same period in 2025. Buried in ITjuzi's figures is a more telling ratio: seed, angel and pre-A rounds accounted for 190 deals, or 58.94% of the total, with angel rounds alone topping 100.

Put another way, more than half the money went to companies with no product, no revenue, and in some cases not even a complete team.

Capital and talent have poured into the same direction within two years, and startups have sprouted in bulk — numerous, scattered, and early. AI tools have cut the amount of prior experience the work requires, and that lower barrier has pushed headcount up further without closing the engineering gap. With entry easier than ever, the mass-production wall standing in front of these companies is higher than ever.

Wang Zhongyuan, president of the Beijing Academy of Artificial Intelligence, has described the gap between research and commercialization as a "valley of death." The prevailing view in the industry is that, measured by technology maturity, embodied AI has not yet reached the stage of mature commercial application.

Multiple Pools of Capital Reach a "Super Consensus"

Since the start of 2026, funding for embodied AI in the primary market has come from an unusually broad front. Dollar funds, renminbi venture firms, national tech-innovation funds, local government industrial funds and corporate strategic investors — nearly every category of capital is converging on the sector.

Cao Wei, a partner at Lanchi Ventures, calls it "a super consensus among multiple pools of capital on embodied AI and robotics," something he says he has rarely seen in decades of investing.

Talent is migrating in organized blocks. "The largest group of new entrants comes from autonomous driving, and they've brought a lot of related software people with them. IT workers are now moving over too," Gu Qiang, founder of the GuYueJu robotics developer community, told reporters.

By incomplete count, nearly 40 senior executives and technology leads from China's intelligent driving sector have moved into embodied AI, spread across more than 20 startups, over 70% of which were founded in the past two years.

The consensus is not without logic behind it. China's embodied AI market has grown from roughly 213.3 billion yuan in 2018 to a projected 1.09 trillion yuan in 2026, a compound annual growth rate of more than 22%. China has the world's most complete manufacturing supply chain, the widest range of industrial use cases, and a decade of accumulated depth in robotics hardware and sourcing.

Capital is betting on two propositions: that embodied AI is a hardware-software innovation built around manufacturing, and that Chinese teams have a shot at taking that innovation to global markets.

"Look at the path of solar, lithium batteries for new energy, electric vehicles, consumer electronics — wherever Chinese manufacturing combined hardware and software, it produced a large number of global leaders," Cao Wei said.

Investor Patience Is Already Thinning

"In the early days, a robotics team was worth backing as long as it was young, learned fast, and was willing to go deep on a use case. But the bar across the industry has risen. Industrial players with systematic engineering capability, like Li Auto and XPeng, are entering in numbers, and competition has intensified sharply," Cao Wei said, adding that investment logic is being rebuilt as the sector matures.

The sector keeps accelerating while the ecosystem around it is not ready. Large numbers of teams have been pushed by the wave of capital and talent into having to answer the question of how to mass-produce, before their engineering capability is complete, their product definition has converged, or their supply chain relationships exist.

Component Prices Are Rising Before Mass Production Begins

The surge in enthusiasm has also made the startup layer far more crowded. When dozens of such teams raise early capital in 2026, start defining products and place supply chain orders at the same time, they face component and engineering-services markets that do not expand automatically just because a team has a strong résumé.

Component price increases are now rippling through the industry. Take memory chips: a founder at one robotics startup said memory was once an almost negligible line item in a robot's hardware bill of materials, "a few dollars, and now it's 40 to 50 dollars."

"Computing power, storage and the circuit boards in a full robot used to come to maybe three or four hundred yuan. The storage increase alone has doubled that," the same person said.

Storage is not the only thing climbing. Power devices, capacitors and other components that have held steady for years are joining the queue. Gu Qiang said some teams that previously specified 8GB of memory are now looking for ways to make do with 2GB.

For full-system manufacturers already shipping at stable volume, this is a cost management problem. For early-stage startups, it can determine whether a product reaches market at all. Gu Qiang said most startups have weak bargaining power and have yet to build a functioning supply chain.

Robots Are Harder Than Cars

Then comes consistency at volume. "I've heard a lot of people who used to work on cars say robots are harder than cars, and mass-producing them is harder still," said Cao Yuelong, vice president of design services, marketing and communications for Asia Pacific at Avnet.

The industry has labeled 2026 the first year of humanoid robot mass production, but a single humanoid contains hundreds or thousands of parts, and traditional assembly depends heavily on manual tuning. Minor differences in screw tightness get amplified into visible deviations in walking and grasping, and the absence of industry-wide component standards means each model has to be tuned separately. Producing consistent performance at scale is extremely difficult.

A further complication: the robotics industry's suppliers have never built robots either. "All we can do is look for experience from similar products and see whether it transfers to robots, and most of the time it doesn't transfer directly," the startup founder said.

That helps explain why a large number of seasoned founders who left big tech companies with teams in tow have still made limited progress getting products into the field. For startups built mainly on software backgrounds with incomplete engineering teams, the difficulty multiplies.

Compliance Is the Barrier You Can't See at the Demo Stage

Export compliance is another dimension almost no one considers at the demo stage, and one that turns into a hard gate at mass production and overseas launch. Interviewees consistently described overseas compliance requirements as extremely complex, and said a failed certification can force a full redesign.

Take the European Union. The Cyber Resilience Act (CRA) applies reporting obligations for vulnerabilities and serious security incidents from September 2026, with the full set of requirements taking effect in December 2027; connected robots, automation systems and edge devices largely fall within scope. The new Machinery Regulation (EU) 2023/1230 folds cybersecurity requirements into the machinery safety framework and becomes mandatory in January 2027. High-risk obligations under the AI Act are also phasing in.

In short, a connected robot entering the European market has to answer questions on machinery safety, cybersecurity and AI compliance at once — another barrier for startups.

Most embodied AI startups are staffed primarily with software and algorithm backgrounds, which means compliance work starts from near zero. Cao Yuelong said many small and mid-sized startups currently lack capability in certification, legal compliance and distribution channels, and all of them need help.

Immaturity Is the Current State, and Crossing It Becomes the Moat

None of these bottlenecks looks easy to solve. The sector's supply chain has yet to be strung together.

Missing component standards, a deeply complex supply chain, a compliance regime not built for embodied AI products, and data and hardware interfaces that each go their own way — stacked together, these make the distance between building one demo and building 10,000 units far greater than in most consumer hardware categories.

Gu Qiang said the sector will eventually need an ecosystem of close to 10 million people, which requires better matching between industry demand and the people who execute. Collaboration is moving: Avnet and GuYueJu recently signed a memorandum of understanding covering technical solution development, product deployment, joint marketing and developer ecosystem building around Physical AI and embodied intelligence.

"Embodied AI is at a critical inflection point, moving from special-purpose robots to general-purpose embodied systems, and that transition depends on deep hardware-software coordination and on the strength of the open source community," Cao Yuelong said.

The gaps in compliance and engineering capability are systemic. If startups have to close them alone, the time cost will filter out the vast majority, which makes ecosystem-level cooperation unusually important. Conversely, whoever builds compliance and engineering capability in at the product definition stage gets an early pass through a narrow door most companies will not get through.

The startup founder summed up the moment in one word: "blue ocean." Immaturity means chaos, but it also means prices for core components have not hardened and supply chain relationships have not been locked up, while the corresponding pricing structures and payment models remain untested in the market. In phones or PCs, the rules are already written and a startup can only place pieces on a board someone else drew. In embodied AI, the board is still being assembled.

"Whoever breaks through on engineering and supply chain first, inside this window where prices and definitions haven't hardened, will be first to hold real competitive advantage in the market," he said.

Hardware companies face long payback cycles, with a far longer path from R&D to mass production to user feedback than software products. The interviewees argued that companies cannot survive on fundraising alone and need to start with small business loops, validating product viability step by step and sharpening their core advantages. Moats built through long-cycle accumulation will naturally filter for genuine long-term operators, who stand to gain durable advantage in the era of physical AI.