China’s AI-driven robotics boom faces IPO reality check

(By Caixin journalist Du Zhihang and Han Wei)

By combining artificial intelligence with robotics to build machines capable of navigating the physical world you arrive at a sleek new buzzword: embodied AI. In China, this concept has triggered an investment boom, emerging as one of the country’s hottest venture capital themes.

Over the past two years, some 370 startups have entered the field and by July, about 50 were pursuing listings in Hong Kong or the Chinese mainland. 

Unitree Robotics, an industry leader preparing to list on Shanghai’s tech-focused STAR Market, is expected to debut with a valuation of around 40 billion RMB (US$6 billion), with many investors betting it will eventually surpass 100 billion RMB.

At least five embodied AI companies are now valued at over 20 billion RMB, and six at more than 10 billion RMB. Several have doubled in value within months.

The appeal is straightforward: if humanoid and other intelligent robots can achieve shipment volumes comparable to the smartphone while selling at prices closer to those of passenger vehicles, the sector could become one of the largest commercial opportunities in modern history.

“Private and public markets alike are exceptionally bullish about this sector right now, primarily because it may be the only industry that combines the scale of two massive markets,” said Alex Zhou, managing partner at Qiming Venture Partners.

That potential helps explain the rush to go public. Startups are competing to list early, hoping to benefit from scarcity before the field becomes crowded. 

“When the first one or two companies in a massive new sector go public, a lack of alternatives allows them to reap oversized capital dividends, pushing stock prices and valuations far beyond conventional logic,” Zhou said.

Yet behind the lofty valuations and claims of record production lies a problem: what tasks can humanoid robots perform that are clearly useful on a commercial scale?

That question is about to be tested. A wave of initial public offerings — and the discipline of quarterly earnings that follows — will soon oblige companies to show that they can turn flashy demonstrations into viable, profitable businesses.

The IPO stampede

Unitree is poised to become the first embodied AI company to list on China’s A-share market, potentially setting a valuation benchmark for the sector. Its path to market was fast-tracked by regulators, with the listing process taking less than a year.

Other companies are close behind. Deep Robotics and Leju Robotics are awaiting approval for A-share listings, while Shanghai-based AgiBot, valued at more than 20 billion RMB, has announced plans for a Hong Kong IPO.

Some 30 to 50 similar companies are also considering Hong Kong listings, according to market participants, in the hope of capturing early investor enthusiasm before the window narrows. They include top-tier startups such as Galbot, AI² Robotics, Galaxea AI and LimX Dynamics. At least three have already filed confidentially with the Hong Kong stock exchange, according to people familiar with the matter. One fund manager said the city’s first embodied AI offering could come as early as late August.

Cao Wei, a partner at Lanchi Ventures, told Caixin the embodied AI sector could eventually produce hundreds of listed companies, just as innovative drugs and smart manufacturing have done. But he said only five to ten companies were likely to emerge as dominant players.

“If you don’t go public, you might be out of the game,” an executive at a humanoid robot startup said. With the technology still too immature for large-scale commercialisation, such companies need public capital to endure a long and expensive research-and-development cycle.

The financial profiles of would-be issuers differ sharply. Unitree, which started with quadruped robots and was one of the earliest companies to commercialise that category, reported a net profit of 278 million RMB on revenue of 1.69 billion RMB in 2025, helped by growing demand and more uses for humanoid robots. 

Rivals such as Deep Robotics and Leju, by contrast, rely more heavily on government subsidies and state-backed contracts. Leju, which still makes a loss, generated nearly 45% of the revenue for its flagship product from government-led data centres.

Much now hinges on Unitree’s market performance after listing. Some investors argue that if its market capitalisation rises to 200 billion RMB or 300 billion RMB, private market enthusiasm could persist for several more months. If it fails to reach 100 billion RMB — equal, by one estimate, to roughly 100 times earnings — investors may begin marking down valuations across the sector, potentially pushing some unlisted startups into funding stress.

Illusion of mass production

The market’s mix of excitement and unease reflects a basic reality: demand for robots may be real, but the technology is still far from ready to meet it.

That tension has left the industry trapped between production milestones and commercial limits.

AgiBot recently said its 15,000th robot had rolled off the assembly line, calling it a global production record. Unitree and UBTech have also touted production capacity in the tens of thousands.

Even so, executives have been strikingly candid about the deployment challenge. In a regulatory filing, Unitree said high-end, general-purpose robots have yet to achieve large-scale commercialisation. Gao Jiyang, chief executive of Galaxea AI, was blunter, saying no robotics company is currently operating effectively in real productive environments and that chasing sales too aggressively can simply create more liabilities.

“A robot’s appearance on a gala stage, in a factory or in a home is not, by itself, proof of commercially meaningful adoption,” one industry source said.

Zhang Yaqin, dean of Tsinghua University’s Institute for Artificial Intelligence, said that global robot production in 2025 was only 20,000 units, negligible beside the scale of the auto industry. 

Today’s robots, he said, have achieved only the first of three necessary cognitive layers: completing specific, standardised tasks with limited flexibility. They still lack continuous autonomous learning and robust cognition across a number of scenarios.

So what can these machines actually do? The market is still searching for a convincing answer. Han Fengtao, founder and chief executive of Spirit AI, recently offered a blunt assessment, saying embodied AI models have barely “hatched” and possess the intelligence of a one-year-old.

“If a perfect robot is a 100 on a 100-point scale, the most mature industrial robotic arms today score about 50, wheeled robots 40 and four-legged robotic dogs 30,” Han said. “Bipedal humanoid robots score only 15, and dexterous hands a mere 5. As for the AI software powering them, that is lingering at about 3.”

The main bottleneck is data. Unlike large language models, which are trained on vast quantities of internet text, robotics data is fragmented and expensive to collect. 

Moving from simulation to the noisy reality of daily life remains a formidable challenge. Many eye-catching demonstrations — from sorting parcels to cooking meals — are still confined to controlled lab settings or rely heavily on human teleoperation. Adapting a robot to a new factory task can require months of coding and calibration, driving deployment costs prohibitively high.

Wang Hao, co-founder and chief technology officer of X Square Robot, said moving a robot from a successful lab benchmark to reliable operation in a messy real-world setting requires substantial capital just to achieve a baseline task-success rate.

Chasing the world model

To address these limitations, the sector is changing its technological narrative. Until recently, Vision-Language-Action, or VLA, models — which allow robots to learn by imitating human demonstrations — were a favourite among venture capitalists. But as VLA systems proved difficult to generalise and costly to deploy, a new term gained momentum: the world model.

Promoted by prominent AI researchers including Fei-Fei Li and Yann LeCun, world models aim to capture the physical rules of reality — such as gravity, motion and spatial relationships — so that robots can better predict and respond to changes in their surroundings. Chinese startups have quickly embraced the language as they seek fresh funding.

Since the start of 2026, a growing list of Chinese companies, including X Square Robot, Astribot and AgiBot, have announced world-model initiatives.

Investors have responded quickly. On 18 June, Manifold AI, a Beijing-based embodied AI startup built around world models, said it had completed six funding rounds within a year of its launch, with pre-A financing totalling nearly 1 billion RMB. Days earlier, GigaAI, another world-model company, said it had raised 1 billion RMB in a B2 round, bringing fundraising over the previous three months to 3.5 billion RMB. ACE Robotics, backed by SenseTime, also said in June that it had raised several hundred million dollars in the first half of the year.

Yet, some researchers remain sceptical. Fu Zipeng, a computer science researcher at Stanford University’s AI Lab, said “world model” functions in part as a fundraising label, while the underlying technical logic remains broadly similar to existing robotics approaches. Others say the industry increasingly views world models not as a replacement for VLA systems, but as a complement.

Even some of the concept’s strongest advocates have urged caution. In a June explainer, Fei-Fei Li said that while demonstrations of VLA and world-action models in embodied AI have been striking, none has been truly validated in real-world conditions, limiting practical value. At a recent World AI Conference forum in Shanghai, speakers from AgiBot, Physical Intelligence and Tencent likewise argued that VLA and world models are more likely to converge than compete directly.

China’s hardware advantage

Despite the technological bottlenecks, global capital continues to flow into Chinese embodied AI, drawn by the country’s manufacturing depth.

While the US still holds an edge in core model algorithms and computing power, China offers a complete hardware supply chain. Roughly 90% of the 1,200 components needed for a humanoid robot come from the Yangtze River Delta and Pearl River Delta, according to industry estimates. The cost of collecting real-world data in China is also about a seventh of that in the US, giving Chinese companies an important training advantage.

Still, fears of a bubble are growing. Some early investors have already begun cashing out. Rumours have circulated of inflated funding rounds in which valuations are marked up on paper without equivalent capital changing hands, allowing companies to ratchet up valuations.

As a market test approaches — one that some analysts expect to run from late 2026 into 2027 after Unitree’s IPO — the sector could face a sharp shakeout. 

In private markets, venture capital strategies are already diverging. Some investors are moving further up the supply chain, betting on the underlying hardware of the embodied AI boom, including edge-computing chips and robotic-vision systems. Others are moving away from general-purpose humanoids toward highly specialised robots that may offer a faster route to commercial deployment.

Qu Yunxu contributed to the story.

This article was first published by Caixin Global as “Cover Story: China’s AI-Driven Robotics Boom Faces IPO Reality Check”. Caixin Global is one of the most respected sources for macroeconomic, financial and business news and information about China.

Related: Is China a threat, or just out-engineering everyone? | How China’s changing the brain-computer interface race