Matrix, EngineAI and the humanoid rush: tens of thousands coming, but functionality lags

Chinese humanoid robots are already doing backflips, directing traffic, and even making coffee, but the market story has a glitch: the people buying them want functionality in messy real life, while many robots are still closer to expensive demos.

That tension shows up fast in the numbers and the business plans. Matrix Robotics, based in Shanghai, sells its flagship “MATRIX-3,” a 1.7-meter (nearly 5.6 feet) tall humanoid with AI-driven, finely controlled hands. It costs around $99,000 per unit. Matrix said it received roughly 1,000 orders and has made only a few hundred robots so far, while aiming to deliver 5,000 units within this year, depending on how many orders it gets. EngineAI, headquartered in Shenzhen, pitches humanoids for security and museum guide work. Its full-sized humanoids also perform routines like dancing and boxing. A basic edition costs 180,000 yuan ($26,600).

Now zoom out to the bigger stake: China is building at factory speed, while demand may not be strong enough or broad enough to justify full-scale, consistent deployment. Experts cited in the reporting argue the use cases are still limited. “The use cases of these robots are still so limited,” said Chibo Tang of the venture capital firm Gobi Partners, which invests in technology startups including robotics companies. And, crucially, “Without the demand and without that scale from the market, these companies are not able to really go into mass production.” Samm Sacks, a senior fellow at the New America think tank focused on Chinese technology, similarly framed the issue as a mismatch between capability and environment: many robots are still performative rather than functional, and they fall short in messy, unpredictable settings.

This is not happening in a vacuum. China already has the industrial ingredients that usually make new robotics waves take off: hardware production, data collection, and speed. The reporting notes that China and the United States dominate research for what Morgan Stanley estimates is a $5 trillion humanoid robots market. By some measures, the U.S. holds an upper hand in developing the “brains,” meaning the high-level computing and AI for humanoids. China leads in mass production capacity, supplies of hardware, and harvesting of data for training.

But regulators are watching the commercialization gap closely. The Ministry of Industry and Information Technology reported that China had more than 140 humanoid robot manufacturers and more than 330 models in 2025. Last year, the Chinese government publicly warned about the risk of a bubble in the industry, pointing to lagging commercialization and applications. In parallel, corporate and academic labs are buying humanoids for research, and Morgan Stanley said that in China, more than 2 billion yuan ($295 million) worth of orders in 2025 came from state-owned enterprises. Those orders were aimed at uses such as power plants, data centers, or entertainment.

That mix matters. It implies demand exists, but not necessarily demand that stresses the robots in the exact conditions that make them truly valuable: highly varied tasks, unpredictable environments, and sustained operational reliability. The reporting summarizes the economics bluntly through Sacks: humanoid robots remain expensive to produce, fragile in operation, and dependent on highly structured environments to function. She also said there is “a long way to go” before people “feel comfortable” having humanoids in homes providing care for elderly or children.

Still, China is not waiting politely in the wings. Over the past year, real-world deployment of humanoid robots in China accelerated, and executives and researchers argue the technology is improving enough to justify new bets. RoboScience founder and CEO Ye Tian, an ex-Apple engineer, said Chinese people are relatively used to rapid technology change. Lian Jye Su of Omdia expects humanoids could perform heavy-lifting and mundane tasks in warehouses, factories, and ports. And there is a strategic logic here: if robots can fill gaps where work is dangerous or repetitive, they can complement or replace labor in specific workflows.

The catch is that many factories already run repetitive single-function work with non-humanoid robotic arms. Sacks argued that the more viable commercial path may be industrial and logistics settings, not the household initially. She also noted Japan and the U.S. have humanoid startups struggling to find buyers even in industrial work settings, suggesting the buyer learning curve is real everywhere. That is why the “where this actually works first” question is so important for executives who are trying to decide whether to fund robotics integration, sign pilots, or wait for a more mature product.

China leads the volume battle, and that’s part of why the bubble warning sticks. Last year, Chinese humanoid robots accounted for around 85% globally, according to a Barclays research report. Omdia data cited in the reporting said that of the more than 13,000 humanoid robots shipped in 2025, AGIBOT and Unitree each shipped over 5,000. U.S. rivals like Figure AI and Tesla shipped a few hundred or less. Morgan Stanley expects China’s humanoid sales to more than double this year to around 28,000 units. Omdia forecasts advanced robot annual shipments could surpass 1 million units by the early 2030s.

Cost dynamics look like they could accelerate adoption, but the economics still have strings attached. Unitree said it made 1.7 billion yuan (around $250 million) in revenue last year, with a profit of over 278 million yuan ($41 million). The reported thesis from Morgan Stanley is that as production increases, costs drop. Locally made parts help Chinese robots be about 20% or more cheaper than foreign models on average. Morgan Stanley estimates the average price could fall to about $21,000 by 2050, from $46,000 last year. Some humanoid robots in China were priced below $6,000.

Yet cost is only half the problem. Data is the other half, and it takes time to scale. For humanoids to learn more than single tasks, they need data from a wide variety of scenarios with reasonable difficulty, not just idealized lab conditions. Wang Xiaogang, co-founder of SenseTime and chairman of ACE Robotics, said his company is collecting human-centric data from factories, retailing, and offices. Eric Guo, founder and CEO of Shenzhen-based AI² Robotics, said that to teach humanoids across many scenarios could take years to massively scale up. “The mass production capability in (the) robotic area is still at the very early stage,” Guo said.

Put it all together, and the strategic stakes are clear for investors, operators, and corporate planners: China can ship thousands, but buyers and regulators are asking whether those robots can do more than perform. The decision is not whether humanoids are real; it is whether functionality, reliability, and data-driven generalization will catch up before budgets tighten or the bubble thesis becomes policy. If you are allocating capital or building an operations roadmap, the playbook is about timing: separating “impressive on stage” from “reliable in the world you actually run.”