China has become the world's dominant manufacturer of humanoid robot bodies, but the machines still lack the artificial intelligence to work reliably without constant human oversight. Unitree Robotics' Shanghai IPO in August 2026 raised roughly $900 million at a $9 billion valuation, then surged more than 460% on its first trading day, signaling massive investor confidence in Chinese robotics. Yet behind the stock surge lies an uncomfortable truth: the robots being deployed by major manufacturers like BYD and Foxconn achieve only 30% to 50% of human productivity on constrained factory tasks.

Why Is China Dominating Robot Manufacturing?

Chinese companies accounted for approximately 90% of global humanoid robot shipments in the first half of 2026, with some analysts putting the figure even higher at over 97%. Unitree alone shipped more than 5,500 humanoid robots in 2025, capturing a 32.4% global market share, while its cumulative quadruped robot shipments topped 33,000 units. In the first half of 2026, the company's revenue reached about 1.15 billion yuan, up 48.54% from the same period a year earlier.

The scale is staggering. China's Ministry of Industry and Information Technology set an official goal of deploying more than 10,000 humanoid robots into commercial use by the end of 2026 across manufacturing, logistics, retail, and healthcare. The government expects humanoid robot production to exceed 100,000 units in 2026 alone. Beijing is backing this ambition with dedicated factories, training bases, and industrial parks designed to accelerate the transition from robotics demos to volume manufacturing.

Other Chinese firms are moving at similar speed. UBTech, AgiBot, Fourier Intelligence, and Galbot have all begun deploying robots into real-world settings, from pharmacy shelves to factory floors. AgiBot recently reached a milestone of 10,000 humanoid robots in production, demonstrating that China has shifted from prototype stages into genuine manufacturing scale.

What's the Actual Productivity Gap?

The headline numbers mask a critical limitation. UBTech's Walker robots, deployed with BYD and Foxconn, operate at only 30% to 50% of human worker productivity on constrained tasks such as box stacking and basic inspection. This is not outright failure; it represents an early-stage pilot wearing an inflated valuation multiple. The robots can perform narrow, repetitive tasks in controlled environments, but they struggle with the unpredictability of real-world work.

Galbot's G1 robot offers a clearer picture of what current humanoid robots can actually do. In Beijing's Haidian district, the robot finds and packs medicines in licensed pharmacy deployments, checking each product along the way. Its shelf-picking success rate exceeds 95% in some company-reported figures. The critical detail: the robot is not running an entire pharmacy like a human worker would. It is performing a single, well-defined task in a managed space with known products on known shelves. That narrow focus is what enables high success rates.

How to Understand the Hardware-Software Split?

• Hardware Dominance: Chinese manufacturers have solved the mechanical problem of building robot bodies that can stand, walk, wave, and dance reliably. Unitree's Superman robot claims to clear a 2-meter standing jump and reach 12.66 meters per second, edging past Usain Bolt's fastest recorded stride, though these claims remain unverified by independent sources.
• Software Bottleneck: Vision-language-action models, which combine camera input, natural language understanding, and physical motion into real-time decisions, lack the training data needed for true autonomy. There is no internet-scale pile of clean robot experience available to be scraped and learned from, unlike the vast text datasets that power large language models.
• Real-World Unpredictability: A robot must learn what happens when a packet slips, a box sits at the wrong angle, or a shelf layout changes. These edge cases cannot be easily simulated or predicted; they require actual experience in varied environments.

The gap between hardware capability and software intelligence is so pronounced that even Selina Xu, China and AI policy lead at Eric Schmidt's office, has noted that humanoids remain far from true autonomy because the hardware is ahead of the software.

Who Is Trying to Solve the Brain Problem?

Nvidia has identified the opening. The company is pushing deeper into physical AI with specialized chips and software models such as GR00T and Cosmos, while working with Chinese robotics customers despite US export limits on its most powerful AI chips. This creates an unusual dynamic: China dominates the robot body market, while Nvidia seeks to own the development stack that helps those bodies act with greater intelligence.

Tesla and SoftBank are pursuing the same opportunity from a different angle. Tesla stated in its second-quarter 2026 filing that it is investing in large-scale Optimus production, with 2026 capital spending exceeding $25 billion across robotaxi, Optimus, semiconductor, solar, and AI compute initiatives. SoftBank's Masayoshi Son told CNBC that physical AI and robotics represent the next trillion-dollar opportunity.

"A $50 billion robot company with weak autonomy is still a hardware company with a very expensive promise attached," the analysis notes.

Startup Fortune reporting on Unitree's valuation and capabilities

The real test ahead is not whether companies can build impressive robot demos or stage successful stock debuts. Unitree's IPO proves that China has made the humanoid body investable. The harder, duller challenge is whether those machines can keep working reliably after the cameras leave and the initial excitement fades. That requires solving the software and autonomy problem, and that race is far from decided.