At the September 17 launch of the Xpeng G9L in Beijing's 798 art district, CEO He Xiaopeng opened the event not with the flagship SUV, but with the IRON humanoid robot. On screen, IRON walked off an automated assembly line, prompting Xpeng to claim a world first: a production line with over 80 percent automation that makes robots using robots. The company's robotics arm, Pengxing Intelligence, has completed a first funding round of over $900 million, valuing it at 43 billion yuan (about $6.1 billion), more than half of Xpeng's market capitalization, according to HUXIU.
Yet the capital markets are turning against humanoid robots. After two years of exuberant funding, investors now demand real orders, yield rates, and repeat purchases. Unitree Robotics, a leading player, saw its stock nearly halve after a stellar debut on the STAR Market, where its IPO price-earnings ratio reached 219 times. Industry insiders say Unitree's fundamentals exposed a harsh truth: fewer than 10 percent of humanoid robots shipped actually work in industrial settings; most go to universities and trade-show demos.
From demo to product
Xpeng is among the fastest carmakers to enter humanoid robotics, leveraging its automotive manufacturing system. He Xiaopeng told media that developing a humanoid robot is about 20 times harder than building a car, requiring new solutions across hardware, software, supply chain, and manufacturing. Xpeng's approach is to transplant its car-making playbook: automotive-grade quality control, automated lines, and 85 percent of its supply chain overlapping with its car business. Its VLA (vision-language-action) and VLM (vision-language model) large models are shared between cars and robots.
This strategy is directionally correct, as the industry's biggest weakness has been a lack of mature manufacturing. But an 85 percent supply-chain overlap does not mean 85 percent of parts are ready to use. Motors, batteries, and thermal management can be reused, but dexterous hands, micro actuators, and flexible skins have no automotive-grade equivalent—these are the bottleneck 15 percent that require manual assembly.
He compared IRON's current stage to Xpeng's G3 launch in 2018, but noted that it took eight years to achieve stable mass delivery of that car. For IRON, Xpeng targets production start by end-December 2026 and global delivery in 2027, leaving only about 100 days for engineering validation.
Three big challenges
First, complexity: IRON has 76 degrees of freedom, with 21 in each hand. A minor change, like adjusting the head camera's field of view, cascades through all motion parameters. Tesla's Optimus has repeatedly delayed production, and the industry's yield curve is steep—improving yield by one percentage point in small-batch production requires enormous time and money.
Second, cost: the dexterous hand is both the core component for real tasks and a cost black hole. Industry data shows the hand accounts for about 25 percent of hardware cost, joint modules about 50 percent, and together with ball screws, 70 percent. No standardized solution exists yet. Xpeng's initial focus on store-guide scenarios—walking, talking, simple gestures—avoids the wear and tear of high-intensity manipulation, but future industrial or service applications will demand better hands.
Third, market: IRON is still a technology showcase with no external paying customers. Its first deployments will be in Xpeng's own stores and parks, a closed environment with limited capacity. Even if each of Xpeng's hundreds of stores takes two robots, total demand is only around a thousand units—far below what the production line can eventually output. The industry lacks proven profitable B2B scenarios; factories demand high reliability, services are fragmented, and consumers are years away.
Carmakers also face unique disadvantages. Embodied data—from hand manipulation and unstructured environments—cannot be transferred from autonomous driving data. And selling to fragmented B2B markets in services and industry requires a completely different sales, installation, and maintenance system than selling cars to consumers.
The cost math
HUXIU estimates that in small batches (under 1,000 units per year), IRON's hardware cost will be 400,000 to 600,000 yuan (about $56,300 to $84,500). At 5,000 units annually, costs could fall to 250,000 to 300,000 yuan, and only above 10,000 units could they approach 150,000 to 200,000 yuan—comparable to mainstream smart EVs. Early yields of 60 to 75 percent, versus over 95 percent for mature car lines, will inflate costs. The 43 billion yuan valuation reflects hopes for a trillion-yuan market, but those hopes must meet cold financial reality.
