XPeng Robotics Raises Over $900M, Sets Record in China's Embodied AI

According to a Leiphone report, XPeng plans to invest more than 2 billion yuan in robotics this year, allocating funds to R&D personnel, high-quality data collection, and physical AI model training. The company has also set up a dedicated 10,000-card compute cluster for robot AI training, a scale rarely seen among domestic embodied intelligence startups, which typically operate hundreds to thousands of cards.

XPeng's robotics roadmap shows that the humanoid robot IRON is scheduled to enter mass production by the end of 2026. The company expects to generate meaningful external customer revenue starting from the second quarter of 2027. The first application scenarios will be XPeng's own stores, parks, and factories, where robots will perform real tasks such as reception, product introduction, and guiding customers. XPeng aims to have IRON handle most of a salesperson's duties in its stores, though it will not be able to close deals or conduct test drives.

The financing will be used for R&D, physical AI model training, data collection, mass production facilities, and global commercialization. A person close to XPeng's management told Leiphone that the robotics unit currently has no IPO plan. Management believes the priority is to push the robot to mass production and real commercial environments before any capital operation. The robotics business is expected to be gradually spun off over the next 18 months, with an independent entity holding assets, IP, and personnel.

On the technology side, XPeng has built a 16,000-square-meter data collection center and designed its own data acquisition equipment. Its data system covers four types: teleoperation demonstration data, first-person video data (UMI), simulation data from world models, and internet human videos. The company uses large-scale, low-cost data for pre-training and high-quality real-world data for post-training. XPeng's CEO He Xiaopeng has said that robots need at least three categories of models at different speeds: a very fast model at 100 frames per second or higher for motion control, a medium-speed model at 10-20 frames per second similar to autonomous driving VLA, and a slow reasoning model at about 1 frame per second. The robot also needs dedicated whole-body motion control and safety models.

IRON is equipped with three self-developed Turing AI chips with a total effective computing power of 2,250 TOPS. It has 76 degrees of freedom, including 21 in each dexterous hand, which uses a direct-drive scheme rather than the tendon-driven approach. More than 85% of IRON's supply chain partners overlap with XPeng's automotive business.

Although the valuation is aggressive, XPeng has not given a specific break-even sales volume. Management said it is too early to predict profitability, but it expects robot hardware gross margin to be higher than autos, and AI models and software services to contribute even higher margins. The company believes robotics can achieve profitability faster than autos once production ramps up, given lower capital expenditure.

Whether IRON can turn into a real revenue generator in 2027 will determine if the $900 million financing marks the beginning of a story or just a capital pricing exercise, the Leiphone report said.