Read a humanoid as a bill of materials and it stops being science fiction. It is roughly forty joints, each one a module: a frameless torque motor, a harmonic or planetary reducer, an encoder and a driver, bolted into a limb. Add a pair of dexterous hands, somewhere between seventeen and twenty-two degrees of freedom each, an IMU, force and torque sensing at the wrists and ankles, a battery, and two computers, one a hard real-time motion controller and one an AI compute board. The proportions are the thing worth knowing: the actuators are about half the cost of the machine, which is why the interesting suppliers in this category sell joints rather than robots.
This product breaks the pattern the rest of this atlas runs on. In earbuds the chipset is the reference design; in action cameras the sensor and ISP pairing is. Here there is nothing equivalent to badge. China had more than 140 humanoid makers and over 330 models released by the end of 2025, and almost none of them will sell you a finished platform to put your logo on the way an audio house hands over a TWS reference. What is reusable sits one layer below the robot, in the joint module, the reducer and the hand, and increasingly one layer above it, in motion stacks that the makers themselves open-source. Treat this page as a components market with some brands attached, not an ODM market.
Be clear about where the delta actually stands, because the volume leaders are not in it. Chinese vendors took about 97 percent of the roughly 19,100 humanoids shipped worldwide in the first half of 2026, but the two at the front, AgiBot and Unitree, are Shanghai and Hangzhou companies. The Pearl River Delta's claim is a different one: it holds the listed pioneer in UBTECH, a dense startup bench inside a few Nanshan blocks, the component supply in Bao'an and Dongguan, and the carmakers who know how to turn a prototype into a production line. XPeng broke ground on a 110,000 square metre robot base in Guangzhou's Tianhe district; Leju's Foshan plant with Dongfang Precision assembles a Kuavo about every thirty minutes. Shenzhen builds the body. The brain is often trained somewhere else.
The money and the policy are pointed here anyway. Shenzhen's 2025 to 2027 embodied AI action plan wants more than ten companies worth over 10 billion yuan each, a 50 billion yuan robotics fund is behind it, and the seventy-odd supply chain firms already clustered, from tactile sensors and dexterous hands through to whole machines, are meant to become 1,200 by 2027. What that buys a buyer is speed: a prototype turn that runs twelve weeks in the West runs ten to fourteen days here. What it does not buy is discrimination between houses. Ask any candidate which parts it makes and which it integrates, ask for its domestic content share, and ask for units delivered to paying customers rather than demo footage. The gap between a company that has shipped a hundred robots and one that has shipped a video is the whole diligence.
For export the gate is unusually heavy, and it is moving. A humanoid placed on the EU market is machinery: CE under the Machinery Directive 2006/42/EC until 20 January 2027, then the Machinery Regulation (EU) 2023/1230, with no grace period on either side of that date and new cybersecurity duties wherever a safety function is affected. ISO 25785-1, the standard being written for dynamically stable robots, is still in development, so the technical file leans on ISO 10218 principles plus the maker's own risk assessment. If the robot ships with an AI system, the EU AI Act sits on top of all of it, and neither framework excuses the other. The tell of a serious partner is a real technical file and a named EU importer, not a CE sticker on the chest plate.