China’s humanoid story is mutating from “cool demo, bro” into something far more dangerous: an industrial system with production lines, data factories, and IPO prospectuses full of numbers that can be audited by people who hate you.
TrendForce is calling a 94% year-on-year output jump in 2026, Unitree is telling investors it can crank out 75,000 humanoids a year, and AgiBot is doing the classic startup move of “we shipped 10,000 units” while the rest of the world argues about whether a robot doing kung fu counts as GDP.
The news hook: the scale claims are now specific
Specific is good. Specific is also scary, because it means the hype has left the land of adjectives and wandered into the land of capacity plans, margins, and throughput claims.
- TrendForce: output growth up to 94% in 2026, with Unitree and AgiBot projected to take nearly 80% of shipments.
- Unitree: STAR Market IPO application, humanoids reportedly surpassing quadrupeds as a share of revenue, and an aspirational capacity target of 75,000 humanoids annually.
- AgiBot: 10,000th “general-purpose embodied robot” (Expedition A3) rolled out in late March, after doubling from 5,000 to 10,000 units in about three months.
What’s actually changing: the bottleneck moved from “build one” to “build the pipeline”
Humanoid robotics is basically three bottlenecks wearing a trench coat:
- Manufacturing: can you build them repeatably, with QA, traceability, and supply-chain discipline?
- Data: can you collect enough high-quality, task-relevant behavior data without turning your operation into a teleop sweatshop?
- Deployment: can you keep the thing useful for more than one camera-friendly afternoon?
China is starting to look like it’s building all three layers at once. That’s not “robots are coming.” That’s “robots are being staffed, scheduled, and measured.”
The manufacturing layer: a production line is a claim you can interrogate
China.org.cn describes an automated production line in Foshan (Dongfang Precision + Leju Robotics) with an annual capacity “over 10,000 units” and a headline throughput claim of one humanoid every 30 minutes.
It also includes the kind of details that matter if you’ve ever met a manufacturing engineer: modular architecture, quality traceability, high process automation, and tight assembly tolerances on transmission components. You can debate whether the “one every 30 minutes” rate holds in practice, but the interesting part is that someone is building the scaffolding where that debate is even meaningful.
The data layer: humanoid scale is starting to look like a data-center problem
Seoul Economic Daily quotes AgiBot describing a very modern worldview: if a robot needs ~1,000 data points per finely segmented motion, then “deployment” is partly a factory-floor question and partly a data-collection logistics question.
AgiBot’s model is explicit: large numbers of robots collecting raw data daily, distilling it into fewer hours of “high-quality” data, then scaling facilities (training centers, warehouses, data centers) to push the pipeline toward million-hour annual data volumes. That’s not how you talk when you’re selling a demo reel. That’s how you talk when you’re selling a production plan.
The Droid Brief Take
Humanoids are entering their most important era: the boring era, where claims collide with yield rates, service schedules, and whether your robot can behave on Tuesday. If you’re still debating “will humanoids happen,” you are now arguing with factories, not futurists.
But here’s the catch. “Shipped” is not “deployed,” and “produced” is not “profitable in the field.” Unitree’s prospectus numbers (via Rest of World and SCMP) suggest a real hardware business, but they also quietly admit that a big chunk of humanoids still go to research and education. That’s not a dunk, it’s context: early-scale markets often start as “labs and schools,” then get brutally selected by real customers.
What to Watch
Deployment evidence: named customers, task-level details, shift-length operation, and failure-rate reporting, not just unit counts.
Data provenance: how “high-quality data” is defined, and what human supervision actually looks like at scale.
Supply chain constraints: joint modules, dexterous hands, sensors, and compute, the parts that decide whether “75,000 units” is a capacity target or a bedtime story.
Sources
TrendForce — China’s Humanoid Robot Output to Surge 94% in 2026; Unitree and AgiBot to Capture Nearly 80% Market Share
Rest of World — The world’s largest humanoid robot maker is going public
South China Morning Post — China’s Unitree Robotics rides humanoid tide as it targets US$610m IPO
China.org.cn (Xinhua) — Automated humanoids production line in place
Seoul Economic Daily — High-Quality Real-World Data Is Key as Agibot Targets 10,000 Humanoid Shipments This Year