Unitree G1 Hits 10k–20k Units; Huixi Bets RMB 2.5B on Robot Chips
Unitree (宇树科技) reports cumulative G1 output of 10,000 to 20,000 units, while Qiyuan Robotics rolled out its 5,000th personal robot on Oct 25 at a stated cadence of one unit every 2.5 minutes. Read together, the two figures place line throughput rather than demo capability as the near-term competitive axis; Unitree's shares rising as much as 9% to a two-month high indicates equity pricing now tracks shipment disclosure more than prototype announcements.
On the compute side, Huixi Intelligence (辉羲智能) has raised RMB 2.5 billion cumulatively in 2026 and is redirecting autonomous-driving chip teams toward embodied AI silicon. Capital is concentrating at the layer with the longest qualification cycle, a supply-side wager that per-unit compute cost falls only if the volume figures above hold. Overseas, Teradyne Robotics will run Sereact's Cortex platform on Universal Robots arms for returns handling and piece picking, evidence that application integrators are buying physical AI software rather than building it.
Data infrastructure is being formalized in parallel. AgiBot (智元机器人) open-sourced AGIBOT WORLD 2026 as a real-world embodied reasoning dataset, and the Beijing-Tianjin-Hebei humanoid super factory applies full-chain SN traceability to stabilize mass production. The engineering bottleneck is shifting from manipulation success rates toward reasoning data and quality traceability, both of which are prerequisites for the volumes cited above.
Financing is diverging. HuanDong withdrew its STAR Market IPO after two rounds of exchange inquiries, a Shanghai court disclosed creditor notices for DeepBlue Robotics, and ASD reported a RMB 142 million net loss for the first three quarters with its humanoid line still early-stage. Capital is separating into compute, volume manufacturing and application integration, leaving demo-stage firms without a clear slot.
Two observation points follow. First, whether G1-class output is matched by disclosed utilization or order backlog, which would confirm demand rather than channel build. Second, whether chip-layer capital converts into shipping silicon within twelve months, the test of whether the RMB 2.5 billion is capacity or option value.