Unitree Robotics has filed for a $620 million IPO on Shanghai's STAR Market, offering a rare window into the finances of a profitable robot maker. The prospectus reveals a stark reality behind the hype: 5,500 humanoid robots shipped in 2025, making it the global leader, yet 74% were sold to universities and research institutions. Only 9% went to industrial applications.
How Unitree Hit 5,500 Units
Founded in 2016 in Hangzhou, founder Wang Xingxing—a self-taught roboticist who built his first quadruped in his apartment—now leads a 480-person firm with about 175 R&D staff. Product lines include quadrupeds (Go2, B2, A2) and humanoids (H1, H2, G1, R1). The G1, 1.32 m tall and 35 kg, is the viral star.
Two years ago Unitree was essentially a robot dog company: humanoids accounted for just 1.9% of revenue in 2023. By the first three quarters of 2025, humanoids generated over half of core revenue. Aggressive marketing helped: two consecutive appearances on CCTV's Spring Festival Gala, the world's most-watched TV show, and a stage appearance with Jensen Huang at GTC 2024.
Shipment numbers are impressive. Approximately 5,500 humanoid units shipped in 2025, the highest globally. China's AGIBOT came closest; U.S. firms like Figure AI and Agility Robotics shipped only hundreds. The five-year target is 75,000 humanoids and 115,000 quadrupeds annually—14 times current humanoid output, underscoring how early the industry is.
Who's Buying: 74% Academic, 17% Showcase, 9% Industrial
The prospectus splits buyers into three categories: scientific/education, commercial/consumer, and industrial. Academic institutions bought 74% of humanoid units, serving as Unitree's backbone since 2022. Commercial/consumer accounts for 17%; buyers mostly use robots for "display"—in retail shops, tourist spots, performances, and exhibitions. Consumer revenue grew nearly 4x year-over-year in the first nine months of 2025, but from a tiny base. A $25,000 humanoid robot today functions more like a promotional gimmick outside a Shenzhen store.
Industrial use accounts for just 9% of humanoid shipments. Unitree admits technology is not mature enough for broad deployment. Within this 9%, 50%–70% go to corporate reception and guided tours, meaning only 3%–4% of humanoid shipments actually perform real work like inspection.
Quadrupeds fare better: about one-third of revenue from scientific, over 40% from commercial, and the rest from industrial. Customers include State Grid, China Southern Power Grid, Sinopec, PetroChina, Baowu, and JD.com—using quadrupeds for real inspection tasks in chemical plants, substations, coal mines, and pipelines.
Vertical Integration Pushes Gross Margin Near 60%
Unitree designs and manufactures most key components in-house: motors, reducers, encoders, joint modules, dexterous hands, lidars, and cameras. Outsourced parts account for only 14%–18% of total cost. Quadruped unit manufacturing cost dropped from ~$3,300 in 2022 to ~$1,800 by mid-2025, a 46% decline; humanoid costs fell from ~$10,800 to ~$9,200 over the same period.
Despite falling average selling prices, gross margins expanded from over 40% in 2022–2023 to nearly 60% in 2025—largely thanks to vertical integration. Revenue grew from $58 million in 2024 to an estimated $252 million in 2025, up 335%. The company achieved GAAP profitability in 2024 with an ~18% margin (nearly 35% adjusted). IPO target valuation is around $6–7 billion.
$300 Million Bet on Embodied AI Models
Unitree plans to spend nearly half its IPO proceeds on software. Of the $620 million, about $300 million will go to AI model training over three years—roughly $100 million annually into what it calls "embodied large models." The prospectus describes two parallel architectures: VLA (Vision-Language-Action), mapping visual and language inputs directly to motor commands for generalization without hand-coded instructions; and WMA (World Model + Action), which builds an internal simulation of physical reality so the robot predicts outcomes before acting, rather than learning purely through trial and error. Initial versions have been released: UnifoLM-WMA-0 (open-sourced September 2025) and UnifoLM-VLA-0 (January 2026).
Unitree understands its hardware lead is real, but lasting advantage may depend on controlling the model layer—the system that decides what a robot does and how it moves. The software push is also a hedge against commoditization: if motors and joint modules eventually become standard components like EV batteries, defensibility shifts to the model layer.