China’s rush to build humanoid robots has moved from optimism to triage. More than a hundred companies entered the race after the spillover of large multimodal AI models, and three years of intense investment and development have produced a stark split: a small first tier accumulating financing and pilot orders, and a long tail of firms struggling to commercialise products.
Investor enthusiasm peaked in 2025, when the sector saw roughly 190 financing events worth about 27 billion yuan, but that momentum has revealed a brutal reality. Universities were the earliest buyers, using robots for research and secondary development, yet throughout 2025 procurement shifted towards industrial clients testing robots on factory floors. That transition has produced a surge in orders for a handful of vendors while leaving many companies that lack scalable solutions exposed.
Market researchers and industry analysts say the defining constraint is not actuators or balance systems alone but the robots’ “brain” — the AI stack that enables general, adaptable behaviour. Firms are deploying visual-language-action (VLA) models that were not designed specifically for embodied agents; high‑quality embodied datasets and purpose-built large models remain scarce. Developing such foundational models is capital intensive and is consolidating in the hands of major tech groups, leaving smaller roboticists dependent on outside breakthroughs.
The split is visible in procurement and shipment figures. Omdia ranked Zhiyuan robot maker first in global deliveries in 2025 with some 5,168 units, largely to industrial customers for repetitive tasks. Research institutes place seven companies in a first tier and a larger second tier of hopefuls, while observers increasingly warn that a substantial cohort of companies — described bluntly as “already not viable” — are merely keeping the lights on without sustainable orders or follow-on funding.
Investment bankers and analysts expect a wave of exits in 2026 as liquidity tightens and clients demand measurable returns. UBS projects global humanoid shipments of about 30,000 units in 2026 — up from roughly 13,000 in 2025 but still modest — and foresees shipments reaching 150,000 by 2030 and 1 million by 2035. Those growth curves assume breakthroughs on the AI front; without a specialised “brain,” mass-market, multipurpose home robots remain a distant prospect.
Hardware gaps persist — dexterous hands, reliable actuators and power efficiency are all works in progress — but the deeper bottleneck is software and data. Robots need long, realistic training environments and extensive multimodal, embodied datasets to acquire robust perception, manipulation and multi‑turn task planning abilities. Building those datasets and training the models to generalise across messy real-world conditions will be expensive and time-consuming, which is why many roboticists are prioritising “small‑brain” problems like locomotion and manipulation while waiting for base models to mature or be open‑sourced.
The commercial implication is that humanoid robotics will follow an uneven industrialisation path where some vendors win by tightly tailoring machines to constrained industrial tasks, while the dream of a general-purpose household humanoid — the industry’s so-called “electric-vehicle moment” — is unlikely within five years. Companies that chase low-barrier applications such as simple guides or tour robots risk commoditisation and fierce price competition, whereas firms that secure deep industrial pilots and recurring revenue streams have a chance to scale.
For international audiences, the Chinese experience offers a cautionary case about AI-driven hardware markets: technological hype and capital can create many contenders, but the economics of building specialised AI brains and the demands of durable field deployment will narrow the field. The outcome matters beyond investors and engineers: if a handful of Chinese firms crack the embodied-AI problem, they could reshape industrial automation, global supply‑chain robotics and the competitive landscape for AI-driven physical systems. If not, the sector will consolidate into fewer, better-capitalised players and slower, task‑specific adoption.