Goldman Sachs Insight: China’s Humanoid Robots Shift to Dual-Model Drive, Real Breakthrough Expected by 2027

China’s embodied intelligence and humanoid robot industry is at a critical inflection point in its technological roadmap. According to the latest report released by a Goldman Sachs analyst team after recently visiting 14 representative Chinese robotics firms, the industry consensus has rapidly shifted from a singular Vision-Language-Action (VLA) model toward a deep integration of VLA or Vision-Tactile-Language-Action (VTLA) models with a World Model, forming a multimodal AI stack. This transition not only redefines the technological barrier to entry but also paints a clearer commercialization timeline for the investment community.

Goldman Sachs analyst Jacqueline Du pointed out in the report that model parameter scales are expanding rapidly, climbing from the early billions toward a massive range of 40 billion to 80 billion parameters. In this new architecture, the World Model no longer exists independently but serves as a functional layer alongside the action model, predicting the next state and validating actions before execution to enhance the robot’s planning capabilities and operational stability in real-world environments. Companies including Galaxea, Galbot, Spirit AI, and One Robotics have all clearly identified this combination as the main thrust of their next-phase R&D.

However, the evolution of technical architecture has not resolved all issues. The report bluntly states that high-quality, multi-dimensional real-world data remains the single largest bottleneck hindering the large-scale deployment of humanoid robots. The industry’s focus has shifted from a vague debate over “data recipes” to how to build a scalable, human-centric data acquisition framework. Two distinctly different paths have emerged for data collection: companies represented by PaXini lean toward building centralized data factories with government support—PaXini currently operates five such factories across China—while firms like Galaxea and Spirit AI are betting on decentralized deployment loops, accumulating data through already deployed systems, VR devices, and client-side collection.

Notably, data itself is becoming a significant revenue stream. Multiple companies expect data-related income to account for a considerably larger share of total revenue by 2026. Among them, UBTech anticipates that government demand for data factories will remain strong, which not only supports its revenue but also accelerates its own data accumulation, creating a virtuous cycle between business and technology.

Regarding commercialization progress, the Goldman Sachs report reveals the industry’s pragmatic attitude. Although application scenarios are gradually expanding into industrial and logistics fields, the vast majority of projects currently remain in the proof-of-concept (POC) stage. Companies typically need to undergo 3 to 6 months of POC, small-batch testing of fewer than 50 units per batch, and a validation period of up to about 12 months before entering a pilot deployment scale of roughly 50 to 100 units per customer. The broad market expectation is that large-scale commercialization will only truly materialize between 2027 and 2029, based on deployable models and after accumulating tens of millions of hours of high-quality data.

On the hardware side, the market is also showing strong pragmatism. Constrained by model capabilities and cost considerations, most manufacturers currently favor a combination of a “wheeled chassis with two- to three-finger grippers.” Goldman Sachs believes this form factor is already sufficient to cover 70% to 90% of industrial application scenarios, while the ultimate form of bipedal locomotion with five-fingered dexterous hands still needs more time. In the logistics sector, companies like Geek+ emphasize a “scenario-first” philosophy, advocating for breaking down complex tasks into clearly bounded sub-tasks and prioritizing reliability over generality at the current stage.

This robotics boom is also fermenting simultaneously in the capital markets and the supply chain. Taiwan’s Shing Hing Industrial (1558.TW) recently approved a cash dividend of NT$5 per share at its shareholders’ meeting, with a payout ratio as high as 91%, and clearly outlined its dual-axis development strategy. Shing Hing’s subsidiary Ulong (2233.TW) has independently developed its “TUF ONE” robotic joint reducer module, which has successfully entered the global humanoid robot supply chain, advancing from a precision component supplier to a modular system application solution provider and becoming a key component concept stock under market scrutiny.

In other segments of the industry chain, companies are also establishing competitive advantages. Linkerbot stated that it has captured a leading share in the global market for high-degree-of-freedom dexterous hands and has achieved pricing significantly lower than overseas competitors through self-developed joint modules. Mech-Mind focuses on 3D vision systems for industrial scenarios, with core customers concentrated in the automotive and battery manufacturing sectors. In the traditional industrial robot space, Estun Automation’s management emphasized that the company’s strategy has shifted significantly from purely pursuing market share and shipment volume toward improving product mix, profitability, and growth quality in response to increasingly fierce price competition within China.

Although large-scale commercialization still faces challenges from data scarcity and long validation cycles in the short term, Goldman Sachs remains highly optimistic about the sector’s long-term investment prospects. The report concludes that advancements in multimodal AI stacks and the establishment of sophisticated data collection systems indicate that widespread application is within reach. However, the analyst also cautioned that investors need to be patient; whether companies can maintain quality stability and continuously reduce costs during the transition from POC to large-scale commercialization will be the critical milestone determining success or failure.