A new Goldman Sachs report argues the humanoid robot industry is shifting its focus from impressive demos to the harsh realities of commercial deployment, with high-quality data emerging as the primary bottleneck.
Back
Goldman Sachs: China's 14 robot makers see 2027 liftoff
A new report from Goldman Sachs pours cold water on short-term hype for humanoid robots, projecting that large-scale commercialization for 14 leading Chinese firms is unlikely to begin before 2027 and is critically dependent on solving a massive data bottleneck. The industry's focus has already shifted from simpler Vision-Language-Action (VLA) models to more complex, execution-oriented AI stacks that fuse VLA with world models.
"Industry discussion has already moved beyond a single VLA framework, toward a multi-modal AI stack oriented around execution," Jacqueline Du, an analyst at Goldman Sachs, said in the report released after visiting 14 Chinese robotics companies. The report notes that for these new models to become deployable, they must be trained on tens of millions of hours of high-quality, real-world data, a challenge the industry is now confronting.
The technical consensus is rapidly converging on this new hybrid architecture, where a world model acts as a functional layer to predict outcomes and verify actions before execution, enhancing real-world robustness. To power this, model parameter counts are climbing from single-digit billions into the 40 billion to 80 billion range. Despite the technical advances, the report states that most projects remain in the proof-of-concept stage, with a clear focus on industrial and logistics applications.
For investors, the report tempers immediate expectations while reinforcing long-term optimism, suggesting the key milestone is the transition from conceptual pilots to scalable, profitable deployments. The complex process of ensuring quality while lowering costs will be the central challenge for the next three to five years, ultimately determining which of the 14 firms will lead the market.
From Model Recipes to Data Architectures
The core challenge for China's robot makers is no longer just about the "recipe" for AI models, but about building the infrastructure to feed them. According to the Goldman report, the industry's focus has shifted to constructing scalable architectures that can reliably produce high-quality, multi-dimensional data from real-world interactions. This marks a significant pivot from simply debating the merits of different model types. As Dr. Yao Maoqing, president of Agibot’s embodied AI unit, noted in a recent interview, there is a "massive gap between lab demos and real-world deployment," and obtaining the physical data of motion, manipulation, and failure is "extremely expensive."
This data acquisition challenge is creating two distinct strategies. Some firms, like PaXini, are building centralized, government-supported "data factories," with five such facilities already in operation across China. Others, including Galaxea and Spirit AI, are pursuing a decentralized approach, collecting data from already-deployed systems and VR simulations. The data itself is becoming a valuable asset, with companies like UBTech expecting government demand for data factories to become a significant revenue driver by 2026.
A Pragmatic Path to Commercialization Kicks Off in 2027
The path to mass deployment, projected to start between 2027 and 2029, is decidedly pragmatic and grounded in industrial reality. The initial opportunities identified by Goldman are in standardized or semi-structured environments like industrial manufacturing and logistics, focusing on tasks such as sorting, material handling, and inspection. This aligns with insights from Agibot's Dr. Yao, who stated that deployment will begin in industrial scenarios with "clearer ROI" before ever entering the home.
This focus on practicality extends to hardware. Instead of pursuing the costly and complex goal of a fully humanoid, five-fingered hand form factor, many manufacturers are opting for a more cost-effective combination of a wheeled chassis with a two- or three-fingered gripper. This configuration is deemed sufficient to address 70 percent to 90 percent of current industrial applications. The adoption process is methodical, typically involving a three- to six-month proof-of-concept phase, followed by small-batch testing of fewer than 50 units for up to a year before larger pilot deployments of 50 to 100 units per customer.
The report underscores a clear trend: the humanoid robot industry is moving past the "what can robots do" phase and into the "can robots create productivity" stage. For investors, this means the most important metric is no longer the impressiveness of a demo video, but a company's demonstrated ability to capture real-world data and secure pilot deployments in high-value industrial workflows.
This article is for informational purposes only and does not constitute investment advice.
