OpenAI turns to in-house robotics: How far has Nanshan Zhishang progressed with its materials strategy?
OpenAI is developing hardware in-house to complete a real-world data loop, while algorithm integration is also advancing. Nanshan Zhishang has begun joint testing and small-batch supply of robot materials, but its recent profit improvement has come mainly from its nylon business.
Original source：数据研选First published on Unhidea：Original source date：
Key takeaway
The robotics industry is moving toward a hardware and data loop, but Nanshan Zhishang’s current operational improvement and its robot materials strategy are at different stages of realization.
Why is OpenAI moving from investing in robotics to building complete robots itself?
The materials say OpenAI has shifted from investing in Figure to developing humanoid robots in-house. It posted 19 robotics positions on its website, covering control systems, actuators, and manufacturing and supply chains. The key change is to build core hardware alongside a real-world data collection system to address the data gap created by the end of its partnership with Figure.
The business logic is to use in-house hardware to gain access to action data from real-world environments, then use it to train and improve models. On that basis, Liu Zejing of Huaxi Securities is optimistic about OpenAI’s robotics supply chain and cites Changying Precision and OPT Machine Vision.
OpenAI plans to introduce a demonstration prototype in 2027. The materials also point to an execution gap between simulation and reality, with Tesla and XPeng ahead in the pace of mass production. A demonstration prototype and progress toward mass production need to be understood separately.
Which part of robot capabilities is being changed by algorithm integration?
Beyond in-house hardware development, another main thread is the integration of action, prediction and generation capabilities. Nvidia and Harvard’s Hydra-0 reduces robot motion error by 90.4% through a unified action-stream interface; Deep Robotics’ DR-WM-1.0 introduces prior knowledge of physical properties.
Agibot’s GE-Act 2.0 integrates prediction and generation and, using 30 thousand hours of data, raises the zero-shot success rate in unfamiliar scenarios to 44.1%. These metrics respectively measure motion error and performance in unfamiliar scenarios, reflecting different capabilities.
Liu Zejing summarizes the advances as follows: in-house hardware development and integration of underlying algorithms together help physical AI establish a technology stack for a “data loop” and “thinking before acting.” The former addresses the source of data, while the latter improves prediction and decision-making before action.
How does Nanshan Zhishang’s robot materials strategy connect with its existing business?
Nanshan Zhishang’s profit improved quarter on quarter in the second quarter, driven primarily by its nylon business. Its flexible transmission materials for cable-driven robots are currently in joint testing and small-batch supply; the current operational improvement and progress in its emerging business have different sources and timelines.
The company’s entry point is to apply its experience in fine spinning, flexible customization and end-to-end quality control from high-end textiles to the production of high-performance fibers. The materials say it has built a scaled production base with annual capacity of 3600 tonnes; its robotics research and development includes application products such as key ropes, tactile gloves and skin suits.
Sun Haiyang of TF Securities is optimistic about the combination of upgrading traditional textiles, expanding high-end materials capacity and developing emerging applications. He forecasts revenue of 2.4 billion, 2.8 billion and 3.2 billion yuan, respectively, for 2026–2028, and net profit attributable to shareholders of 150 million, 180 million and 230 million yuan, respectively. He identifies the pace of the capacity ramp-up, consumer demand conditions, tariff changes and market competition as the main constraints.
