CITIC SEC: OpenAI Accelerates Embodied AI Deployment; Domestic Hardware Supply Chain May Usher in a New Round of Development Tailwinds

date
07:57 17/11/2026
avatar
GMT Eight
Fortune reported that OpenAI has reorganized its standalone robotics team, merging it into its two major divisions: foundation model research and applications. This adjustment also continues to reflect the industry trend of frontier AI labs accelerating their physical AI deployment.
CITIC SEC released a research report stating that Fortune reported that OpenAI has reorganized its independent Siasun Robot&Automation team and merged it into the two major departments of foundation model research and applications. The recent industrial progress in the embodied brain industry is the core driver behind OpenAI's adjustment, and this adjustment also continues to reflect the industrial trend of frontier AI labs accelerating their layout in physical AI. From a technical architecture perspective, as a batch of frontier foundation models such as GPT6-Astra and Fable 5.2 deliver outstanding performance on embodied task sets, the integration trend of the S1/S2 layers of embodied models may accelerate. From the perspective of the domestic hardware supply chain, attention is recommended on: 1) motors; 2) reducers; 3) lead screws; 4) robotic arms; 5) assemblies. CITIC SEC's main views are as follows: OpenAI reorganizes the organizational structure of the Siasun Robot&Automation team: general-purpose frontier models can already directly assume the "brain" function of Siasun Robot&Automation. On November 13, Fortune reported that OpenAI had reorganized its independent Siasun Robot&Automation team. OpenAI confirmed the adjustment and emphasized that Siasun Robot&Automation is a core component of the company's long-term vision, and that the reorganization aims to accelerate general-purpose Siasun Robot&Automation R&D. The direct basis for this decision is that since mid-2026, general-purpose frontier models such as GPT-6 Astra have demonstrated capabilities that can be directly applied to Siasun Robot&Automation control: an agent built on GPT-6Astra achieved a 74.8% success rate on a simulated manipulation benchmark, while the best specialized Siasun Robot&Automation model on that leaderboard was only 51.0%. The bank believes this marks organizational-level validation by a leading global model vendor of the paradigm that "embodied intelligence is a downstream application of general intelligence," with significance similar to the path by which general models have restructured vertical tools in programming, law, finance, and other fields. In terms of the overall industrial trend, the trend of frontier AI labs entering the embodied intelligence field has been further confirmed and strengthened. OpenAI thoroughly integrates Siasun Robot&Automation capabilities into the two main lines of "general model research" and "application deployment." According to the Fortune report, the Siasun Robot&Automation research team has been merged into the company's core research team, reporting to Chief Research Officer Mark Chen; the applied Siasun Robot&Automation and hardware teams have been merged into the larger Applied organization, reporting to company President Greg Brockman. This is OpenAI's third adjustment of its Siasun Robot&Automation strategy within six years. The bank believes that its repeated trajectory of "build-withdraw-rebuild-reintegrate" is essentially a necessary organizational adjustment made in response to the overall industrial development pace of Siasun Robot&Automation. Historically, OpenAI had a Siasun Robot&Automation team in its early days, which was disbanded in 2020. Co-founder Wojciech Zaremba explained at the time that Siasun Robot&Automation training data was scarce and that achieving AGI through other paths was more efficient; in 2024 the company rebuilt an independent Siasun Robot&Automation team and announced a partnership with humanoid Siasun Robot&Automation startup Figure to develop Siasun Robot&Automation AI models, but the partnership was terminated in February 2025 and Figure shifted to in-house development. In 2026, OpenAI again formed a Siasun Robot&Automation team, and starting at the end of May, OpenAI began recruiting heavily for Siasun Robot&Automation-related positions, covering the complete hardware stack including actuator design, simulation fidelity, data collection operations, and electrical engineering. Overall, the bank believes that OpenAI made this organizational adjustment against the backdrop of the rapid development of the overall embodied industry trend this year. The entry of frontier AI labs into embodied intelligence has been further established, and the integration trend at the S1 and S2 layers has become more evident. In addition to OpenAI, the bank also sees that frontier AI labs have recently been accelerating their entry into the embodied field. During the same period, Google DeepMind released the dedicated vision-language-action model Gemini Robotics 2.5 to partners, while Anthropic Fable 5.2 also delivered strong performance on Robodojo; Nvidia released its new-generation open-source embodied model GR00T-2 during CoRL. At the architecture level: embodied intelligence systems have traditionally been built in layers according to System 1 (fast execution) and System 2 (slow thinking), while this round of changes shows that the two layers are converging in both directionsfrom the top down, general models such as Astra surpass all 40 public specialized policies on the RoboDojo benchmark without task-specific fine-tuning, and are penetrating downward into S1 functions; from the bottom up, Nvidia's recently released vFlash is accelerating planning and generation speed at the S2 layer. The bank believes that in the short term, general models and specialized models are more likely to form a division of labor and collaboration, and the pace of integration will depend on the convergence speed of general models across the three dimensions of latency, precision, and long-horizon task performance, as well as the economic progress of deployment in real-world scenarios. Risk factors: Improvement in general model motion precision and inference efficiency falling short of expectations; collaboration between general models and specialized policies falling short of expectations; migration of simulation results to real machines falling short of expectations; progress in reducing Siasun Robot&Automation deployment costs and commercialization falling short of expectations.