Guotai Haitong: Physical AI data collection is the substantive bottleneck in the mass-production closed loop

Guotai Haitong released a research report stating that the capacity milestones for Tesla's Optimus have largely been settled. The key variable determining its transition from "capable of mass production" to "capable of doing work" is shifting toward real-machine data collection and the speed of building the data flywheel. The firm believes that the demand elasticity in the data collection segment is more likely to materialize first on the side of domestic embodied AI OEMs and data collection centers. Guotai Haitong's main views are as follows: Capacity milestones settled; data pathway becomes the next key variable. On October 21, 2026, Tesla held its Q3 2026 earnings call. The company disclosed that early production of Optimus at the Fremont factory has already begun, and most equipment for the production line planned at an annual capacity of one million units has been put in place. Musk expects that ramping to approximately 20,000 units per week will take at least a year, possibly longer, mainly because some designs are still being iterated. External deliveries are expected to begin in about a year, with the first batch targeting hazardous and repetitive industrial scenarios. In addition, the second Optimus factory in Austin has broken ground. On the data pathway, Musk proposed the concept of "Optimus Academy" in a February 2026 interview: deploying at least 10,000-20,000 Optimus units to repeatedly practice in the real world and conduct "self-play." The company also disclosed at its Q2 2026 earnings call that early output from Fremont will be deployed to Optimus Academy to accelerate the collection of training data. The firm believes that the capacity milestones lock in delivery certainty; whether generalized capabilities that can be reused can be formed after the robots are delivered depends on the startup speed of the data flywheel. Why data collection is the key constraint: real-machine data can only be actively "manufactured." Data sources for embodied intelligence roughly fall into three layers. The top layer is real-machine data, which has the highest precision and also the highest cost. The middle layer is simulation data, which is low-cost and can be scaled, but there is a simulation-to-reality gap (Sim2Real gap). The bottom layer is internet video and human behavior data, which has broad sources and strong generalization, but lacks 3D motion and force-sensing annotations. Autonomous driving can rely on naturallying data from on-road fleets, but humanoid Siasun Robot&Automation does not have such an existing fleet and can only actively build data sources: first build scenarios, then build data, and only then discuss external sales. At the same time, the task space for humanoid Siasun Robot&Automation is open-ended. Data demand is therefore not one-time collection but continues to expand with task complexity. "Dual-wheel drive" of "core components + domestic data collection infrastructure." The core components side benefits from the certainty of capacity ramp-up, and dexterous hands, precision transmission, and force/tactile sensing remain the core variables determining the ramp-up speed. On the data collection side, demand parties need to be distinguished. Tesla mainly relies on its own Siasun Robot&Automation for self-play in the real world, and its training focus has already tilted toward human video, so its dependence on external data collection equipment is relatively limited. Therefore, order elasticity in the data collection segment is more likely to materialize first domestically. Domestic embodied AI OEMs and various local data collection centers and training grounds are in the stage of "first building scenarios, then building data," and segments such as motion capture and teleoperation equipment, data annotation and simulation synthesis, and force/tactile and 3D vision are expected to enter the procurement cycle before large-scale external sales of complete machines. In addition, if AI5 is introduced into Optimus around April 2027 as planned, it will be able to support higher degrees of freedom and more complex tasks, while also raising requirements for data quality. Tracking milestones: first, the weekly production ramp-up progress and design freeze pace at Fremont; second, the deployment scale and training progress of Optimus Academy; third, the structure of the first batch of external industrial customers in 2027; fourth, the timing of AI5 introduction into Optimus; fifth, the tender and order implementation progress of domestic data collection centers. Risk warnings: embodied intelligence/AI development falling short of expectations; geopolitical risks.