Cathay Pacific Haitong: Physical AI data collection is a real bottleneck in the closed loop of mass production

The Zhitong Finance App learned that Cathay Pacific Haitong released a research report saying that the production capacity node for Tesla Optimus has basically been settled. The key variable that determines it's moving from “energy production” to “capable work” is shifting to the speed of data acquisition and data flywheel construction for real machines. The bank believes that demand elasticity in the digital acquisition process is more likely to be realized first on the side of domestic entity manufacturers and digital acquisition centers.

Cathay Pacific Haitong's main views are as follows:

The production capacity node is settled, and the data path becomes the next key variable. On October 21, 2026, Tesla held a performance exchange meeting for the third quarter of 2026. The company revealed that early production of Optimus at the Fremont plant has already started, and most of the equipment for the planned production line with an annual production capacity of one million units is already in place. Musk predicts that it will take at least a year to climb the slope to weekly production of about 20,000 units, possibly longer. The main reason is that some designs are still being iterated. Foreign transactions are expected to commence in about a year, and the first batch targets dangerous and repetitive industrial scenarios. Additionally, construction of the second Optimus plant in Austin has commenced. In terms of data paths, Musk proposed the idea of “Optimus Academy” (Optimus Academy) in an interview in February 2026: deploy at least 20,000 Optimus units, practice repeatedly in the real world, and carry out “self-play” (self-play). The company also revealed at the 2026 second quarter results meeting that Fremont's early output will be invested in Optimus Academy to accelerate the collection of training data. The bank believes that the capacity node is locked in the certainty of delivery; whether the main body can form a generalization capacity that can be reused after delivery depends on how fast the data flywheel starts.

Why is digital acquisition a key constraint: Real machine data can only be actively “manufactured”. There are roughly three layers of intelligent data sources. The top layer is real machine data, which has the highest accuracy and the highest cost. The middle layer is simulation data, which is low in cost and can be scaled, but there is a migration gap (Sim2Real gap) from simulation to reality. The bottom layer is Internet video and human behavior data. It has a wide range of sources and strong generalization, but lacks three-dimensional motion and kinesthetic labeling. Autonomous driving can rely on the natural return of data from road fleets. Humanoid robots do not have such an existing fleet; they can only actively build a data source: first build the scene, then build the data, and finally talk about export sales. Meanwhile, the humanoid robot's mission space is open. As a result, data requirements are not collected all at once, but continue to expand with the complexity of the task.

“Body parts + domestic digital infrastructure” two-wheel drive. The body side benefits from the certainty of productivity climbing. Dexterous hands, precise transmission, and force/tactile sensing are still the core variables that determine climbing speed. The digital acquisition side needs to distinguish between the demand side. Tesla mainly uses its own robots to play games on its own in the real world. Its training focus is already skewed towards human video, and its dependence on external digital acquisition equipment is relatively limited. Therefore, the flexibility of orders in the digital collection process is more likely to be fulfilled domestically first. Domestic console manufacturers, local digital acquisition centers, and training grounds are in the stage of “building the scene first, then reconstructing the data”. It is expected that they will enter the procurement cycle before large-scale export of the whole machine. Furthermore, if AI5 introduces Optimus around April 2027 as planned, it will be able to support a higher degree of freedom and more complex tasks, while also increasing the requirements for data quality.

Tracking points: First, Fremont's weekly climbing progress and design freeze pace; second, Optimus Academy's deployment scale and training progress; third, the structure of the first batch of external industrial customers in 2027; fourth, the timing of AI5's introduction to Optimus; and fifth, the execution of tenders and orders for domestic digital acquisition centers.

Risk warning: Embodied intelligence/AI development falls short of expectations; geopolitical risk.