Four industry themes: the robotics data gap, the dairy cycle and defensive healthcare

Why has robot data become a business?

The industry estimates that a mature embodied AI foundation model requires at least tens of millions of hours of training data, while high-quality real-world data available globally amounts to only hundreds of thousands to a million hours, leaving a supply-demand gap of approximately 90%. Data must also link tasks, environments, robot states, actions and outcomes over time, making collection costly.
Real-robot data sells for approximately 500–1000 yuan per hour; collecting data without a physical robot costs approximately 50–100 yuan per hour. These are selling prices and costs, respectively. Real-robot, robot-free and simulation approaches are developing in parallel, and mixed data is expected to become the mainstream training method.
Guojin Securities believes data collection services have relatively stable order visibility and cash flow. In the short term, the focus is on orders, capacity and delivery. Long-term value depends on whether collection, feeding data back into training and retesting on physical robots can form a closed loop, and whether data can be reused across different robot bodies.
Haitian Ruisheng's 2026 capacity for real-robot and robot-free data is approximately 100 thousand and 500 thousand hours, respectively. Assuming full utilization, it could generate 120 million yuan in revenue, equivalent to approximately 32% of 2025 revenue.

Huaxi Securities believes dairy combines the characteristics of a consumer staple and a cyclical sector: investors can focus on expansion in high-margin categories during an upcycle, and on operational improvements and dividends in volatile markets.
After five years of declines in raw milk prices, the price of bulk milk has bottomed out and rebounded. The firm is optimistic about a cyclical reversal in 2026–2027, citing easing downstream competition and improving impairment conditions in upstream dairy farming.

China Securities believes running inference locally on devices such as phones, home appliances and robots can reduce latency, protect privacy and save cloud token costs. Established devices will be upgraded first, while shipments of newer devices will gradually ramp up.
Running large models locally requires both smaller models and sufficient memory bandwidth. According to the report's theoretical calculations, at an effective bandwidth of 50 GB/s, a quantized 3B model can generate approximately 30 tokens per second; under the same conditions, 8B and 30B models can generate approximately 11 and 3 tokens per second, respectively.
Rockchip's coprocessor integrates logic chips and high-bandwidth memory through 3D stacking. Its first customers, in areas including cleaning and companion robots and smart homes, have reached product launch or mass production. Device makers' user access points and scenario data still need to be combined with model developers' compression and deployment capabilities.

Tebon Securities is focusing on improving fundamentals following the exit of pharmacy supply. The number of retail pharmacies nationwide fell from a peak of 684 thousand in 2024 to 656 thousand at the end of 2025. In the second quarter of 2026, revenue in the Shenwan offline pharmacy sector rose 4.1% year over year, while net profit attributable to shareholders rose 11.7%.
The defensive case for traditional Chinese medicine includes lower valuations and higher dividends. Fund holdings accounted for 2.5% at the end of the second quarter of 2026, a historically low level; as of 22 October, 17 listed traditional Chinese medicine companies had dividend yields above 3%.
Costs and inventories are also improving: sector inventories fell 3.6% in the first half of 2026 from the end of 2025; as of 21 October, the traditional Chinese medicinal materials price index had fallen 12.7% since the start of the year. Tebon Securities believes lower raw material prices and inventory clearance are gradually feeding through to gross margins and profits.