To The Moon

On October 20, Hong Kong's robotics sector staged a notable rally. Two major catalysts drove the move: OpenAI publicly confirmed a comprehensive push into humanoid robot R&D, listing robot intelligence systems as a core research priority; and Tesla, in written materials for its shareholder meeting, formally announced that its third-generation Optimus humanoid robot is ready for mass production. Against this backdrop, SEER TECH (06106) surged as much as 16% intraday, with its share price touching HK$57.25. What matters more than the single-day move is the narrative shift behind this round of capital bets. Guosheng Securities' Zhang Yiming argues that from the fourth quarter through next year, physical AI will see a wave of major players entering the field, with the flywheel of "brain, data, and applications accelerating." Great Wall Securities' Deng Yuliang, tracking the industry chain, further pinpoints the truth: the core bottleneck in the robotics industry is no longer hardware bodies and motion control—the real test is whether large models can drive robots to complete generalized tasks. Tesla's push to ramp up Optimus quickly is fundamentally aimed at internal training and real-machine data collection; what it is racing for is data, not just selling the robot itself. These two industry judgments reinforce each other, and the main thread in capital markets has quietly shifted. Capital's focus is no longer solely on hardware components such as motors and reducers, but is tilting toward robot brains and real-machine training data. Across the Hong Kong market, SEER TECH is one of the few listed targets that genuinely takes the "robot brain" as its business core. What kind of industrial value is this company actually delivering?

The Inflection Point of Embodied Intelligence Logic: Value Shifts from Hardware Manufacturing to Brains and Real-Machine Data

Over the past few years, public attention on the humanoid robot industry has mostly centered on hardware bodies. Lead screws, harmonic drives, and servo motors took turns as market hot topics, and amid the wave of localization, hardware supporting components matured rapidly. But the industry's pendulum is shifting. The real question of embodied intelligence has never been about completing a robot's limbs, but about how to convert large models' text-understanding capabilities into stable, reliable physical actions in the real world. Model outputs must go through real-time control, motion planning, and hardware interface conversion before they can drive robots to complete complex tasks. The hardware side has already gone through three years of rapid development, and localization of core components is largely in place. As the threshold for bodies continues to fall, the real pain point that is hard to standardize begins to surface: whether a robot can interpret real-world tasks and execute them in a stable closed loop under complex conditions. To iterate stronger embodied large models, high-quality physical datasets are an unavoidable prerequisite. Text, images, and video cannot fully reproduce the laws of the physical world; simulation environments generally suffer from the Sim2Real gap; first-person body data is constrained by hardware configuration; and teleoperation collection is difficult to scale. Dongwu Securities International notes in a research report that only data generated by autonomous real-machine operation can fully record the complete closed-loop feedback of "state-action-next state-result," naturally accumulating large volumes of anomaly and extreme-condition samples. The complete-machine segment will remain fragmented over the long term. Different industries, factories, and production lines differ enormously in materials, space, production rhythm, and safety standards, making it hard to move toward full monopoly by a few oligarchs. By contrast, control software and model algorithms naturally have huge potential for standardized reuse across machine types. Vertically integrated complete-machine makers prioritize control systems for their own hardware systems, with limited openness; end customers developing their own control systems must bear extremely high long-term costs for algorithm iteration and maintenance, which is commercially unattractive. Neutral third-party robot brains thus move into a key industry position: not competing in complete-machine hardware, while serving complete-machine makers, integrators, and end customers alike, with high R&D costs shared across massive installed scale. Dongwu Securities International estimates that in the short to medium term alone, the addressable market for embodied intelligence in scenarios such as industrial handling, light logistics operations, and commercial cleaning, shelving, and delivery could reach RMB 2.5 trillion to RMB 8 trillion; over the long term, as model capabilities iterate and upgrade, the stock of equipment that can be transformed could expand to tens of trillions. Across the entire industry chain, the robot brain is a necessary node spanning all scenarios.

Fifty Thousand Distributed Terminals: SEER TECH Builds a Low-Marginal-Cost Data Production Network

The so-called robot brain is not a single piece of hardware, but an integrated software-hardware coordination system of "model + cerebellum + nerves." The upper-layer model handles environmental perception, task decomposition, and intelligent decision-making; the middle-layer cerebellum handles motion planning, trajectory generation, real-time execution, status feedback, and functional safety; the bottom-layer "nerves" rely on unified interface protocols to connect sensors, joints, and all actuators. The SRC series controllers serve as a unified hardware carrier, hosting AI compute units, real-time control kernels, and multi-bus interfaces; the Roboshop, RDS, M4, Meta, and Robocare software toolchain extends single-machine control capabilities to multi-machine collaboration and fleet management. The company has completed adaptation for more than 400 types of core components, and its Nebula platform has over 2,000 SKUs. Manufacturing and assembly, as well as project implementation, are handled by ecosystem partners, while the company focuses on R&D definition, testing and acceptance, and technical support, pursuing a light-asset platform route. About 50,000 devices equipped with SEER TECH's brain are now deployed at customer sites worldwide. The significance of this number goes far beyond a simple tally of business scale. The devices cover more than 2,500 customers, over 20 industries, and more than 40 countries and regions, spanning hardware forms including humanoid robots, robot dogs, embodied forklifts, composite robots, and traditional AMRs. About 20% of the devices have compliance authorization for routine data transmission back to the company; by the end of 2025, roughly 500,000 hours of cross-configuration real-machine multimodal data had been accumulated, including point clouds, vision, motion trajectories, and actuator parameters—significantly higher than the tens of thousands of hours typical of overseas peers. The core advantage of this system lies in data consistency. All devices share the same brain, and data is collected through the same control chain, naturally aligning timing, actions, and results, greatly reducing secondary labeling costs and allowing direct use for model training. Business and data thus form a two-way self-reinforcing flywheel. New devices enter real production scenarios, continuously transmitting operational data back to feed model iteration and optimization; upgraded model capabilities improve the competitiveness of the entire brain product line, driving new device shipments and further expanding data supply. The entire data loop reuses existing devices, customer networks, and upgrade channels, pushing the marginal cost of new data extremely low. Dongwu Securities International points out that its expansion logic is: operating installed base × transmission coverage rate × effective data output per machine—all three variables are currently on an upward trend. In other words, the company is betting on the scaling law of real-machine data, anchored in genuine commercial shipments, rather than betting on a single model architecture. Current shipments mainly consist of single-configuration VLA deployment, while the company is jointly developing a cross-configuration WAM model with frontier teams and plans to release related results and a demo in 1H27.

Earnings Signal: Operating Leverage Opens, a Value Re-Rating Window Has Emerged

Industrial logic must ultimately land in business execution. Looking at the 1H2026 financial report, SEER TECH's growth gear has clearly stepped up. In 1H2026, the company achieved revenue of RMB 264 million, up 67.5% year over year, compared with a 30.2% compound growth rate from 2022 to 2025—a clear upward jump in the growth curve. Breaking it down, intelligent robot business revenue was RMB 185 million, up 70.0% year over year; robot brain business revenue was RMB 59 million, up 56.1% year over year, with brain product shipments exceeding 8,000 units, up more than 80%. This acceleration in growth is the most direct financial confirmation that the embodied intelligence industry is entering an accelerated phase. More important than revenue figures is the change in profit quality. In 1H2026, overall gross margin rose 1.2 percentage points to 46.6%, with gross profit of RMB 123 million, up about 72% year over year; the robot brain business maintained a gross margin of around 80%. In 2025, that segment contributed only about 25% of revenue but generated 42% of gross profit. Selling expense ratio, administrative expense ratio, and R&D expense ratio increased 34.9%, 36.3%, and 43.8% year over year respectively, all significantly below the 67.5% revenue growth rate, with all three expense ratios declining simultaneously. R&D investment continues to increase, but operating leverage is already visibly emerging. During the period, the overall loss narrowed from RMB 51 million to RMB 38 million, and the adjusted net loss narrowed 50.1% year over year to about RMB 10.97 million. Order and customer data further solidify the visibility of growth. In 1H2026, newly signed order value exceeded RMB 467 million, up more than 60% year over year, reaching 1.8 times current-period revenue; new customers exceeded 410, up 65% year over year, while the repeat purchase rate of existing customers remained above 60%. Looking back from 2023 to 2025, annual signed customers grew from 620 to 1,114, and the contribution rate of signed orders from existing customers rose from 50.9% to 60.6%. With new customers continuously expanding and existing customers deepening cooperation, growth is not driven by a one-off pulse. Overseas business has become the sharpest second growth curve in this round of expansion. In 1H2026, overseas revenue was RMB 65.806 million, up 197.5% year over year, with its revenue share rising from 14.0% in the same period last year to 24.9%; new overseas orders exceeded RMB 124 million, up 550% year over year, and the number of signed customers rose 150% year over year. Its business covers 43 countries and regions, with a customer list including NV, TSLA, BYD, Foxconn, and Caterpillar. Even more noteworthy is the profit structure: overseas business gross margin is about 67%, far above the overall blended level, while payment terms are more favorable. Each increase in the overseas revenue share simultaneously lifts overall revenue growth and consolidated gross margin. Market disagreement is often the starting point of value re-rating. Dongwu Securities International initiated coverage in September with a Buy rating, using a sum-of-the-parts weighted valuation: the robot complete-machine-related business is assigned 7x 2027E P/S; the robot brain, with its around 80% gross margin, standardized replication capability, and imagination space from extending into data and model layers, is assigned 18x P/S, resulting in a weighted overall 2027E P/S of about 9.8x and a 12-month target price of HK$100.55; Guotai Haitong issued an Overweight rating on September 23, assigning 13x 2026 P/S, with a target price of HK$92.74, and forecast that the company could turn net profit attributable to shareholders positive in 2027. As a reference, peers such as Yuejiang, Huayan Robotics, and UBTECH trade at an average 2027E P/S of about 6.1x. Most comparable targets use complete-machine robot business as their valuation anchor, but more than half of SEER TECH's gross profit comes from high-margin brain software business, and it also holds a real-machine data asset that the capital market has not yet priced separately—500,000 hours of real-machine data, targeting 1 million hours by 2027. While the market is still comparing the locomotion performance of various prototypes, this company has already turned 50,000 robots distributed across global production lines into a continuously operating data production network.

Final Thoughts

The investment narrative for SEER TECH can be broken into progressively layered dimensions of observation. In the short term, the focus is on equipment shipment scale and gross margin changes; in the medium term, on the rising share of overseas business and higher device data transmission rates; and in the long term, on whether the WAM world model can convert massive accumulated real-machine data into substantive premiums at the product and model levels. Behind each layer of logic lies a clear major industrial trend. The label of "first robot brain stock" is only appearance; what is truly scarce is those 50,000 robots scattered across real global scenarios, running day and night, continuously accumulating physical-world experience for the platform. The embodied intelligence race has already moved past the stage of prototypes flexing their muscles. As the industrial wave rolls forward, the future competition is no longer just about prototype parameters, but about who can continuously obtain high-quality real-machine physical data.