Tesla’s Optimus is set for mass production next year, Xpeng’s IRON is already on the road, and Figure 02 is open for pre-orders… As robots transition from laboratories to factories and homes, the competition for computational power is becoming the key factor in determining success.
1. Introduction: The Year of Mass Production Has Arrived, and the ‘Brain’ is the Biggest Bottleneck
By 2025, humanoid robots will officially bid farewell to the ‘PPT era’.
- Tesla announced that Optimus will startsmall-scale deliveries in Q1 2026, aiming for an annual production of thousands;
- Xpeng IRON completed its ‘muscle cutting’ demonstration on November Tech Day, planning to enterfactory training in 2026;
- Figure AI‘s Figure 02 robot is now open forcommercial pre-orders from BMW and Amazon;
- Domestic manufacturers like UBTECH, Yushun, and Leju have launched products priced around ten thousand yuan, targeting education and elderly care scenarios.
However, all these ambitions point to the same underlying question: How can robots run large models with billions of parameters in real-time locally (on the edge)?
In the past, we discussed motors and reducers; today, the real ‘bottleneck’ has shifted to AI chips. This battle for the ‘brain’ not only concerns the technical route but will also reshape the semiconductor industry landscape for the next decade and give rise to a new wave of investment opportunities.
2. The Battle of Technical Routes: End-to-End vs. Layered Architecture, Determining Chip Demand
Currently, mainstream players are divided into two camps:
1. End-to-End Camp: All-in on Large Models, Power Above All
- Representatives: Tesla (Optimus), Google (RT-2)
- Characteristics: Vision + Language + Action are directly output by a single neural network, without intermediate modules.→ Advantages: Strong generalization ability, capable of handling unknown tasks;→ Disadvantages: Requires massive data + super strong computing power, with extremely high chip requirements.
- Chip Demand: Must support real-time inference of Transformers, computing power > 500 TOPS, memory bandwidth > 200 GB/s.
2. Layered Decision-Making Camp: Brain + Cerebellum Collaboration, Efficiency First
- Representatives: Figure 02, Xpeng IRON, UBTECH Walker X
- Architecture:
- Brain (Top Layer): Runs large models like GPT-4V, responsible for understanding instructions and planning tasks;
- Cerebellum (Middle/Bottom Layer): Dedicated neural networks or traditional controllers, responsible for high-frequency motion control at 200Hz–1kHz.
- Chip Demand: Requires a heterogeneous computing platform—NPU for large models, GPU/CPU for perception, and MCU for low-level control.
✅ Conclusion: In the short term, layered architecture is mainstream (cost-effective, mature technology); In the long term, end-to-end is the direction (higher intelligence ceiling). Regardless of the path, high computing power, low power consumption, and strong real-time performance of edge AI chips are essential.
3. Global AI Chip Landscape: Who is Building the ‘Brain’ for Robots?
| Chip Manufacturer | Representative Product | Partner Robot Companies | Core Advantages |
|---|---|---|---|
| NVIDIA | Jetson Thor | Tesla (early testing), several humanoid robot startups | Unmatched CUDA ecosystem, mature Isaac Sim simulation platform |
| Qualcomm | RB5 / Next-Gen Robot Platform | UBTECH, Leju, some Chinese startups | 5G + AI integration, high energy efficiency, supported by Arduino ecosystem |
| Intel | Core™ Ultra (Brain-Cerebellum Fusion Solution) | Yushun Technology, domestic industrial robot manufacturers | Heterogeneous integration of CPU + NPU + GPU, complete OpenVINO toolchain |
| Huawei | Ascend Atlas Series | Not publicly disclosed, but has full-stack capabilities | Preferred choice for domestic alternatives, MindSpore framework adapted for embodied intelligence |
| Horizon Robotics | Journey 5 / Journey 6 | Xpeng Motors (XNGP), expected to extend to IRON | Reuse of autonomous driving chips, strong cost control capabilities |
| Cambricon | SiYuan 590 | Research institutions, university robot projects | Domestic NPU IP, supports large model compression |
🔍 Key Insight: NVIDIA remains the current developer’s choice, but Qualcomm and Intel are quickly catching up through vertical integration + ecosystem binding. In the Chinese market, the window for domestic chip alternatives is opening.
4. Opportunities in China: Three Major Investment Lines to Capitalize on the ‘Brain’ of Robots
Main Line 1: Domestic AI Chip Design Companies (Direct Beneficiaries)
- Horizon Robotics (not listed) is backed by Xpeng, Li Auto, and other car manufacturers, its Journey series chips have verified automotive-grade reliability, and transitioning to humanoid robots is highly likely. Keep an eye on its IPO progress.
- Cambricon (688256.SH) SiYuan 590 supports INT4 quantization and can run large models below 30B. If it collaborates with UBTECH, Yushun, etc., it will open up new spaces.
- Black Sesame Intelligence (planned Hong Kong IPO) Huashan series chips reach 106 TOPS and are expanding into the robotics field.
Main Line 2: Upstream Chip and Supporting Enterprises
- Chipone Technology (688521.SH) provides NPU IP licensing (e.g., Vivante series), many domestic robot chips use its IP. This is the ‘selling shovel’ logic.
- Guoxin Technology (688262.SH) Self-controllable CPU cores have application potential in robot controllers.
- Lanqi Technology (688008.SH) is a leader in memory interface chips, and the demand for high-bandwidth memory (HBM) in robots will drive its technology upgrades.
Main Line 3: Software and Toolchains (Invisible Champions)
- Thunder Software Technology (300496.SZ) has deep cooperation with Qualcomm, providing robot operating system (ROS) optimization solutions, serving multiple leading clients.
- CloudWalk Technology (688327.SH) ‘Calm Large Model’ is exploring embodied intelligence applications and may provide industry-specific robot models.
- Tuolsi (300229.SZ) has accumulated in the combination of knowledge graphs + large models, which can be used for common sense reasoning in robots.
5. Summary of Investment Opportunities
| Category | Core Logic | Key Focus Targets |
|---|---|---|
| Domestic AI Chips | Humanoid robots are the best testing ground for domestic chips | Cambricon (688256), Horizon Robotics (primary market) |
| IP and EDA | Chip design cannot be separated from underlying IP | Chipone Technology (688521), Huada Jiutian (EDA) |
| Operating Systems and Algorithms | Software defines the intelligence ceiling of robots | Thunder Software Technology (300496), CloudWalk Technology (688327) |
| Advanced Packaging | Chiplet technology enhances computing power density | Changdian Technology (600584), Tongfu Microelectronics (002156) |
⚠️ Risk Warning:
- The technical route has not yet converged, posing a risk of betting on the wrong track;
- Robot deployment may be slower than expected, leading to potential delays in chip orders;
- The U.S. has upgraded export controls on high-end AI chips.
6. Conclusion: He Who Controls the ‘Brain’ Controls the World
The ultimate competition for humanoid robots is not about whose legs are faster, but whose ‘brain’ is smarter, more energy-efficient, and more reliable.
In this trillion-dollar entry battle, chip manufacturers are no longer supporting roles but are the strategic core defining product forms. For China, this is not only a technical challenge but also a key springboard to break overseas monopolies and achieve semiconductor independence.
We will continue to track the birth and evolution of every ‘robot brain’, marking those value points that are building the future intelligent foundation.Frontier Observations, Awaiting Opportunities.🚀Follow us for valuable insights