According to research modeling by Morgan Stanley, the emerging market triggered by embodied AI will become a core growth engine for the semiconductor industry. By 2045, the global market size for semiconductors dedicated to embodied AI robots is projected to reach as high as $305 billion. A more critical industry trend is that, although the price of complete machines is expected to plummet from the current average of $131,000 to approximately $23,000 by 2045, the depreciation of the hardware "body" has not diminished its semiconductor value. On the contrary, the value proportion of semiconductors in the complete machine BOM will surge against the trend from the current 4%-6% to 24% by 2045. This trend indicates that mechanical hardware is moving towards commoditization, while chips and semiconductors will become the absolute dominant profit pool.
Currently, leveraging policy dividends and supply chain cluster advantages, the Chinese market has taken the lead in initiating explosive shipments, becoming the optimal window to observe the restructuring of the "silicon sand super cycle." So, how can we unearth the "golden shovels" that have already entered the supply chains of leading complete machine manufacturers?
I. In-Depth Breakdown of China's Complete Machine Shipments: Explosive Growth and the Evolution of the "Two Superpowers and Many Strong Players" Pattern
Chinese manufacturers have already secured a dominant global position in the mass production of embodied AI, contributing over 85% of global shipments. Regarding shipment data, there is a significant divergence in statistical calibers in the market: in the first half of 2026 (1H26), based on the narrow caliber of SAG, global humanoid robot shipments reached 19,100 units; whereas data based on the broad caliber of the "China Embodied AI 100 Forum" (including service-oriented and wheeled bodies) shows that shipments in the same period have exceeded 40,000 units. Despite the different calibers, the fact that Chinese manufacturers account for as high as 97.4% establishes their absolute leading position.
The market has already shown the embryonic form of a "duopoly" with Agibot and Unitree competing for the top shipment spot, with their combined share reaching approximately 80%.
Table 1: Comparison of Shipments and Market Shares of Major Chinese Embodied AI Complete Machine Manufacturers (1H26 Data)
Core drivers of shipment growth: 1. Special support from local governments: The "15th Five-Year Plan" and the special practical training plan for embodied AI have promoted the government procurement of tens of thousands of robots, with data collection centers in Beijing, Shanghai, Shenzhen, and other places becoming the largest single buyers. 2. Rigid demand for data collection: The generalization of physical AI large models requires over 10 million hours of high-quality physical operation data, forcing complete machine manufacturers to increase their distribution efforts. 3. Price "sweet spot": The BOM cost of complete machines is approaching $28,000 (approximately RMB 200,000), greatly stimulating the willingness of downstream small and medium-sized enterprises to try them out.
II. Scenario Map: Cascading Explosions in Manufacturing Training, Logistics Handling, and Scientific Research Penetration
Embodied AI is undergoing a critical transition from "laboratory demonstrations" to "workshop trial deployments." However, Morgan Stanley's analysis suggests that the adoption logic in different regions exhibits significant ROI differences.
Manufacturing collaboration (industrial factory entry wave): This is currently the most certain sector. In automotive assembly and 3C electronics, robots are replacing processes such as screw tightening. A key "strategic insight" is: due to the relatively low salaries of manufacturing workers in China, the ROI payback period for humanoid robots in the current Chinese market is as long as 6.7 years, whereas it only takes 2.7 years in the US market. This means that although China leads in shipments, the deep penetration of high-value scenarios may first cross the 2.5-year adoption threshold in the US market, which has a shorter payback period.
Logistics handling (generalized operations): Compared to traditional AGVs, robots with "large model + dual-arm" capabilities, such as those from Galbot, can cope with unstructured environments. The extreme pursuit of sorting efficiency by giants like Meituan and JD.com has driven the release of demand at the scale of tens of thousands of units.
Scientific research and education (sinking market): Unitree's ultra-low-priced body at RMB 29,900 has lowered the development threshold through "open-source geekification," accounting for approximately 65% of the current shipment stock.
Home companionship (long-term potential): Although the long-term market space is huge, constrained by irregular environments and safety challenges, it is still in a 5-10 year technology convergence period. There is a severe mismatch between the pricing of UBTECH U1, which is as high as RMB 880,000-990,000, and its current functional value.
III. Chip Supply Chain Mapping: "Dual-Track" Procurement Strategy for AI Chips and the Localization Path
Embodied AI robots are essentially distributed semiconductor systems. Currently, Chinese complete machine manufacturers exhibit a clear "dual-track coexistence" strategy at the computing power layer.
"Dual-Track" Strategy for AI Chips:
Reliance on overseas high-end computing power: For flagship models running multi-modal models with tens of billions of parameters, NVIDIA (Jetson Orin/Thor) remains the core reliance. Galbot has been among the first to equip the Thor platform, and Agibot's Yuanzheng A2 and Genie G2 models also highly depend on NVIDIA's computing power.
Domestic cost-effective path: To achieve cost reduction, high cost-performance models have deeply integrated Rockchip (RK3588, etc.). Rockchip's share in the "cerebellum" SoC (System on Chip) of domestic humanoid robots has exceeded 75%. In addition, SemiDrive and Galbot have signed a strategic agreement to jointly develop the next-generation dedicated processor, marking the dawn of the era of domestic customized ASICs.
Table 2: Mapping Table of Chinese Embodied AI Complete Machine Manufacturers and Core Chip Suppliers
IV. Competitiveness Comparison of Core Suppliers: Technological Moats, Capacity Bottlenecks, and TCO Advantages
Against the backdrop of "silicon computing power as the dominant profit pool," the competition among chip companies has shifted to a game of ecosystem and cost.
Table 3: Competitiveness Comparison Matrix of Core Chip Suppliers
TCO advantage analysis: Morgan Stanley's research shows that although the single-item performance of domestic chips still has a generational gap compared to overseas ones, in large-scale inference scenarios, through efficient multi-die stacked packaging, the actual Total Cost of Ownership (TCO) of domestic intelligent computing solutions is 30%-60% lower than overseas solutions.
V. Outlook for the Next 24 Months: Capturing the Profitability Nodes in the Silicon Sand Cycle
Morgan Stanley maintains its judgment that 2026 will be the "first year of mass production" for Chinese robots. Using the Di(t) quantitative modeling model, we have calculated the explosion nodes of semiconductor components in the next two years:
Quantitative forecast of component demand (Di(t) model):
- 2026 (forecast shipment of 80,000 units): The total demand for joint control MCUs (Microcontroller Units) will reach 5.52 million units; the demand for perception sensor chips will be 3.68 million units.
- 2027 (forecast shipment of 180,000 units): The demand for joint control MCUs will climb by an order of magnitude to 12.42 million units; the demand for perception sensor chips will reach 8.28 million units.
Key catalysts:
1. IPO of pure-play targets: Unitree's listing on the STAR Market will reconstruct the valuation anchor of the sector.
2. Breakthroughs in large models at the edge: Physical AI large models (such as the fine-tuned version of DeepSeek-V3) achieve ultra-high token throughput at the edge.
3. Batch deployment by automakers: XPeng and BYD achieve thousand-unit level deployment of IRON robots on their production lines.
Investment recommendations:
1. Adopt the "shovel and gold digger" strategy: Avoid the red ocean of complete machine competition, and overweight high-certainty components such as perception magnetic encoders, GaN drivers, and motor MCUs.
2. Lock in "core domestic substitution" targets: During the game of advanced process nodes, focus on allocating resources to Rockchip, which holds a dominant position in the "cerebellum" SoC, and Nations Technologies, which has a deep foothold in joint main control and physical security chips.
3. Mine the dividends of the "perception layer": Focus on allocating resources to Orbbec (visual perception middle platform) and Hesai Technology (miniature LiDAR). As shipments scale up, the ASIC-ization of the perception layer will bring huge profit elasticity, becoming the first stop for physical AI interaction.
Appendix: Forecasts of Global Top Investment Banks on Complete Machine Shipments and Core Chip Demand in the Next 12-24 Months
JPMorgan and Morgan Stanley's complete machine shipment forecasts: Morgan Stanley predicts that the shipment volume of humanoid robot complete machines in China will reach 50k units in 2026.
The institution further predicts that, as pilot projects transition to larger-scale deployments, the shipment volume of humanoid robots in China will reach 100k units in 2027 and climb to 446k units by 2030.
DBS Group's forecast for leading manufacturers: DBS Group expects that, as a leading Chinese complete machine manufacturer in global shipments, Unitree's humanoid robot shipments will reach 15,000 units in FY26F. With the improvement of product cost-performance and scenario expansion, Unitree's shipments are expected to accelerate to 40,000 units and 90,000 units in FY27F and FY28F, respectively.
UBS Securities' breakdown of BOM costs and chip demand: Research by UBS Group (UBS) points out that actuators account for over 40% of the total BOM cost of humanoid robots. In terms of core semiconductors and perception components, complete machine manufacturers currently rely heavily on external procurement to obtain high-computing-power edge computing modules, power battery cells, camera components, and high-performance sensors.
Huatai Securities' outlook for edge computing chip demand: Huatai Securities points out that as embodied AI large models evolve towards multi-modal and agent directions, the demand for edge inference computing power will show exponential growth. This trend will directly drive up the per-unit installation volume and unit price of high energy-efficiency AI inference SoCs, high-speed bus interconnect chips, high-bandwidth memory, and high-capacity enterprise-grade SSDs.
Market Catalysts and Potential Headwinds Analysis
Industry catalysts pointed out by Morgan Stanley: Morgan Stanley believes that key short-term market catalysts include the IPO issuance of pure-play embodied AI targets such as Unitree, the holding of the World Robot Conference (WRC) in Beijing, and the World Humanoid Robot Games. In addition, the release of Tesla's Optimus Gen3 and small-batch mass production nominations are also important events driving the sector's valuation repair.
Mass production and application catalysts emphasized by DBS Group: DBS Group points out that Chinese automakers (such as XPeng, BYD, and Xiaomi) conducting trial training and deployment of humanoid robots in their own auto factories is a landmark milestone for embodied AI moving towards B2B practicality. Among them, XPeng plans to achieve mass production of its IRON humanoid robot by the end of 2026 and launch commercial sales and external deliveries in 2027, which will serve as a strong catalyst for the synergy between complete machine manufacturing and the automotive supply chain.
Policy drivers focused on by Cinda Securities and Hong Kong Economic Times: Hong Kong Economic Times points out that the guidelines for the construction of the national standard system issued by China's Ministry of Industry and Information Technology and other departments, as well as the special industrial funds and data collection centers set up by local governments (such as Beijing, Shanghai, and Shenzhen), provide the industry with highly certain policy support.
Bank of America Merrill Lynch's warning on geopolitical risks: Bank of America Merrill Lynch warns that the US Federal Communications Commission (FCC) and Congress are accelerating access restrictions on Chinese intelligent robots, such as promoting the "Guarding American Robotics from Chinese Domination Act" (GUARD Act). Although these geopolitical restrictions have limited impact on models already authorized by the FCC in the short term, they will hinder future new models from entering the high-value US market and prompt US humanoid robot manufacturers to accelerate the "de-Chinization" supply chain restructuring.
Huang Mao and CLSA's assessment of multiple legislative risks in the US: CLSA points out that US sanctions on Chinese robots have formed a multi-departmental coordinated situation, including the Department of Defense listing some Chinese complete machine manufacturers on the 1260H military enterprise list and the Senate promoting the "Safe American Robots Act." These bills intend to implement de facto market bans by imposing high Section 232 tariffs on non-compliant robots and restricting government procurement, forcing overseas complete machine manufacturers to transfer core component orders to non-Chinese suppliers in advance.
References:
1. Morgan Stanley: 20260817 - Morgan Stanley Research - China Industrials The Humanoid 2. AlphaSense: Research on Shipment Volume of Complete Machines and Chip Supply Chain of China's Embodied AI Robots 3. Gemini Notebook Deep Research: In-depth Research Report on China's Embodied AI Robots and Semiconductor Supply Chain: Silicon Sand Super Cycle and Industrial Pattern Reshaping Driven by Physical AI.
#ShenXinMeng #BayChipExpo #SemiconductorIndustryResearch #AI #Robots #SiliconSand
Republished with permission.（Translated and republished with permission.） Copyright remains with the original author.