WAIC 2026 shows China’s robotics mass-production inflection point has arrived, but demand and the data closed loop still need validation

Overview

This report summarizes Nomura’s observations from the 2026 World Artificial Intelligence Conference in Shanghai. Its core conclusion is that China’s robotics industry has moved from stage demonstrations into the phase of mass production, delivery, and deployment validation, but the demand side has not yet been fully proven. The report argues that the industry narrative is shifting from TAM to shipments, ASP, margins, and real deployments, with the manufacturing side resembling the mass-production inflection point of the 2019-20 new energy vehicle supply chain; however, current humanoid robot demand still mainly comes from entertainment/performances, consumer use, government procurement, education, data collection factories, and pilot demonstrations, while productive industrial replacement has not yet become mainstream.

Core views

The report’s core views include: first, mass-production capability has improved, with nearly 60 humanoid robots performing actual services on site at WAIC 2026 and AgiBot’s cumulative output reaching 15,000 units, but this validates manufacturability rather than demand durability. Second, industrial and logistics scenarios are becoming the direction for technology-route convergence, with vendors placing greater emphasis on predictable, repeatable environments and even adopting wheeled forms that sacrifice anthropomorphism in exchange for reliability and cost advantages. Third, industrial customers base procurement on ROI, MTBF, and cycle time, and for most workstations the current ASP of humanoid robots versus labor cost still falls short of the adoption threshold. Fourth, the real bottleneck in embodied intelligence is shifting from hardware to high-quality real physical interaction data, with a significant gap between existing data stock and training demand. Fifth, dexterous hands are the fastest-iterating link in the hardware chain, but supply progress is still running ahead of end-demand validation.

Analysis framework

The report uses conference observations, industry research, expert interviews, and cross-validation with supply-chain data to compare on-site deployments at WAIC 2026, vendor technology routes, demand structure, funding flows, data collection capabilities, and dexterous-hand hardware routes. The focus is not simply on assessing demonstration effects, but on judging whether shipments can convert into sustainable orders, whether industrial customers are willing to procure using operating budgets, and whether the data closed loop is sufficient to support the maturation of robot “brains.”

Methodology notes

• industry_lifecycleManufacturing Mass-Production Inflection Point and Demand Validation Framework

  Frames the robotics industry’s shift from concept demonstrations to mass-production delivery as analogous to the manufacturing-side inflection point of new energy vehicles in 2019-20.

  The report believes that mass production, delivery, and integration capability are the metrics of the new stage, but order quality, paying-customer mix, repeat orders from non-affiliated customers, and unit economics must be further validated before shipments can be confirmed to translate into profits.

• unit_economicsIndustrial Adoption Framework Based on ROI, MTBF, and Cycle Time

  Industrial buyers judge whether robots can replace labor or robotic arms based on return on investment, mean time between failures, and production cycle time.

  The current ASP of humanoid robots relative to labor cost still leaves the payback period for most workstations insufficient to trigger large-scale adoption; grippers are already sufficient for structured workstations, and five-finger dexterous hands may be overengineered in some scenarios.

• ai_data_loopCollection-Training-Deployment-Feedback Closed Loop

  The maturity of robot “brains” depends on forming a closed loop of real physical interaction data, training, deployment, and actual feedback.

  The report believes that architecture and computing power are not the core constraints; high-quality real interaction data is the bottleneck. Even if data prices decline, that does not mean the data gap will automatically close or that the feedback loop has already formed.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• China robotics/embodied intelligence supply chain

  Thematic asset of the report

  Strengths

  Improved mass production, delivery, and integration capabilities; ample capital investment; policy support and data-factory construction provide early demand.

  Weaknesses

  Low share of productive demand, industrial ROI not yet validated, large real-data gap, and shipments may not necessarily convert into profits.

  Comparison

  The manufacturing side is compared with the 2019-20 inflection point of the new energy vehicle supply chain, but the demand side has not yet shown a similarly clear commercial closed loop.

  Risks

  If repeat orders from non-affiliated industrial customers fail to emerge for an extended period, the industry may remain at a supply-side inflection point rather than a demand-side one.

• Humanoid robot OEMs

  Beneficiaries of shipments and mass-production capability

  Strengths

  Vendors such as AgiBot have demonstrated strong mass-production capability, and on-site deployments at WAIC have shifted from performances to actual services.

  Weaknesses

  Current demand mainly comes from entertainment, consumer use, government procurement, education, and pilots, with insufficient procurement from industrial operating budgets.

  Comparison

  Compared with traditional industrial robotic arms, humanoid robots do not yet have clear cost and reliability advantages in structured workstations.

  Risks

  Falling ASPs, declining rental rates, weaker pilot budgets, and a possible slowdown in data-collection procurement in 2027F may pressure profitability.

• Dexterous hands and core components

  The fastest-iterating hardware subsegment

  Strengths

  Rich technology routes, with low-DOF linkage hands already scaled, continued iteration in high-DOF tendon-driven and direct-drive solutions, and OEM in-house development driving demand.

  Weaknesses

  High-DOF solutions face lifespan, thermal management, cost, and overengineering issues, with supply ahead of end demand.

  Comparison

  Linkage hands are more commercialized, tendon-driven is more anthropomorphic, direct-drive is better suited to force control and reinforcement learning, and hybrid architectures are increasing.

  Risks

  If end industrial scenarios do not require complex five-finger manipulation, demand for high-end dexterous hands may come in below technology-route expectations.

• Robot data collection and training infrastructure

  A key constraint for the maturity of robot “brains”

  Strengths

  Data collection equipment, exoskeleton gloves, teleoperation, simulation, and evaluation platforms are becoming independent product categories, while national and industry standards are beginning to be established.

  Weaknesses

  High-quality real physical interaction data is scarce, and deployment feedback still accounts for a low share of the current data mix.

  Comparison

  China’s route is more oriented toward real-machine teleoperation, while the U.S. route is more simulation-oriented; the report believes real-data demand can decline but cannot disappear.

  Risks

  Declining data prices may improve costs, but if the collection-training-deployment-feedback closed loop cannot be formed, the maturity of robot “brains” will still be delayed.

Key data

• Number of humanoid robots on site at WAIC 2026Nearly 60 unitsThese robots performed actual services in the venue hosting 1,100+ exhibitors.
• AgiBot cumulative output15,000 unitsThe report cites this as evidence of improved mass-production capability.
• 2026E industry shipment forecast45,000-50,000 unitsBelow the media narrative of about 100,000 units.
• 2026E demand mixEntertainment/performances about 30%, consumer about 30%, government procurement about 20%, education about 15%, industrial/commercial about 5%Demand is still dominated by non-productive use cases.
• YoY change in humanoid robot daily rental feeDown 50-60% to CNY4,000-20,000Falling prices indicate the economics of rental demand are relatively fragile.
• 1H26 humanoid robot OEM financingCNY70.5bn, 98 dealsAccording to GGII, accounting for 53% of total industry financing.
• Embodied foundation model financingCNY22.3bn, 40 dealsFunding flows show capital is betting early on a model-driven capability inflection point.
• World model financingCNY19.1bn, 17 dealsThe report emphasizes that competition in robot “brains” is accelerating.
• Data infrastructure financingCNY4.4bn, 19 dealsData collection and infrastructure are becoming important funding directions.
• High-quality physical interaction data gap2026E demand about 10mn hours, global high-quality stock about 0.5mn hours, about a 20x gapEach deliverable skill requires about 2,000-5,000 hours of data.
• 2026F China data collection market sizeAbout CNY4-5bnThe report expects more than 20 city-level data factories to be under construction, with unit prices potentially falling within 12-24 months.
• Linkage dexterous hand ASP and shareCNY12,000-13,000, more than 80% of industry shipmentsLow-DOF linkage hands are the incumbent mainstream route currently being commercialized.
• Tendon-driven dexterous hand15-19 degrees of freedom, about CNY38,000 ASP, lifespan about 2 monthsCloser to the human hand, but with tendon creep and lifespan issues.
• Direct-drive dexterous hand20+ degrees of freedom, CNY33,000-34,000, expected volume ramp in 4Q26Algorithm teams prefer it for per-joint force control and reinforcement-learning friendliness, but it is constrained by motor thermal management.

Impact & implications

From an investment perspective, the report believes the robotics theme has entered the order-validation stage, and shipment growth alone is insufficient to support profit realization. More important variables to track include order composition, whether paying customers are independent, the share of to-G and related-party demand, unit economics, repeat orders, real training efficiency after data collection costs decline, and whether robot “brains” can form a sustainable feedback closed loop. Supply-chain opportunities may concentrate in mass-producible hardware, dexterous-hand iteration, data collection infrastructure, and world-model integration, but if demand remains concentrated in demonstrations, consumer use, and government projects, valuations and financing may run ahead of fundamentals.

Risks

• Current humanoid robot shipments may mainly reflect manufacturing capability and policy funding support rather than sustainable market demand.
• Industrial customer ROI, MTBF, and cycle time have not reached adoption thresholds, which may cause productive-scenario penetration to be slower than expected.
• Daily rental prices have fallen sharply year over year, indicating fragile economics in some non-productive demand.
• If government procurement and data-factory demand weaken in 2027F, near-term shipments may come under pressure.
• The gap in high-quality real physical interaction data is large, and robot “brains” may take years to mature.
• High-end dexterous-hand technology routes face lifespan, heat dissipation, cost, and overengineering risks.
• Capital formation is clearly running ahead of demand validation, and industry valuations may be pulling forward the capability inflection point.

What to watch

• Whether non-affiliated industrial customers begin placing first repeat orders, and whether those are paid from operating budgets rather than pilot budgets.
• Changes in the share of paying customers, to-G customers, related parties, and display-oriented procurement within the order mix.
• Changes in humanoid robot ASPs, daily rental fees, gross margins, and payback periods.
• MTBF, cycle time, and ROI data in industrial scenarios versus labor/robotic-arm replacement.
• Whether government procurement and data-factory procurement slow in 2027F.
• Changes in hours of high-quality physical interaction data, per-skill data requirements, and data collection unit prices.
• Progress in VLA and world-model integration, and the share of real deployment feedback data within training data.
• Shipment share, ASP, lifespan, and thermal-management improvements across linkage, tendon-driven, direct-drive, and hybrid dexterous-hand architectures.