WAIC 2026 shows China’s Physical AI has entered an early stage with greater emphasis on deployment

Overview

This report summarizes Morgan Stanley’s observations from WAIC 2026, with the core theme that Physical AI in China continues to iterate rapidly. The report covers industrial AI solutions, humanoid robots and robotics, and new AI hardware, arguing that AI is moving from a purely software form toward the perception layer, embodied form, and execution layer, thereby providing incremental support for industrial capital expenditure.

Core views

The report’s core views include: first, industrial AI agent systems are becoming an emerging application layer, currently used mainly for operational monitoring and decision support and gradually expanding into engineering automation; second, industrial AI deployment requires equipment digitalization and data connectivity, which may drive a longer-cycle upgrade of industrial equipment; third, humanoid robot deployment is this year’s main theme, but early-stage scenarios are concentrated, generalization is limited, and deployment is engineering-intensive; fourth, data volume, data diversity, and data quality are key bottlenecks for the development of robotics foundation models; fifth, large technology companies are entering the embodied intelligence ecosystem and may in the future focus more on the brain and capability layers.

Analysis framework

The report is based on on-site observations at WAIC 2026, the structure of participating companies, exchanges with solution providers and technical personnel, and cross-sectional comparisons across themes such as industrial AI, humanoid robots, data infrastructure, world models, and sensors, in order to assess industry deployment progress, business model changes, and investment implications for the supply chain.

Methodology notes

• Industry chain analysisPhysical AI layered framework

  From the perception layer and embodied form to the execution layer

  The report interprets Physical AI as the process by which AI extends from pure software into hardware and robotics, first manifesting in the perception layer and embodied carriers, and then moving into the execution layer of robots or other hardware.

• Business model analysisMigration from project-based to subscription-based models

  Recurring revenue from industrial AI agents

  The report believes that hardware sales combined with ongoing post-training models and token consumption provide a foundation for higher annual subscription and service revenue, although this transition is still at an early stage.

• Technology path comparisonComparison of VLA/RL and world models

  Trade-off between deployment efficiency and generalization capability

  The report believes world models have better generalization potential, but at present VLA combined with real-world post-training and reinforcement learning is more practical for specific tasks, and the industry may maintain a hybrid path before sufficient data accumulation.

Asset mapping & comparison

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

• China industrial equities/industrial automation sector

  Theme directly covered

  Strengths

  Industrial AI deployment requires equipment digitalization, data interconnection, and automation upgrades, which may extend the industrial capital expenditure cycle.

  Weaknesses

  Deployment is still at an early stage, and limited customer data openness, poor equipment interoperability, and project-based delivery constrain the pace of expansion.

  Comparison

  Compared with pure software AI, industrial AI relies more on hardware retrofits and scenario deployment, but is also more likely to generate demand for equipment upgrades.

  Risks

  Weaker-than-expected demand validation, customer budget fluctuations, and long equipment connectivity retrofit cycles.

• Humanoid robotics and embodied intelligence supply chain

  One of the core beneficiary directions

  Strengths

  Real-world deployment has become industry consensus, and scenarios such as handling, sorting, inspection, wiring harnesses, and insertion already have ongoing exploration.

  Weaknesses

  Generalization capability is limited, the early-stage market is fragmented, and deployment requires extensive engineer involvement.

  Comparison

  At present, VLA plus post-training/RL is closer to real-world deployment, while world models are more oriented toward medium- to long-term breakthroughs in generalization.

  Risks

  Data bottlenecks, high costs, non-convergent technology paths, and slow commercialization pace.

• Data infrastructure and robotics model companies

  Critical bottleneck segment

  Strengths

  Egocentric video, UMI, and outsourced data collection services help expand the supply of training data, while proprietary data pipelines become a source of differentiation.

  Weaknesses

  Outsourced data collection services are highly homogeneous and may commoditize quickly, leading to price declines.

  Comparison

  General data service providers have limited differentiation; proprietary pipelines with capabilities in annotation, task decomposition, context analysis, and quality rating are more valuable.

  Risks

  Insufficient data quality, price deflation, customer concentration, and compliance and privacy constraints.

Key data

• Number of participating companies at WAIC 2026>1,000 companiesThe conference was held in Shanghai from July 17 to 20, 2026.
• Share of robots/smart hardware among participants24%According to Jazzyear, this was the most crowded segment.
• Share of computing, chips, and infrastructure among participants20%This was the second-largest participation category.
• Share of industry AI applications among participants13%Reflecting that industrial and sector-specific applications are still accelerating.
• IndustrialNext production switch timeAs low as 8 hoursIts flexible manufacturing model combines ViT and VLA, enabling adaptive manufacturing processes across different products and customers.
• Robbyant supported hardware platforms20+ robotic hardware platformsAnt Group subsidiary Robbyant launched a model suite covering VLA, VA, and world models.

Impact & implications

From an investment perspective, the report points to potential opportunities in industrial automation, robot bodies, data collection and labeling, industrial software, sensors, AI hardware, and equipment digitalization upgrade chains. More visible short-term gains come from deployable industrial AI, equipment connectivity retrofits, and robotic applications for specific tasks; medium- to long-term opportunities depend on data accumulation, model generalization capability, the intensity of entry by major technology companies, and whether subscription-based business models can scale.

Risks

• Industrial AI and humanoid robots are still at an early stage, and productivity gains and commercialization scale still need validation.
• China’s installed industrial equipment base has insufficient connectivity capability and poor cross-device interoperability, which may lengthen the AI deployment cycle.
• Robotics foundation models are constrained by data volume, diversity, and quality, and insufficient generalization capability may limit the scope of applications.
• Humanoid robot deployment is highly engineering-intensive, and market fragmentation may suppress short-term scale efficiency.
• The cost of vision-tactile sensors remains high, and technical issues such as durability, size, and algorithm interoperability have not been fully resolved.
• The entry of large technology companies may alter the competitive landscape and compress the profit margins of startups or traditional integrators.

What to watch

• Whether industrial AI agents expand from operational monitoring to engineering automation and closed-loop execution.
• The capital expenditure intensity of customer equipment digitalization upgrades and data connectivity retrofits.
• Whether the industrial AI business model shifts from one-off project fees to annual subscriptions and recurring service revenue.
• The progress of scaled deployment of humanoid robots in mainstream scenarios such as handling, sorting, inspection, wiring harnesses, and insertion.
• Whether egocentric video, UMI, and proprietary data pipelines can alleviate the training data bottleneck for robotics.
• The convergence of performance and cost across world model, VLA, VA, and reinforcement learning approaches.
• The embodied intelligence strategies of major technology companies such as Ant Group, JD.com, Tencent, and StepFun.