Physical AI and the Global Race for Sovereign Robotics

Introduction

Robotics has moved well beyond the fixed-arm assembly lines that defined factory automation for half a century. Physical AI — the application of artificial intelligence to machines that perceive, reason and act in the real world — is now the subject of direct industrial policy in Washington, Beijing, Tokyo, Seoul and Brussels. Governments are no longer simply funding research; they are setting unit deployment targets, backing sovereign foundation models, and treating robotics supply chains as a matter of national competitiveness. For enterprise leaders and institutional investors, understanding this shift matters because the winners will not necessarily be decided by whoever builds the most capable robot first, but by whoever can scale physical AI reliably across real industrial and service environments. This article maps where each major economy stands, what the verified numbers actually show, and where the genuine bottlenecks remain.

Key Takeaways

• Physical AI has become a matter of explicit state industrial policy, not just private-sector innovation, with the US, China, Japan, South Korea and the EU each pursuing distinct strategic approaches.
• China’s Ministry of Industry and Information Technology, together with the state asset regulator SASAC, has mandated 10,000 humanoid robot deployments and more than 100 high-value industrial applications by the end of 2026.
• Japan’s Noetra consortium, backed by Sony, SoftBank, NEC and Honda with government funding of up to $6.2 billion, is targeting 10 million AI-equipped robots deployed across 18 sectors by 2040.
• Goldman Sachs Research has revised its humanoid robot market forecast sixfold, from $6 billion to $38 billion by 2035, citing faster-than-expected AI progress and falling manufacturing costs.
• The primary technical constraint remains hardware, not AI models: precision actuators, harmonic reducers and tactile sensors are concentrated among a small number of specialised suppliers worldwide.
• Long-range market forecasts vary enormously by bank, timeframe and definition, and should be treated as scenarios rather than settled projections.

The United States: Foundation Models and Direct Deployment

The US retains a significant lead in the underlying intelligence layer, driven by multimodal Vision-Language-Action (VLA) models that convert visual and spatial data directly into physical movement commands. Commercial pilots led by companies including Tesla, Figure AI and Boston Dynamics are focused on moving humanoid platforms out of controlled laboratory demonstrations and into logistics hubs and automotive assembly environments. The US approach has generally been characterised by private capital and corporate R&D leading deployment, rather than a single centralised government mandate, though public investment in AI infrastructure more broadly continues to support the underlying compute layer these models depend on.

China: Industrial Scale and State-Directed Deployment

China’s strategy is built around manufacturing density and state-coordinated deployment rather than model capability alone. In June 2026, the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission jointly mandated a target of 10,000 humanoid robot deployments and more than 100 high-value industrial application scenarios across Chinese industry by the end of 2026, requiring local governments and state-owned enterprises to submit implementation plans on a fixed schedule. The programme is explicitly designed to feed large volumes of real-world operational data back into robot foundation models, accelerating the transition from laboratory demonstration to genuine commercial reliability.

China’s manufacturing base gives this strategy real teeth. The country hosted the inaugural World Humanoid Robot Games in Beijing in August 2025, in which 280 teams from 16 countries competed with more than 500 humanoid robots across 26 events, including athletics, football and household task scenarios. Beyond the spectacle, the event served a genuine industrial purpose: benchmarking dynamic stability, balance recovery and task completion under conditions closer to real deployment than most laboratory testing allows.

Japan: Sovereign Foundation Models for an Ageing Workforce

Japan’s approach is distinct in being explicitly framed around demographic necessity. In July 2026, Noetra Corp. launched full-scale research and development of a Japan-developed multimodal foundation model for physical AI and robotics, backed by founding members Sony, SoftBank, NEC and Honda, with investment from 44 companies in total. Japan’s Ministry of Economy, Trade and Industry has committed up to ¥1 trillion (approximately $6.2 billion) over five years to back the programme, with an initial tranche of ¥387.3 billion allocated for fiscal 2026. The government’s stated ambition is genuinely large in scope: 10 million AI-equipped robots deployed across 18 sectors, including eldercare, hospitality, food service and medical support, by 2040. Japan currently holds a substantial share of the global industrial robotics market, giving the initiative a credible manufacturing base to build from.

South Korea and the European Union: Targeted Industrial Strategies

South Korea’s robotics strategy centres on its Fourth Intelligent Robot Basic Plan (2024–2028), announced by the Ministry of Trade, Industry and Energy. The plan targets over $2.24 billion in combined public and private investment by 2030, with a specific goal of raising the domestic manufacturing rate of core robot components from 44% to 80% over the same period. South Korea’s approach leans heavily on its existing strength in semiconductor manufacturing, embedding compute directly into collaborative robots (cobots) designed for high-speed manipulation tasks.

The European Union’s strategy is more narrowly focused on precision engineering and regulatory compliance than on mass deployment targets. High-precision cobots for automotive and industrial human-robot interaction remain the EU’s primary area of competitive strength, developed within the compliance framework established by the EU AI Act, which applies horizontally across AI systems including those embedded in physical robots.

Applications Across Enterprise and Public Sectors

Physical AI’s commercial relevance now extends well beyond factory floors. In defence, unmanned ground vehicles and quadruped platforms are increasingly used for logistics and reconnaissance in contested environments, reducing personnel exposure to risk. In semiconductor fabrication, cleanroom-compatible robotic systems handle wafer transport and precision chemical processes that require tolerances beyond typical human manual capability. In healthcare, robotic-assisted surgical systems and motorised exoskeletons for rehabilitation are among the more mature commercial applications of precision robotics, while in agriculture, vision-guided autonomous systems are being deployed for selective harvesting tasks. Each of these applications depends on the same underlying hardware and AI stack, which is why bottlenecks in one sector tend to ripple across all of them.

Comparison Table: National Robotics Strategies at a Glance

The Real Bottleneck: Hardware, Not Intelligence

The industry’s dominant technical narrative has shifted. Multimodal Vision-Language-Action models can now convert unstructured environmental data into coherent action plans with genuine sophistication, and this capability is generally not the limiting factor it was even two years ago. Execution is throttled instead by the physical hardware layer: high-precision strain wave gears and harmonic reducers, high-torque frameless motors, multimodal tactile sensor arrays, and energy-dense battery systems remain concentrated among a small number of specialised component suppliers globally. This concentration is precisely why several governments, particularly China and South Korea, have made domestic component manufacturing capacity an explicit policy target rather than leaving it purely to market forces.

Market Forecasts: Treat Long-Range Numbers as Scenarios

Long-range market size projections for humanoid robots vary enormously depending on the bank, the timeframe and precisely what is being measured, and readers should treat every figure below as a scenario rather than a settled fact. Goldman Sachs Research has revised its total addressable market forecast for humanoid robots sixfold, from an original $6 billion to $38 billion by 2035, citing faster-than-expected progress in AI training methods and a roughly 40% reduction in manufacturing costs. The bank’s base case anticipates more than 1.4 million unit shipments by 2035. Other banks, including Morgan Stanley, have published considerably larger figures for a longer horizon and broader definition, projecting a total humanoid robotics ecosystem — including hardware, supply chains and downstream services — reaching several trillion dollars by 2050, a fifteen-year longer timeframe and a much wider scope than Goldman Sachs’s narrower 2035 hardware estimate. The gap between these figures illustrates why comparing headline numbers across reports without checking the underlying definition and timeframe is a common and consequential mistake.

Common Mistakes Organisations Make

Comparing market forecasts without checking their definitions. A $38 billion hardware TAM by 2035 and a multi-trillion-dollar “ecosystem” projection for 2050 are not comparable figures, yet they are frequently cited interchangeably in industry commentary.

Assuming AI model capability is now the binding constraint. Component-level hardware bottlenecks, particularly in precision actuators and sensors, remain the more significant near-term limitation on deployment scale.

Underestimating the role of state industrial policy. Treating physical AI purely as a private-sector technology race overlooks how directly China, Japan and South Korea are using state coordination, funding and deployment mandates to shape outcomes.

Overlooking data collection as a competitive advantage. China’s state-directed deployment programme is explicitly designed to generate real-world operational data at scale, which compounds as a competitive advantage independent of raw model capability.

Ignoring regulatory divergence. The EU’s compliance-first approach under the AI Act creates a materially different deployment environment from China’s deployment-first, data-driven model, with direct implications for how quickly products can reach market in each region.

Future Trends: The Next Three to Five Years

Expect state industrial policy to become more explicit and more binding, following the pattern set by China’s MIIT/SASAC deployment mandate and Japan’s METI-backed Noetra funding, rather than remaining a matter of general research grants. Hardware component supply, particularly for precision actuators and sensors, is likely to see significant new investment as governments recognise it as the genuine bottleneck ahead of AI model capability. Cross-border standards and safety frameworks will likely need to reconcile increasingly divergent regulatory philosophies, particularly between the EU’s compliance-first model and China’s deployment-first approach. Market forecasts will almost certainly continue to be revised, in both directions, as real-world deployment data either validates or challenges the more optimistic scenarios currently in circulation. Enterprise adoption is likely to concentrate first in structured, high-volume, lower-liability environments such as warehousing, semiconductor fabrication and controlled manufacturing settings, before expanding into less predictable service environments.

Frequently Asked Questions

What is physical AI, and how is it different from traditional robotics? Physical AI refers to AI systems, typically built on multimodal Vision-Language-Action models, that allow robots to perceive their environment and generate appropriate physical actions dynamically, rather than following pre-programmed, fixed sequences as traditional industrial robots do.

Is China really deploying 10,000 humanoid robots by the end of 2026? Yes, this is a verified target set jointly by China’s Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission in June 2026, requiring local governments and state enterprises to submit implementation plans on a fixed timeline.

What is Japan’s Noetra project? Noetra is a Japanese sovereign AI consortium backed by Sony, SoftBank, NEC and Honda, developing a homegrown multimodal foundation model for physical AI and robotics, with government funding of up to $6.2 billion and a target of 10 million deployed AI-equipped robots by 2040.

How big is the humanoid robot market expected to become? Estimates vary widely by definition and timeframe. Goldman Sachs projects a $38 billion total addressable market by 2035 for humanoid hardware specifically, while broader “ecosystem” forecasts covering hardware, supply chains and services from other banks extend to considerably larger figures over a longer horizon to 2050.

What is the biggest obstacle to scaling humanoid robots? Current evidence points to hardware, not AI software, as the primary constraint. Precision components such as strain wave gears, harmonic reducers and tactile sensor arrays remain concentrated among a small number of specialised global suppliers.

Why is South Korea investing in robotics component manufacturing specifically? South Korea’s Fourth Intelligent Robot Basic Plan aims to raise the domestic manufacturing rate of core robot components from 44% to 80% by 2030, reducing reliance on imported parts and leveraging the country’s existing semiconductor manufacturing strength.

How does the EU’s approach to robotics differ from China’s? The EU prioritises regulatory compliance and precision engineering within the framework of the EU AI Act, while China pursues a deployment-first, state-coordinated model designed to generate real-world operational data at industrial scale as quickly as possible.

Final Thoughts

The physical AI race is genuinely global, but it is not being run on a single track. China is betting on scale and state-directed data collection, Japan on sovereign models addressing a demographic problem it cannot solve any other way, South Korea on semiconductor-enabled component manufacturing, and the EU on precision and compliance. The United States retains the clearest lead in underlying model capability but has left deployment largely to private capital rather than centralised mandate. For enterprise leaders and investors, the more useful question is not which nation or company currently leads, but which combination of state support, hardware supply security and real-world deployment data proves most durable once the current wave of ambitious targets meets the harder reality of manufacturing at scale.