Physical AI is becoming a contest over who controls the intelligence inside machines. In the West, infrastructure companies are building model, chip and software platforms. In China, manufacturers are combining growing software capability with an enormous robotics supply chain. The two strategies are converging on factories, vehicles and other systems that act in the real world.
Nvidia chief Jensen Huang has called the shift a “ChatGPT moment for robotics.” The comparison marks a move from specialist research toward broad commercial deployment. A Deloitte survey cited by the source found 58% of more than 3,200 global business leaders already using physical AI in some form, rising to 80% when plans for the next two years are included.
The question is increasingly not whether companies will adopt physical AI, but which layers they will depend on when they do.
Western companies compete to own the platform
Nvidia has released Cosmos and GR00T models for robot learning alongside its Blackwell-based Jetson T4000 module, which it says delivers four times greater energy efficiency for robotics computing. Arm has created a dedicated Physical AI business unit focused on chip design for robots and intelligent vehicles.
Siemens and Nvidia plan an Industrial AI Operating System and an adaptive, AI-driven manufacturing site. Google has moved Intrinsic from Alphabet’s experimental portfolio into its core business, bringing robotics deployment software closer to DeepMind models and Google Cloud.
The strategy resembles an operating-system contest. A platform provider does not need to manufacture every robot if models, simulation tools and deployment software become the common layer across many machines. Google’s internal Android analogy captures the ambition: make the environment on which an ecosystem builds.
For industrial customers, shared platforms can reduce programming and integration work. They also create long-term dependencies. Once workflows, data and safety processes are tied to one stack, changing provider may be far more difficult than replacing a conventional software tool.
China combines scale with a complete supply chain
China’s advantage is more physical. The country accounted for over 80% of global humanoid installations in 2025 and more than half of industrial robots. It controls roughly 70% of the lidar sensor market and leads production of harmonic reducers, the precision gears central to robot movement.
More than 140 domestic humanoid manufacturers have unveiled over 330 models. Competition and manufacturing volume can push hardware costs down in the same way China’s electric-vehicle industry achieved rapid scale.
Alibaba has joined the foundation-model layer with open-source RynnBrain, designed to help robots understand physical environments and identify objects. It competes conceptually with Nvidia Cosmos and Google DeepMind’s Gemini Robotics while giving developers another open option.
Public demonstrations, including humanoids performing complex routines during China’s Spring Festival Gala, are partly spectacle. They also signal that coordination and hardware quality are improving quickly. The more consequential work will happen in logistics and factories, where reliability and cost can be measured over thousands of shifts.
Industrial control becomes strategically important
Western software platforms and Chinese manufacturing strength are not separate stories. Their collision will determine supply-chain dependencies, data sovereignty and who sets interfaces for a new generation of automated equipment.
Historically, installing industrial robots required months of custom programming and specialist expertise. New platforms aim to compress that work into days. Vention, which raised $110 million in January, claims its tools can make that change. If such timelines become routine, automation could reach smaller manufacturers and more variable tasks.
Lower barriers also make platform governance more important. A robot operating layer can influence which models, chips and data formats a factory uses. Updates or restrictions imposed upstream may affect physical production rather than only an office application.
Businesses should therefore evaluate openness, portability and support alongside headline capability. They need to understand where operational data travels, which components can be replaced and how safety evidence survives a model update. Governments will view the same choices through industrial resilience and national capacity.
Physical AI’s current momentum comes from several advances arriving together: capable models, efficient edge compute, synthetic simulation and cheaper hardware. No single company controls all of them. The winners may be those that connect the pieces without making customers surrender too much flexibility.
This is why the race matters beyond impressive robots. The software and hardware standards taking shape now could sit beneath how products are made, goods are moved and infrastructure is maintained for years. Physical AI is becoming industrial infrastructure—and control of that infrastructure will carry economic and geopolitical weight.
