Physical AI refers to systems that move beyond generating content into operating in real environments: robots, autonomous machines and the foundational models that teach them how to behave. The category has been building momentum for years, but 2026 marks a credible inflection point. At CES in January, Nvidia CEO Jensen Huang declared that the ChatGPT moment for robotics had arrived, suggesting the combination of AI models and computing infrastructure could soon unlock large-scale commercial adoption.
New Battleground
The most consequential investments in physical AI right now are not in the robots themselves. They are in the models that teach machines how to interpret and respond to the world around them. Nvidia has released Cosmos and GR00T, open models built specifically for robot learning and reasoning, alongside the Blackwell-powered Jetson T4000 module designed to bring that intelligence to the industrial edge. The company is building what amounts to a robotics operating system, one that partners including Boston Dynamics, Caterpillar and LG Electronics are already deploying on top of.
Google brought its robotics software unit Intrinsic fully in-house from Alphabet earlier this year, positioning the company to offer a vertically integrated stack running from foundation models through deployment software to cloud infrastructure.
The platform that lowers the expertise barrier the furthest will accumulate the most developers, the most hardware partners and ultimately the most operational data, which is precisely where the AGI argument connects back to the infrastructure race.
Data Advantage
The foundation model race is running in parallel with a harder-edged competition over deployment volume and hardware economics. Global humanoid robot installations reached approximately 16,000 units in 2025, with China accounting for more than 80% of deployments across logistics, manufacturing and automotive applications. Morgan Stanley data shows China filed 7,705 humanoid patents over five years, five times the U.S. total and accounted for 54% of global industrial robot installations.
Those numbers reflect structural advantages in component manufacturing. Chinese robotics firms produce motors, sensors and harmonic reducers domestically and at costs that enable rapid iteration. Unitree shipped roughly 36 times more units last year than U.S. rivals Figure and Tesla combined.
The dynamic mirrors what happened in electric vehicles: Domestic scale creates a cost floor that is difficult to undercut once established, and the data generated by those deployments compounds over time.
That last point is where the AGI thesis becomes concrete. Robots deployed at scale in real environments generate the data that makes the next generation of models smarter. The intelligence layer does not improve in isolation; it improves through contact with the physical world. Whoever deploys the most robots accumulates the most training signal, which strengthens the models, which makes the next generation of robots more capable.
Tesla’s Optimus robot program, for example, is expected to begin deployments in Tesla factories where machines can perform tasks while collecting data used to refine software systems.
A Deloitte survey of more than 3,200 global business leaders found that 58% are already using physical AI in some form, with that figure expected to reach 80% within two years. The adoption curve is real, and the intelligence layer being built on top of it will determine which platforms run the next phase of industrial AI. Musk’s AGI comment would be easy to dismiss. The infrastructure being built around that thesis is considerably harder to ignore.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.