The competition for artificial intelligence is entering a new phase, and this time it's not just about who has the biggest computers. While the United States still leads in training large language models (LLMs), China is building a commanding advantage in what experts call "embodied AI" or "physical AI," the systems that allow robots and machines to learn from the real world. The shift matters because the substrate for training these systems is fundamentally different: instead of text and code, physical AI systems need streams of sensory data generated by robots actually working in factories, warehouses, and other environments.

What's the Difference Between Language AI and Physical AI?

Large language models, the technology behind ChatGPT and similar systems, work by processing vast amounts of text, code, and images to recognize patterns and generate responses. They reason about the world, but they don't act in it. A language model can describe how to assemble a machine or optimize a workflow on paper, but it cannot learn from the friction, errors, and variations that occur during actual assembly.

Physical AI systems, by contrast, learn by doing. They perceive their environment through sensors, predict what will happen next, take action, and adjust their behavior based on feedback. This approach, sometimes called "world models" or "spatial intelligence," gives these systems access to causal signals of success or failure that static datasets cannot provide. The technical literature converges on a working definition: world models are internal representations of environmental dynamics that allow an agent to perceive, predict, and act.

The practical implication is straightforward: to train physical AI systems at scale, you need data from physical systems working in real environments. Every movement, adjustment, error, and correction feeds back into learning. This kind of data accumulates only through deployment.

Why Is China Pulling Ahead in the Physical AI Race?

China has built a structural advantage in the substrate that physical AI systems require: deployed robots generating training data. According to the International Federation of Robotics, China installed a record 295,000 industrial robots in 2024, representing 54 percent of global deployments and nearly nine times as many as the United States. The gap in the installed base is even starker. Fully 2,027,200 factory robots operate in China, roughly five times the 393,700 in the United States.

This massive deployed base means Chinese companies have access to vastly more real-world sensory data for training physical AI systems. Chinese firms such as Unitree are also commercializing humanoid platforms at prices that no Western equivalent can match, further accelerating deployment and data accumulation.

The strategic significance was not lost on German Chancellor Friedrich Merz when he visited China in February. Chinese President Xi Jinping greeted him with a demonstration of humanoid robots dancing, sparring, and moving in perfect synchrony. The carefully scripted performance reflected what a concerted industrial and political system can build and deploy at scale.

How Are the U.S. and Other Nations Responding?

The U.S. government is mobilizing to address the gap. The National Science Foundation has opened a proposal window for its X-Labs Initiative, seeking teams to develop platform technologies that advance intelligent, adaptive AI systems capable of perceiving, acting, and learning within complex physical environments. The solicitation targets early-stage robotics, sensing, and human-robot interface platforms, with written proposals due through January 7, 2027.

The funding structure reflects the urgency. Awards will follow a milestone-based framework reaching $50 million per year, bypassing standard federal procurement to grant teams high operational flexibility. Instead of disbursing funds on a fixed calendar schedule, the agency ties all payments directly to documented progress and verifiable exit criteria.

• Phase 0 (9 to 12 Months): Selected teams will receive up to $1.5 million to refine technical concepts, establish governance structures, and prepare for full operational deployment.
• Phase 1 (24 to 36 Months): Following a Go/No Go selection during Phase 0, advancing teams can secure up to $50 million annually to scale full-time personnel, execute critical progress milestones, and overcome technical bottlenecks.
• Phase 2 and Beyond: High-performing teams that accomplish Phase 1 milestones may receive Phase 2 funding to mature platform solutions toward widespread adoption or commercial transition.

Key focus areas for the NSF initiative include distributed learning and swarming, edge AI, bio-inspired or bio-hybrid robotics control, and training data development for real-world variability, with targeted applications spanning healthcare, national defense, emergency response, advanced manufacturing, and scientific discovery.

Meanwhile, private capital is flowing toward physical AI startups. Yann LeCun, formerly Meta's chief AI scientist, raised $1.03 billion at launch for his new venture AMI, describing its mission as building world-model-based AI "that understands the real world," with applications in industry, robotics, healthcare, and automation. World Labs announced a separate $1 billion round in February to advance spatial intelligence and build world models for robotics and scientific discovery.

What Role Will Nvidia Play in Physical AI?

Nvidia, which controls an estimated 80 to 90 percent of the AI GPU (graphics processing unit) segment, is positioning itself to play a similarly central role in physical AI's training, simulation, and deployment. During its August earnings call, CEO Jensen Huang pointed to "physical AI coming online" as one of the forces accelerating demand for AI infrastructure.

Nvidia's strategy extends well beyond supplying chips for robots. The company is assembling infrastructure, simulation, models, software, edge computing, and safety technology into what it describes as a full-stack physical AI platform. The company's three-computer architecture involves training, simulating, and deploying physical AI systems. The first part is the computing infrastructure used to train AI models. The second uses Nvidia's Omniverse and Cosmos platforms to simulate physical environments, generate data, and test AI systems before deploying them in the real world. The third brings computing to the edge, where platforms such as Jetson run AI workloads inside robots and other autonomous systems.

"Nvidia provides the full-stack platform for physical AI. We bring together AI infrastructure, open simulation frameworks, open models, edge computing and safety so developers and companies can train, test, deploy and continuously improve robots and autonomous systems," said Sasa Docca, head of robotics product marketing at Nvidia.

Sasa Docca, Head of Robotics Product Marketing at Nvidia

Nvidia expanded its edge computing portfolio in August with Jetson Orin Nano 2, an entry-level robotics computer that runs AI workloads at the edge. The company is targeting the platform for robots, delivery and inspection drones, and vision AI systems, with the module and developer kit expected to be available in the first half of 2027.

However, replicating its dominance in generative AI may prove challenging. Physical AI brings Nvidia into a more fragmented market with established semiconductor and industrial technology competitors, while real-world deployments must overcome challenges involving safety, power, latency, integration, and cost.

What Are the Critical Bottlenecks in Physical AI Development?

Despite the investment and momentum, significant technical and logistical challenges remain. While China leads in deployed robots, it still depends on Western technology for critical components. Software for training, sensing, and control remains dominated by firms based outside China, most importantly Nvidia, whose ecosystem underpins nearly every major robotics company in the country. High-precision mechanical components represent another dependency. Bosch Rexroth, Schaeffler, THK, and NSK control roughly 90 percent of the high-end ball screw and precision motion market on which Chinese manufacturers rely.

At the level of autonomous function, the gap is wider still. Chinese humanoids are mostly deployed in narrow tasks and site-specific trials, a far cry from full autonomy in unpredictable environments.

Another layer that cuts across both language AI and physical AI is critical minerals and energy. Rare earths are embedded in AI infrastructure, from the permanent magnets in data center cooling systems to the actuators that give robots their range of motion. China is the dominant refiner of 19 of the 20 minerals in the International Energy Agency's Global Critical Minerals Outlook 2025, with an average market share around 70 percent. In permanent magnets, the concentration is near total. China's share of sintered magnet production has climbed from roughly 50 percent two decades ago to 94 percent today.

How Should Organizations Prepare for Physical AI?

For companies and research institutions looking to participate in the physical AI ecosystem, several practical steps can help position them for success:

• Develop Real-World Data Collection Capabilities: Organizations should invest in systems and processes to collect sensory data from deployed robots and physical systems, since this data is the primary fuel for training physical AI models and represents a structural competitive advantage.
• Build Partnerships Across the Technology Stack: Physical AI requires expertise spanning AI model development, robotics hardware, sensors, safety systems, and edge computing. Organizations should identify and collaborate with partners who can provide complementary capabilities rather than attempting to build everything internally.
• Engage with Simulation and Synthetic Data Tools: Platforms like Nvidia's Omniverse and Cosmos can reduce the cost and time required to generate training data by creating physically based simulations and synthetic scenarios, allowing organizations to accelerate development without waiting for real-world deployment.

The NSF's X-Labs solicitation also provides guidance on what the U.S. government considers strategic priorities. Proposals should target Technology Readiness Levels 2 through 4, ruling out both immature theoretical concepts and technologies nearing commercial scale-up. The agency excludes efforts that duplicate substantial existing public or private research and development investment, and it explicitly flags narrow single-product efforts that do not generalize into reusable platforms, simulation-only work without a sim-to-real transfer plan, and large language model wrappers that control off-the-shelf robots without new embodied capabilities as nonresponsive.

What Does This Mean for Strategic Competition?

The divergence between language AI and physical AI represents a fundamental shift in how technological advantage accumulates. The United States leads on the language AI path, driven by compute and data, but China is ahead in building the substrate for physical AI. Whichever side can integrate both and translate them into deployed capability at scale will gain a structural industrial and military advantage in a contest in which strategic primacy and technological supremacy are becoming one and the same.

The next phase of AI competition will not be decided solely by who trains the largest models or who controls the most powerful chips. It will be decided by who can turn AI systems into machines that learn, adapt, and act reliably in the physical world at scale. For now, that advantage belongs to China, but the U.S. government and private sector are mobilizing to close the gap.