Nvidia delivered another blockbuster earnings report this week, but one line in its outlook was easy to miss. The company said its guidance assumes no data center compute revenue from China.
That is different from saying Nvidia expects no revenue from China. Far from it. At almost the same moment, The Wall Street Journal reported that China has emerged as an eager customer for Nvidia’s growing physical AI business—the chips, software, models and development tools used to build robots, autonomous vehicles and other machines that perceive and act in the physical world. Nvidia says that business already produces about $10 billion in annual revenue and could grow tenfold over the next decade.
Those two stories beside each other raise an obvious question. If the United States wants to deny China access to cutting-edge American technology in strategic industries, why is it helping China’s robotics industry build on the world’s leading physical AI platform?
The answer reveals more than an inconsistency over which Nvidia chips can be sold in China. It exposes an export-control complex that is still looking at individual components and benchmark thresholds while the technology industry is building integrated platforms. Washington is regulating the pieces. China is assembling the stack.
A Real Distinction That Is Becoming Obsolete
There is a legitimate technical distinction between a data center accelerator and an embedded robotics computer. Thousands of high-end GPUs can be connected into an AI factory capable of training frontier models and supporting military simulation, intelligence analysis, cyber operations and other strategic work. An Nvidia Jetson module normally sits inside one robot or vehicle, operating within strict limits on power, cooling and space.
One Jetson is not a rack of B300s. That part of the policy logic is reasonable.
The problem is that physical AI is not created by one chip performing one isolated task. It is a continuous system. Developers simulate environments, generate synthetic data, train models, test robot policies, deploy those models onto machines and collect real-world operating data that feeds the next development cycle. The intelligence may be trained in a data center, but it is refined through its interaction with the physical world.
Nvidia participates across that loop. It offers Jetson computers for onboard inference, CUDA as the underlying development environment, Isaac for robotics simulation and development, Omniverse for digital twins, Cosmos world models and GR00T robot foundation models. This is not a parts catalog. It is an architecture for developing, training, testing and operating intelligent machines.
Washington’s rules were designed to count chips. Nvidia’s strategy is designed to control platforms.
That strategy should be familiar. Nvidia introduces the hardware, surrounds it with software and libraries, attracts developers and becomes progressively more difficult to replace. The value is not confined to the processor. It accumulates in the tools, trained engineers, development workflows, models and data pipelines built around it.
I have called this the Indispensability Trap: A company supplies a critical layer so effectively that customers build the rest of their operations around it. Before long, replacing the supplier means changing far more than a product. Chinese robotics companies are not simply purchasing American silicon. Their engineers are learning to build physical AI through Nvidia’s way of doing things.
China Has the Bodies. Nvidia Can Help Supply the Brains.
In our recent Techstrong Special Report, “Bodies, Brains and the Real World: The State of AI-Powered Humanoid Robots,” we proposed that a commercially useful humanoid requires three systems to mature together.
The body includes the actuators, hands, sensors, batteries, onboard computing, balance, durability and manufacturing capacity needed to produce a reliable machine. The intelligence includes perception, task understanding, planning, manipulation, learning and error recovery. Deployment encompasses workflow integration, safety, maintenance, fleet management, human exception handling and sustainable economics.
No country currently leads across all three layers. China, however, has moved fastest on the body and manufacturing layer. It benefits from supply chains built for electric vehicles and consumer electronics: motors, batteries, actuators, sensors and precision components available at enormous scale and increasingly aggressive prices.
The report found that China had more than 140 humanoid manufacturers in 2025. AgiBot and Unitree together represented roughly 70% of estimated global humanoid shipments. China also accounted for 54% of new industrial robot installations in 2024, giving it a reservoir of manufacturing experience and potential customers that no other country can match.
That does not mean Chinese humanoids are already the world’s most capable autonomous workers. They are not. The industry is scaling bodies faster than it is proving autonomous labor. Most commercially relevant humanoids still perform bounded tasks in controlled environments with human supervision and exception handling. Chinese platforms generally trail the leading American programs on foundation-model intelligence, adaptable control and independently verified on-the-job performance.
But that is precisely why Nvidia’s role deserves scrutiny. China has become extraordinarily good at building robot bodies. Nvidia can help provide the brains.
The strategic prize in robotics does not belong to the country that ships the most machines or publishes the most impressive demonstration videos. It belongs to whoever first combines capable bodies, adaptable intelligence and credible deployment at commercial scale. Allowing China to build its already formidable manufacturing ecosystem around Nvidia’s intelligence platform could help it close the part of the equation where the United States still holds its clearest advantage.
Accelerate Now, Substitute Later
China plainly considers robotics strategic. Its government has designated embodied AI as an industry of the future. Official plans describe humanoid robots as a potentially disruptive platform comparable to computers, smartphones and electric vehicles. They call for advances not only in robot bodies but in the “brain,” “cerebellum,” specialized chips, operating systems, development tools and a secure domestic supply chain.
There is no public evidence that the Chinese Communist Party issued an explicit instruction saying, “Use Nvidia until our alternatives are ready.” We should not pretend there is. But the observable strategy looks very much like accelerate now, substitute later.
Chinese manufacturers can use Nvidia to shorten development cycles, train engineers, deploy more machines and begin collecting the scarce physical-world data needed to improve robot intelligence. At the same time, Huawei, Horizon Robotics, Black Sesame and other domestic suppliers can work on the processors and software intended eventually to replace Nvidia.
Beijing does not have to choose between Nvidia and technological sovereignty today. It can use Nvidia to accelerate the industry while financing the companies intended to replace it tomorrow.
That arrangement creates very different clocks for the two sides. Nvidia receives revenue and platform adoption now. China receives factories, trained developers, deployed robot fleets, operating experience and data. Even if a domestic processor eventually replaces Jetson, those assets remain in China. Once an industry and its workflows have reached scale, changing the compute module is a more manageable problem than creating the industry from scratch.
Strategic Enough to Keep Out, but Not to Help Build?
The contradiction becomes harder to explain when viewed from the other direction. The United States has moved to restrict new Chinese humanoid and quadruped robot imports because of cybersecurity, surveillance and supply-chain concerns. Washington apparently considers Chinese autonomous machines sufficiently strategic to keep out of American infrastructure.
Yet American technology can still help improve the ecosystem producing those machines.
If the finished robot is strategic enough to keep out, why is the platform helping create it not strategic enough to examine as a whole?
The likely answer is not that policymakers believe robots are unimportant. It is that technology policy remains divided into administrative categories. Advanced computing chips are handled through performance thresholds and export licenses. Robotics modules are treated as embedded products. Open models are treated as research resources. Development frameworks are software. Autonomous vehicles, industrial machines, military end users and foreign investment fall under still other rules and agencies.
Physical AI crosses all of those boundaries. The regulators see separate boxes because regulations require definable boxes. The technology sees one stack.
There is a serious counterargument. Embedded processors are less capable than data center accelerators. Commercial robots have overwhelmingly legitimate civilian applications. Keeping Chinese developers dependent on CUDA could preserve American influence and provide Nvidia with continuing leverage. Broader restrictions could injure American suppliers while accelerating the same Chinese alternatives Washington hopes to hold back.
Those are not trivial concerns, and this is not an argument for reflexively banning every chip or software package that might find its way into a Chinese machine. Export controls can impose costs on the country applying them as well as the country being targeted. Poorly designed restrictions can surrender markets without preventing technological progress.
But dependence and acceleration are not opposites. China can depend on Nvidia today and still become a stronger robotics competitor because of that dependence. Nvidia’s platform can accelerate Chinese development during the formative years of the market even if Beijing replaces portions of that platform later.
The strategic calculation therefore cannot stop at who supplies today’s chip. It must consider who owns the factories, developer base, deployment experience and physical-world data created while that chip is being used.
The Technology No Longer Respects the Policy Boundary
Nvidia can simultaneously assume no Chinese data center compute revenue and view China as a major physical AI opportunity because government policy treats those businesses separately. The technology increasingly does not.
The United States is trying to deny China the infrastructure needed to build the most advanced AI while allowing an American company to help China embody AI in the machines that may define the next industrial era. That may comply perfectly with the letter of the current rules. It should still force a much deeper examination of whether those rules reflect the architecture of the technology they are intended to govern.
China is approaching robotics as an integrated national industry: bodies, intelligence, chips, software, manufacturing, deployment and data. The United States is still deciding which individual processors cross a technical threshold.
Washington is looking for the strategic technology inside the data center. Increasingly, that technology is walking out of the data center on two legs.
