Nvidia’s $10B Physical-AI Business Bets on a Full Stack—and an Open China Trade Lane
The company’s hardware, robot models, simulation tools and startup investments form a platform strategy that can serve competing manufacturers. Its ability to keep doing so depends partly on policy that remains unsettled.
Listen to this story
The audio brief
Story brief
3 key pointsNvidia is building physical AI as a cross-company platform rather than a single robotics product, combining Jetson computers, Thor, robot models, and simulated training environments. The business spans robotics, autonomous vehicles, and drones and is estimated at roughly $10 billion annually, but its projected tenfold growth is only a management target. Nvidia’s investments in robotics startups deepen the strategy,...
- 01
Jetson Orin Nano 2, released August 25, adds an entry-level product to Nvidia’s physical-AI hardware range.
- 02
GEAR, led by Jim Fan, is developing simulated environments and world models to generate embodied-AI training data.
- 03
Customers include AgiBot, Unitree, Fourier, Waymo, Tesla, and Chinese robotaxi fleets—spreading Nvidia across competing approaches.
Nvidia’s physical-AI business, spanning robots, cars and drones, is estimated to generate about $10 billion in annual revenue. Jensen Huang expects it to grow tenfold within a decade, a target that now rests on a stack extending from robot computers to simulated training worlds—and on continued access to a politically exposed China market.
The immediate buildout is broader than a chip sale. Nvidia’s physical-AI lineup includes Jetson robot computers, the high-end Thor platform, Jetson Orin Nano 2, GR00T robot models and Cosmos world models. Jetson Orin Nano 2 was released August 25, adding an entry-level product to a range that reaches into higher-end systems.
A platform assembled around the machine
The logic is to combine the computer inside a robot with software used to train it. Nvidia’s GEAR embodied-AI research group, led by Jim Fan, is pursuing simulated environments and world models to generate training data. That approach aims to let developers produce data for embodied AI, rather than rely only on data gathered from physical machines.
That makes the customer strategy unusually horizontal. Nvidia supplies technology to robotics companies including AgiBot, Unitree and Fourier, while its automotive customers include Waymo, Tesla and Chinese robotaxi fleets. The arrangement lets Nvidia sell into rival approaches to autonomy rather than making its own bet on which robot or vehicle maker prevails.
Capital follows the software and hardware
Nvidia is also investing in companies that could become customers of the stack. Its reported robotics investments reach across humanoid makers and foundation-model developers, adding a financing layer to the company’s role as a technology supplier.
- Nvidia invested in Figure AI’s 2024 financing.
- Skild AI raised $1.4 billion with Nvidia among the participants.
- Field AI reached a reported $2 billion valuation with Nvidia on its cap table, while Neura Robotics raised up to $1.4 billion with Nvidia participating.
The trade lane becomes the next decision
The platform’s reach creates its own vulnerability. Potential U.S. expansion of export restrictions to robot and automotive chips could limit Nvidia’s physical-AI strategy in China. Chinese manufacturers are already testing alternatives from Huawei, Black Sesame and Horizon Robotics, while FCC import bans on Chinese robots could reduce the U.S. market for machines that use Nvidia technology.
The unresolved question is whether Nvidia can preserve its position as the shared layer beneath a fragmented robotics market. Its technical breadth gives it multiple ways to participate in adoption. But that breadth is most valuable if robot makers, vehicle companies and developers can continue to buy and build across the same hardware and software ecosystem.
Editorial analysis
Our Read
The important signal is not a single robot product. It is Nvidia’s effort to occupy several control points at once: the computers in machines, the models that guide them, simulated environments that generate training data, and capital for startups building the category. Serving Waymo and Tesla illustrates the appeal of that position: Nvidia does not need to choose a vehicle winner. The next test is whether Washington broadens restrictions to robot and automotive chips, or whether Chinese makers accelerate their alternatives. Either outcome would pressure a strategy built around being a common supplier on both sides of the Pacific.
Citation desk / original work
Cite this
Citation desk / original work
Cite this
The important signal is not a single robot product.
/posts/nvidia-s-10b-physical-ai-business-bets-on-a-full-stack-and-an-open-china-trade-lane#finding-1
