NVIDIA sells no robots, and 110 robot makers build on its stack

A keynote stage built for other companies' robots

Fact At GTC in March 2026, NVIDIA did not launch a robot. It listed partners: ABB Robotics, FANUC, Yaskawa, KUKA and Universal Robots on the industrial side, and AGIBOT, Agility, Figure, 1X, Boston Dynamics, Skild AI and NEURA Robotics among humanoid and robot-brain developers. The Robot Report counted 110 companies in the announcement, and NVIDIA put the combined installed base of FANUC, ABB, Yaskawa and KUKA at more than 2 million robots.6

Fact The same announcement covered healthcare (CMR Surgical, Johnson & Johnson MedTech, Medtronic), logistics and manufacturing (Foxconn, KION, GXO, Teradyne Robotics) and four clouds, including Microsoft Azure and Alibaba Cloud. Hugging Face said it would fold Isaac and GR00T into its LeRobot open-source framework.6

Fact NVIDIA sized the funnel behind the slide in the same announcement: about 2 million robotics developers on its platforms, more than 40,000 members in its Inception startup program and 13 million AI builders on Hugging Face. NVIDIA's Rev Lebaredian put investment in humanoid robots at $20 billion.6

Our read For anyone selling robots, this is the main fact about NVIDIA. It has made itself the shared supplier for companies that compete with each other, and it uses its keynotes to show that. A humanoid startup that appears on that slide gets borrowed credibility with buyers. A startup that is missing has to explain why.

Open models as the top of the funnel

Fact NVIDIA announced Isaac GR00T N1 on March 18, 2025, calling it the first open, fully customizable foundation model for humanoid reasoning and skills. It pairs a slower vision-language model that plans with a faster action model that produces movement. Agility Robotics, Boston Dynamics, Mentee Robotics and NEURA Robotics had early access, and 1X showed a humanoid running a post-trained GR00T N1 policy on the keynote stage.1

Fact The release leaned on synthetic data. NVIDIA said it generated 780,000 synthetic trajectories in 11 hours, equal to 6,500 hours of human demonstrations, and that mixing synthetic with real data improved GR00T N1's performance by 40% over real data alone.1

Fact Two months later, at Computex on May 18, 2025, NVIDIA released GR00T N1.5 and a GR00T-Dreams blueprint for generating synthetic motion data. It said NVIDIA Research built N1.5 in 36 hours with the blueprint, against nearly three months of manual data collection. AeiRobot, Foxlink, Lightwheel and NEURA Robotics were named as early adopters of the GR00T N models, and Foxconn and Foxlink were using the GR00T-Mimic blueprint for manipulation data.8

Fact By GTC 2026 the model line had moved on. GR00T N1.7 entered early access with commercial licensing, and NVIDIA previewed GR00T N2 for release by the end of 2026, claiming it succeeds at new tasks in new environments more than twice as often as leading vision-language-action models.6

Our read The sequence is the business model. The first version was free and open so labs would try it. Commercial licensing arrived once robot makers had post-trained on it and wanted to ship. Each version also assumes NVIDIA hardware for training in the data center and inference on the robot, so the model is a reason to buy the chips around it.

Jetson Thor puts a price on the robot brain

Fact Jetson AGX Thor went on general sale on August 25, 2025, with the developer kit starting at $3,499. NVIDIA rated it at up to 2,070 FP4 teraflops with 128GB of memory in a 130-watt envelope, and said it offers up to 7.5 times the AI compute of Jetson Orin with 3.5 times the energy efficiency. Production T5000 modules are sold through distribution partners.2

Fact NVIDIA named Agility Robotics, Amazon Robotics, Boston Dynamics, Caterpillar, Figure, Hexagon, Medtronic and Meta as early adopters, with 1X, John Deere, OpenAI and Physical Intelligence evaluating the chip. It said more than 2 million developers use its robotics stack and that Jetson Orin had reached more than 7,000 customers.2

Our read A published price matters more than it looks. A robotics team can budget for compute before talking to NVIDIA sales, and an engineer can buy a kit on a card. That is the same bottoms-up path developer tools use: the engineer picks the platform, and procurement inherits the decision when module volumes arrive.

Our read The adopter list also shows who NVIDIA refuses to choose between. Amazon Robotics, which builds robots for Amazon's own warehouses, sits next to Agility, which sells humanoids into logistics. Meta and OpenAI appear on the same list as Figure. NVIDIA's neutrality is the product, and it holds only as long as NVIDIA does not ship a branded robot of its own.

Simulation and world models sit upstream of every deployment

Fact NVIDIA launched Cosmos at CES on January 6, 2025, as a platform of world foundation models released under its open model license. It said the Cosmos data pipeline could process and label 20 million hours of video in 14 days on Blackwell, against more than three years on a CPU-only pipeline. Early adopters included 1X, Agility, Figure AI, Skild AI, XPENG, Waabi and Uber.3

Fact At the same GTC where it introduced GR00T N1, NVIDIA announced Newton, an open-source physics engine developed with Google DeepMind and Disney Research. Isaac Lab 3.0, shown in early access at GTC 2026, is built on Newton 1.0 and PhysX, and NVIDIA launched Cosmos 3 in its fiscal second quarter of 2027, describing it as a fully open frontier model for physical AI.165

Our read This is where NVIDIA's position is strongest. A robot that is trained in Isaac Lab, tested against Cosmos-generated scenes and run on Thor carries NVIDIA in its development history, not just its bill of materials. Switching the chip is an engineering project. Switching the simulation, data and training pipeline means redoing the work that got the robot to a pilot.

Our read What NVIDIA cannot generate is the last mile of data from a real site: the specific totes, lighting, floor wear and exception cases of one customer's building. That data comes from deployments, and deployments belong to the robot maker. It is the one input to NVIDIA's own pipeline that its partners control.

Robotics is still small in NVIDIA's accounts

Fact NVIDIA reports robotics inside its Automotive segment. In the fiscal year ended January 25, 2026, that segment brought in a record $2.3 billion, up 39%, out of total revenue of $215.9 billion, according to its February 25, 2026 earnings release. Data Center revenue was $193.7 billion.4

Fact In its fiscal 2027 second-quarter release on August 26, 2026, NVIDIA reported revenue of $96.2 billion and Edge Computing revenue of $7.2 billion, and did not break out a separate robotics or automotive figure. The same release announced an Isaac GR00T reference humanoid design built on Jetson Thor and Halos for Robotics, a safety system for physical AI.5

Fact NVIDIA also invests in the companies it supplies. Figure listed NVIDIA among the investors in its Series C, which closed above $1 billion at a $39 billion post-money valuation on September 16, 2025.7

Our read Owning a stake in Figure while selling Thor to Agility and Boston Dynamics is a small conflict NVIDIA can live with, because its equity checks are tiny next to its chip revenue. For the startups, an NVIDIA check works more as a signal to other investors and buyers than as capital.

Our read On those numbers, the whole Automotive segment, robotics included, was about 1% of revenue in fiscal 2026, and NVIDIA can afford to give models away to grow it. The payoff is upstream: every humanoid team that trains on synthetic data rents GPUs. A reference humanoid design also lowers the cost of entry for new hardware makers, which adds more customers for Thor and puts pressure on hardware margins for the companies already selling.

What this means for companies that sell robots

Our read NVIDIA's GTM can be copied in parts. It gives developers something free that leads to something paid. It publishes prices so engineers can start without a sales call. It puts partners in its keynotes so their success reads as its own. None of this requires being a chip company. A robot maker can open-source a simulation model of its robot, publish a dev-kit price, and showcase integrators who build on it.

Our read The risk runs the other way too. When the model, simulator, data pipeline and compute all come from one supplier, the robot maker's own differentiation shrinks to hardware, deployment and service. Companies on NVIDIA's partner slides should treat that as a reason to own the customer relationship, the site data and the service contract, the parts NVIDIA does not sell.