NVIDIA and Physical AI: omnipotent, omniscient and omnipresent?

I checked 16 humanoid makers’ announcements, spec pages and filings: 14 of them mention an NVIDIA component. The two exceptions: for Leju, I found nothing, and XPENG says its IRON runs three in-house Turing chips.

Unitree uses NVIDIA’s Jetson Thor, a computer module for robots, in its H2 Plus humanoid. That robot is the body of NVIDIA’s open research reference humanoid, announced on 1 June. Agility says NVIDIA will be the exclusive supplier of onboard compute for Digit 5 once it is generally available to customers. Early access is planned for the first half of 2027 and general availability by the end of 2027.

Control is harder to prove than compute. Tesla designs its own onboard chip; Elon Musk says AI5 will go into Optimus, but volume production remains a target. Tesla still trains on NVIDIA at data-centre scale: Data Center Dynamics reported that its Cortex cluster came online in Q4 2024 with 50,000 NVIDIA H100 GPUs, specialised chips used to train AI models. Figure’s deal with data-centre provider Nscale is worth $3.5 billion initially for computing capacity and could cover up to 100,000 of NVIDIA’s next-generation Vera Rubin GPUs. That compute is for training Figure’s Helix AI model in data centres. It is the same kind of infrastructure business that made NVIDIA central to generative AI.

The robot-side numbers are smaller. Research firm IDC counted about 25,000 humanoid shipments in the first half of 2026. Jetson Thor modules were priced at $2,999 apiece at launch (August 2025) for orders above 1,000. Even if every one of those robots carried a Thor, the total would be about $75 million. Many do not. So the money is in training compute; the chips in robots are a small slice. Then there is Washington. In June, SCMP reported, the Pentagon added Unitree to its list of companies it says support China’s military. Unitree was also barred from Defense Department contracts. That came a week after NVIDIA announced a reference humanoid built on Unitree’s H2 Plus.

Nothing in my research supports omniscience, the idea that NVIDIA sees everything these robots do. The closest case is Agility. Agility’s SEC-filed presentation says Agility is building NVIDIA’s Halos, its robot-safety software, into Digit’s system for detecting people, and using NVIDIA’s own “accredited” inspection lab to prepare for third-party safety certification. Agility does not name the accreditor in the filing, and certification is still to come.

How far does NVIDIA reach? In the ones I checked, it reaches as far as chips, software or training compute. It stops short elsewhere. The deepest it goes is at Agility. By Agility’s own account, NVIDIA will supply Digit 5’s computer and safety software, and NVIDIA’s lab helps prepare the robot for certification. The final sign-off is meant to come from outside, and until it does, the hardware, safety software and preparation for certification all lead back to the same company.

The question “who sells the most humanoid robots?” has several answers if we look at the first half of 2026.

Smart Analytics Global (SAG) counted 19,100 shipments, with AGIBOT at about 8,400 and Unitree at 5,900. Counterpoint Research counted more than 22,000, with AGIBOT at 9,700 and Unitree above 7,000. IDC put the total near 25,000. Frost & Sullivan counted 36,000 units sold. A shipment is a robot a maker has sent to a customer. The totals do not show where the robots end up or whether they do useful work, and the firms disagree about that too.

IDC says about 70% of first-half shipments went to display and validation: research, education, performances and government data collection. SAG says more than 70% went to industrial and commercial uses. The Frost & Sullivan report includes small, medium-sized and remote-controlled robots and, when it calculates revenue, leaves out sales between a maker and companies tied to it by shareholdings. SAG leaves out some models, including Leju’s Aelos and Roban lines, because they do not meet its definition of a functional humanoid robot. Each research firm defines its own version of what counts as a humanoid before it counts the robots.

Take Unitree, for example. SAG and Counterpoint both rank it second by shipment volume, behind AGIBOT. Unitree mostly sells half-size G1 robots to research and education buyers. Each relatively inexpensive machine counts as one unit alongside a full-size robot intended for factory work. In Frost & Sullivan’s first-half revenue ranking, Unitree comes third at ¥580 million, behind UBTECH at ¥620 million and AGIBOT at ¥610 million. UBTECH’s reported revenue has surpassed Unitree’s. But according to UBTECH’s interim results filed with HKEX, it reported a net loss of ¥311 million (about $43 million). Unitree, on the other hand, reported a net profit of ¥278.21 million for full-year 2025, although that includes its quadruped robot business, and the periods differ. None of these numbers tell the full story on their own.

Different rankings reward different things: unit volume rewards the number of machines counted, while revenue reflects what those machines bring in. A cheap research robot can move a company up the shipment leaderboard without establishing how many robots are doing productive work. So when investors and pitch decks repeat the word “leader”, they have picked the metric that works for them.

NVIDIA touches nearly every humanoid robot maker on the market, and nobody can agree on the actual numbers in that market. A company can be everywhere without being in control, and a market leader can change depending on what counts as a sale. In this space, the neatness of the headline often hides the messiness of the world it describes.