Nvidia has introduced two humanoid robot training platforms within months, and both depend on mechanical hands made by Chinese companies. The first, unveiled May 31 with a Unitree body and Sharpa hands, is scheduled for delivery to research institutions including Stanford and ETH Zurich in late 2026. The second, called Agile One SN and shown October 21, pairs a body from Agile Robots with Wuji's Hand 2 and runs on Nvidia's Isaac software and Jetson Thor computing platform, with availability planned for early 2027, according to CNBC.
Wuji, founded in 2019 by Pan Yunzhe and backed by Ant Group, HSG and Future Capital, is headquartered in Shenzhen. Sharpa is based in Singapore with manufacturing R&D in Shanghai and was co-founded by David Li, who also runs the lidar company Hesai. One industry figure cited by CNBC, JoyIn founder Zhu Mingxuan, described Wuji's hands as lightweight and capable but prone to overheating, while Sharpa's are stronger and more expensive.
Why the hand is the hardest part to replace
Humanoid robotics has largely solved locomotion. Balancing, walking and basic manipulation are now demonstrated routinely by companies including Boston Dynamics, Figure and Unitree itself. The harder unsolved problem is dexterous manipulation: a hand with enough actuators, tendons and sensors to grip a tool, a fork or a fragile object with human-like control, packed into a housing the size of an actual hand.
That density requirement has pushed much of the supply chain toward Chinese manufacturers, who have scaled precision actuator and tendon-driven component production faster than counterparts elsewhere. Nvidia does not build humanoid hardware itself; it supplies the compute and software layer, Jetson Thor chips and Isaac simulation tools, and leaves body and hand sourcing to partners. That means its reference platforms function as a visible proxy for where the rest of the industry is already sourcing components, whether or not any individual robot maker advertises it.
A supply chain question that is now also a policy question
The timing sharpens the stakes. Washington has moved to treat robot imports as a national security concern, and tariff and import-control debates now explicitly target the robotics sector rather than leaving it under general electronics rules. A chipmaker whose own humanoid reference designs depend on Chinese-made hands is a harder case to regulate around than a single finished robot, because the dependency sits inside a subcomponent that most buyers never see broken out on a spec sheet.
The trade-off for Nvidia is speed versus exposure. Sourcing from Sharpa and Wuji lets it ship working reference systems to universities and developers faster than waiting for a domestic or allied hand supplier to mature. But it also means any future export restriction or tariff on Chinese robotic components could delay delivery schedules Nvidia has already set for late 2026 and early 2027. Rival chip and compute platforms pursuing humanoid robotics, including efforts tied to Tesla's Optimus program and various US-based manipulation startups, face the same underlying scarcity of non-Chinese dexterous hand manufacturers, which is why Nvidia's choice reads less as a one-off vendor decision and more as a signal of where the entire sector currently has no alternative.