A Chinese Startup Just Beat Nvidia at Its Own Robotics Game

A young company from Hangzhou has pulled off a surprise win in the race to build smarter robots. Spirit AI’s model, Spirit v1.6, scored 1,924 on the RoboArena benchmark, edging out Nvidia’s Cosmos3-Nano-Policy at 1,881. It’s the first time a Chinese model has topped this test, which measures how well AI controls physical machines in messy, real-world tasks—squeezing a ketchup bottle without bursting it, closing a drawer smoothly, stapling papers accurately.

Founded in 2024, Spirit AI has moved fast. It raised $73 million early, then closed a 1.5 billion yuan round (about $222 million) this week—its fourth funding round in three months. Investors include battery giant CATL and e-commerce leader JD.com, both of which already use Spirit’s Moz1 humanoid robot on their factory and warehouse floors. That real-world deployment gives the startup an edge that pure lab performance can’t match.

Still, one benchmark win doesn’t rewrite the industry. Nvidia built much of the infrastructure—Isaac, Cosmos, Omniverse—that makes these advances possible. It continues to release open models and simulation tools. But the gap closed faster than many expected. Spirit v1.6 outperformed Nvidia on a test Nvidia helped design. That stings, and it signals broader momentum in China’s robotics push, where government strategy, domestic chips, and tight corporate-research loops accelerate progress.

What matters now is whether Spirit AI can turn this benchmark success into reliable products at scale. Can its models transfer across different robot hardware? How quickly will competitors close the gap? Nvidia has resources and a massive installed base of GPUs. Chinese firms counter with speed, strategic investors who double as customers, and direct industrial partnerships. The race for physical intelligence is on, and the results are already surprising.