Spirit AI, a robotics start-up headquartered in Hangzhou, Zhejiang province, China, has become the first Chinese company to lead the RoboArena global leaderboard for embodied intelligence models, surpassing Nvidia just two days after the US chip giant released its latest physical AI foundation model. The milestone, announced on Wednesday, 4 June 2026, signals an intensifying China-US contest over the 'brains' that will power next-generation robots.

The benchmark breakthrough

Spirit v1.6, the company's foundation model for embodied intelligence, recorded a score of 1,924 on the RoboArena benchmark — edging out Nvidia's Cosmos3-Nano-Policy, which placed second with a score of 1,881. Third place went to DreamZero, another Nvidia project unveiled in February, with a score of 1,763.

The RoboArena benchmark evaluates how effectively generalist robot policies translate to real-world physical actions. It was co-developed by Nvidia alongside elite institutions including Stanford University and the University of California, Berkeley.

Why it matters

The timing is striking: Nvidia launched its Cosmos 3 model — designed to help physical AI 'think before it acts' — just two days before Spirit AI's announcement. For a Chinese start-up to immediately displace a Nvidia model atop a benchmark co-designed by Nvidia itself underscores how rapidly the competitive landscape in physical AI is shifting.

Embodied intelligence — the capacity for AI systems to perceive, reason, and act in the physical world — is widely regarded as the next critical frontier after large language models. Whoever leads in foundation models for robotics stands to shape the hardware-software stack that factories, logistics networks, and consumer robotics will depend on for decades.

The competitive backdrop

The rivalry extends well beyond Spirit AI and Nvidia. Chinese players including AgiBot and XYZ Embodied AI are also active in the embodied intelligence space, while Nvidia has been deepening partnerships across Asia. On Monday, 2 June 2026, Nvidia announced collaborations with Unitree Robotics of China and Sharpa, a Singaporean robotic-hand pioneer, further illustrating that the physical AI supply chain is deeply intertwined across geopolitical lines.

The pattern mirrors earlier contests in large language models and AI chips: Chinese firms absorbing open benchmarks and rapidly iterating to reach or exceed Western performance thresholds, even as export controls limit access to the most advanced semiconductors.

What's next

The leaderboard position is a snapshot, not a verdict — Nvidia and other US labs are expected to respond with updated models. Investors and enterprise buyers will be watching whether Spirit v1.6's benchmark lead translates into real-world deployment wins in manufacturing and logistics. As physical AI moves from research labs to factory floors, the model layer is becoming as strategically contested as the chip layer.