Ubtech and Muxi form $14M JV to build embodied AI chips

China's two leading embodied AI companies — robot maker Ubtech and GPU designer Muxi — are joining forces to develop edge chips for physical AI, combining hardware and intelligence in a single 100 million yuan ($14 million) joint venture.

"By integrating Muxi's chip design expertise with Ubtech's robot deployment data, we can optimize the AI inference pipeline from silicon to servo motor," a person familiar with the partnership said, speaking on condition of anonymity because the terms are private.

The JV, Xixuan Chuangzhi Technology, registered in Wuxi on June 11 with 100 million yuan in capital. Ubtech and Muxi each hold 35.01% as co-largest shareholders, with auto parts maker Fenglong holding 4.7%. The venture targets edge-side inference chips for embodied intelligence — the class of AI that perceives and acts in the physical world, rather than processing data in cloud data centers.

The partnership reflects a strategic bet that China's 138 billion yuan state AI fund and its national embodied intelligence push through 2030 will create a domestic chip market large enough to support a homegrown alternative to Nvidia's Jetson platform, which currently powers most humanoid robot prototypes worldwide.

Why China's chip makers are racing into embodied AI

Embodied intelligence appeared in China's Government Work Report for the first time in 2025 and is a centerpiece of the 15th Five-Year Plan. The country installed 295,000 industrial robots in 2024 — more than the rest of the world combined — creating the factory-floor training data that humanoid developers depend on. Unitree, the Chinese robot maker that partnered with Nvidia on the Isaac GR00T humanoid, saw revenue grow 335% year over year in 2025.

Muxi, a GPU startup founded by former AMD engineers, brings chip design capability that Ubtech lacks. The company's chips are built on domestic foundry processes, insulating the JV from US export controls that have restricted Nvidia's advanced AI chips from entering China. Ubtech contributes the robot body and the deployment data — the real-world sensor inputs and motor commands that train embodied AI models.

The JV's focus on edge inference chips is strategic. Cloud-based AI requires constant data transmission, introducing latency that makes real-time robot control difficult. Edge chips process AI workloads locally on the robot, cutting response times from milliseconds to microseconds. Nvidia's Jetson Thor, the onboard chip for its Isaac GR00T platform, performs at 200 teraflops of FP8 compute — a benchmark the JV will need to match or exceed to compete.

What this means for Nvidia and the competitive landscape

Nvidia's strategy for physical AI is to act as the arms dealer — selling the compute platform while letting partners build the robot bodies. The company's Isaac GR00T platform uses a Unitree body with Nvidia's Blackwell GPU as the brain, and Huang has estimated the total addressable market for physical AI at $40 trillion.

The Ubtech-Muxi JV represents the Chinese alternative to that model: a vertically integrated domestic stack that controls both the chip and the robot, free from US export controls. If the JV can deliver edge chips competitive with Nvidia's Jetson series, it could capture a meaningful share of China's domestic humanoid robot market — which Beijing has designated a strategic priority.

For investors, the key question is whether the JV can achieve the performance density of Nvidia's 4nm Blackwell chips using domestic foundry nodes, which typically lag TSMC by one to two generations. Muxi did not disclose the process node or performance targets for the JV's first chip. Ubtech shares trade on the Hong Kong exchange; Muxi remains privately held.

This article is for informational purposes only and does not constitute investment advice.