300名截肢患者，正在变成机器人的触觉训练师

The bottleneck of embodied AI is shifting from algorithms to data, especially high-quality manipulation data. In an experiment, Nvidia enclosed 8 AIs and 8 robots in an experimental environment; they autonomously learned to conduct experiments without any human intervention. This demonstrates the potential of autonomous data generation, but it remains far from general-purpose scenarios.

Meanwhile, 300 amputee patients are training robots with their own tactile data. This 'human data collection' model is not only costly—each patient requires long-term cooperation—but also involves complex ethical and privacy issues. Data scarcity has become the primary obstacle to the commercialization of embodied AI.

Every time a model iterates, the demands on computing power and chip architecture can be upended. Cuisi Technology, a new startup, secured its first-round funding led by Monolith and Qiying Tongchuang, with a mission to use AI throughout the entire chip design process, significantly shortening the development cycle of custom chips. This is a direct response to the trend of 'models driving chips.'

Li Auto, from an application perspective, proposed that embodied AI requires 'one chip, one brain, and a new paradigm.' Its self-developed chip is not a general-purpose GPU but a dedicated architecture for robot motion control and perception. The synergy between hardware and algorithms has become unprecedentedly tight.

Zhipu's stock price surged 18 times within the year, with a market cap approaching one trillion Hong Kong dollars, but revenue was only 700 million yuan. This script of 'valuation first, revenue lagging' has been repeatedly played out in the large model track. The market's frenzy is built on the belief that it will become infrastructure in the future, but the reality is that commercial scenarios have not yet been validated.

Comparing Nvidia's and Li Auto's approaches—the former building heavy-investment experimental platforms, the latter emphasizing hardware deployment—can Zhipu's asset-light model sustain its high valuation? Evidence is insufficient, but bubble signals are already apparent.

Nvidia's enclosed experiment demonstrates the possibility of 'autonomous research' in embodied AI, but Li Auto points out that true embodied AI needs to step out of the lab and deeply integrate with supply chains and application scenarios. Cuisi's AI-driven chip design is an attempt to bridge the 'last mile' from models to hardware.

The three paths—pure autonomy, hardware collaboration, and design automation—are not mutually exclusive but together form the ecological puzzle of embodied AI. However, each path faces significant constraints in data, cost, or time.

The heat around embodied AI is undeniable, but the three hard truths cannot be avoided: the ethical and economic costs of data collection, the lag of chip design behind model iteration, and the divergence between capital markets and business reality. Nvidia and Li Auto choose heavy investment, Zhipu chooses asset-light, and Cuisi chooses a middle ground.

In the short term, no single model can solve all problems simultaneously. What the industry may need is not faster pace but more precise focus—finding breakthroughs in data, hardware, and business models.