At the 2026 World Robot Conference forum, held in Beijing from August 19-22, Zhang Jianzhong, founder, chairman, and CEO of Moore Threads, stated during his speech that embodied intelligence is approaching its own GPT moment. He noted that many are currently debating when this pivotal moment for embodied intelligence will actually arrive.

By reviewing the development trajectory of large language models, one can see a path and growth curve remarkably similar to that of today's embodied intelligence. Over the past five years, from the early GPT iterations to the official release of GPT-3.5, many have observed language models beginning to exhibit comprehension abilities, enabling multi-dimensional communication with humans and assisting in handling repetitive, complex tasks. When GPT-3.5 debuted, its model parameter count had already reached the hundred-billion scale. For both language and vision models, as parameter counts continue to rise, the emergence of intelligence is closely tied to the trinity of compute power, data, and algorithms.

Zhang pointed out that current embodied intelligence models remain relatively small in scale, still an order of magnitude behind the hundred-billion-parameter early language models. "Looking at training resources, many models today are trained using only dozens, or even just a few, GPU cards. From this perspective, we still have a way to go before reaching the true GPT moment for embodied intelligence," he said.

He further explained that the bottleneck is not confined to compute power alone. We also require vast datasets substantial enough to support models in learning richer knowledge of the physical world. Only through the coordinated advancement of compute, data, and algorithms can embodied intelligence models reach their GPT inflection point. "The primary reason the GPT moment hasn't arrived today is the lack of sufficient, high-quality, and large-scale data. Moreover, parameter models lack the massive compute clusters needed to process them," he added.

Based on lessons from language model development, Zhang offered a prediction: the next-generation world model for vision and embodied intelligence will likely see its GPT moment emerge on compute clusters comprising thousands, or even tens of thousands, of cards. He also called upon partners across the embodied intelligence industry chain to jointly build domestic large-scale training compute clusters, leveraging these resources to train the physical and world models essential for embodied intelligence.