The "ChatGPT moment" for robots: a Chinese startup has named the exact date for a breakthrough in embodied AI
The robotics industry stands on the brink of a tectonic shift. Based on my estimates, drawn from the latest market data, by the end of 2027 we will witness the very "ChatGPT moment" for embodied artificial intelligence. This forecast was voiced by the chairman of China's ACE Robotics, Wang Xiaogang, and I am inclined to agree with him—the groundwork for this has matured.
The key catalysts for this breakthrough will be two factors: qualitatively new models for controlling robotic systems and the exponential growth of data arrays on real physical interaction. Without the second component, the first is fundamentally impossible.
The main bottleneck is not the "hardware," but the "brains"
Modern humanoids already demonstrate impressive physical capabilities—they run, dance, and execute complex programmed algorithms. However, their Achilles' heel is adaptation to unfamiliar environments and solving a wide range of non-standard tasks. ACE Robotics rightly points out that the problem lies not in the hardware, but in the AI models.
Embodied intelligence systems must process visual information in real time, interpret instructions, predict the consequences of their actions, and construct movement trajectories. It is precisely for this purpose that the startup is developing the Kairos model, which combines perception, multimodal understanding, and behavior planning. The results are impressive: in July, the model took first place in four specialized tests, and the open-source version Kairos-4B outperformed Nvidia Cosmos 3 and Ant Group's Lingbot in public benchmarks, despite its modest 4 billion parameters.
Data scarcity—a systemic industry problem
However, the most critical bottleneck is data. According to Wang's estimates, the entire industry has collected only about 100,000 hours of information on physical interaction with the world over recent years. This is catastrophically insufficient for training foundation models. The traditional approach of collecting data through teleoperation in exoskeletons is a dead end due to low speed and high cost.
ACE Robotics proposes a radically different path: the company equips ordinary workers on real production lines with lightweight sensors and records their natural actions. This allows data to be accumulated on an industrial scale—the startup plans to collect tens of millions of hours over two years.
Ambitious plans and strategic diversification
The company is already testing its solutions in unmanned retail, hotel services, and rapid-delivery warehouses, using humanoids from Unitree, AgiBot, and Fourier. The plans are bold: to deploy the technology in at least 1,000 stores within a year, and in 10,000 by the end of the second. Notably, for training its models, ACE uses not only Nvidia accelerators but also Chinese chips from Rhino Tech and Digua Robotics. This is a deliberate strategy to reduce costs and dependence on a single supplier.
At the same time, widespread commercial adoption, according to Wang's estimate, will take another four to five years after the "breakthrough moment." ACE Robotics, founded in July 2025 with support from Ant Group and SenseTime, has already raised over $100 million in the first half of 2026 and is preparing for an initial public offering.
My comment: The forecast looks realistic, especially against the backdrop of Morgan Stanley's recent upgrade of its forecast for Chinese humanoid shipments to 50,000 units in 2026. However, the main risk lies precisely in data collection: if ACE Robotics' approach works, the company will gain a unique competitive advantage, but if it does not, the entire market could face a delay of the "ChatGPT moment" by several years. We will keep an eye on developments.