Physix FrontierPhysix Frontier · News Briefing Card (Leiphone · Aug 21, 2026)
Xinghaitu's Gao: Embodied AI's Endgame Is Selling Physical Tokens
KEY FACTS
- Gao Jiayang predicts that the business model for embodied intelligence will shift toward token-based sales charged per action.
- Future complete machines may become negative-margin carriers, with core value migrating to the embodied brain.
- Xinghaitu employs a unified autoregressive architecture to achieve zero-shot generalization and long-horizon tasks.
- Reinforcement learning has improved manipulation precision to the sub-millimeter level, achieving a 99.9% success rate.
- Generalizability is defined as training cost, with the goal of compressing new task training time to one hour.
KEY DATA
40%-60%Current gross margin on complete machines
20%Gross margin during solution subscription phase
99.9%Task success rate after RL
10 hoursPost-training time for new tasks
PHYSIX OBSERVATION
This perspective reveals a paradigm shift in embodied intelligence from "selling hardware" to "selling services." Hardware becomes a traffic entry point, while algorithmic capability emerges as the core of pricing power. For the industry, this means that simply stacking mechanical performance offers no way out; the competitive focus will completely shift to model generalization efficiency and data closed-loop capabilities. Whoever can enable robots to adapt to more scenarios at the lowest marginal cost will hold the commercial lifeline of the next-generation computing platform.
Source: Leiphone report