Physical Token Economics: Cheaper Next Capabilities Are Key to Robot Scaling

At the WRC 2026 'Physical AI Leaders Forum' held in Beijing on August 19, Jia Kui, founder of Cross-dimensional Intelligence, delivered a keynote titled 'The Economics of Physical Tokens,' proposing the concept of converting the production capacity of robotic intelligence into a measurable object. Physical Token is defined as 'the unit of reliable physical intelligence that a physical agent can produce,' and its economics focuses not on pricing individual actions but on the human effort, data, and compute required to produce a unit of reliable physical intelligence.

This concept arises from the practical contradiction facing embodied intelligence: the actions, contacts, forces, and failure experiences robots need are mostly acquired through real-world interaction, and each collection and trial-and-error incurs equipment, compute, and engineering labor costs. If learning each new skill requires roughly the same investment, the increase in robot capabilities is merely an increase in the number of projects, failing to achieve economies of scale.

Cross-dimensional attributes costs to three aspects: whether infrastructure can be transformed from repetitive human investment into long-term tool assets; whether a small amount of expensive real data can yield more effective training experience; and whether data and models can be reused across embodiments, scenarios, and tasks. These investments ultimately point to a single metric: the Marginal Cost of New Capability.

Only when the additional investment required for new skills, scenarios, or embodiments decreases with accumulation does Physical AI achieve economies of scale. Therefore, measuring whether general physical intelligence is viable cannot rely solely on the success rate of individual demos; it must also consider whether the capability boundary can expand continuously with a downward cost curve.

Cross-dimensional summarizes its technology stack into four stages: Experience, Leverage, Condensation, and Expansion. Experience emphasizes unified data representation and pretraining, making each learning a reusable capability asset; Cross-dimensional achieves multi-source data alignment through unified coordinate systems and Dexterity-BEV.

Leverage uses Real2Sim and generative simulation to transfer trial-and-error to virtual environments, leveraging a small amount of real data to generate more training experience. According to Cross-dimensional, their self-developed DexVerse engine can generate high-fidelity simulation environments, and the GS-World engine reduces simulation task environment setup time from days to just over 10 minutes. Condensation addresses how to convert general capabilities into reliable skills with less task-specific data; Expansion uses a closed loop of real-world deployment feedback, allowing deployment data to flow back into models and reduce the cost of the next deployment.

At WRC, Cross-dimensional showcased three types of tasks: phone assembly quality inspection, autonomous checkout scanning in supermarkets, and continuous operations for ice cream and popcorn, drawn from industrial manufacturing, retail, and commercial services. These demos point to the same question: can a unified data, model, and simulation system enter different scenarios with low incremental cost? But exhibition demos cannot directly equate to scaling validation; the key is whether the incremental investment truly decreases when transferring across scenarios.

Regarding Zero-shot, Cross-dimensional believes it may not be a prerequisite for scaling, but rather a result gradually approached after sustained economies of scale. When the marginal cost of acquiring new capabilities is low enough, Zero-shot transitions from a model goal to an industry state. This also explains why commercialization and model research cannot be separated: commercialization provides scarce real-world data, and enhanced models reduce the cost of the next scenario, forming a closed loop.