In essence, inference consumes electricity and computing power, so the inference results themselves are a form of asset.

90% of inference requests, due to high similarity, do not trigger the GPU but directly read existing inference results.
If each read incurs a fee much lower than that of independent inference, such as 0.0000001 ETH, it aligns well with the behavior pattern of rational economic agents like AI agents.
Such high-frequency, small-amount payments can only be completed on-chain.