An inference is essentially consuming electricity and computing power, so the inference result itself is a type of asset. 90% of inference requests, due to high similarity, won’t trigger the GPU but directly retrieve existing inference results. If each retrieval costs a very small fee, like 0.0000001 eth, it perfectly aligns with the behavior model of rational economic agents like AI agents. Such high-frequency, micro-payments can only be completed on-chain.