The inference efficiency of running GPT open-source models on Blackwell GPU has significantly improved within just one month—token processing capacity per unit cost increased by 33%. This breakthrough is thanks to the optimization work of the vLLM project and NVIDIA's hardware support, directly lowering the cost barrier for deploying large language models. For the Web3 application layer, this means the cost of AI inference infrastructure continues to decrease, further expanding the feasibility boundaries of on-chain AI applications and smart contracts.