DeepInfer 开源引热议：155 万亿 token/天、每 GPU 小时约 38 美元

Jakub Przybylski @jacobbuilds_原推发布于 2026.10.21

在 X 上打开原推

deepseek in april 2025: we're too small to maintain a big open source inference engine deepseek today: here's the whole engine, apache 2.0 18 months. that's the turnaround the number i can't stop looking at in the deepinfer q&a: 155 trillion tokens a day ~$38 per GPU-hour, at half price renting a b200 costs $4-7/h. "AI is unprofitable" takes are going to have a rough week open weights = you get the recipe open engine = you get the kitchen the model decides what AI can do. the engine decides what it costs caveat before the hype: deepinfer runs one model (v4.1 flash), the cache system, PD split and load balancer aren't in it, and every benchmark is deepseek's own recommended setup is 128 GPUs. i'm only 128 short