从算力优先级解释 ChatGPT 端 Astra 为何被降配

Patrick C Toulme @PatrickToulme原推发布于 2026.09.30

在 X 上打开原推

I want to explain this a bit in terms of OpenAI’s serving compute priority. In order of what they care about 1. Internal RSI loops of Astra/Bel agents optimizing their workloads and auto research of next GPT 1.5. Rollouts for their RL samplers 2. Internal OpenAI users burning billions of tokens per day 3. API customers paying API money by the token 4. Consumers paying for subscriptions. At the end of the day, the most important customer is the RSI loop itself. Consumer money in the short term is a rounding error compared to potential RSI compounding.

这条是引用着另一条写的：「OpenAI 重开 200 美元 Pro 订阅并改额度算法」