OpenAI uses ~20% compute for inference during training

eliebakouch · x · 2026-08-19

Tweet discusses compute allocation during model training. OpenAI mentioned 20% of "inference compute" (likely including RL training) is used for training, evaluation, and internal usage. The author hopes other labs share similar monitoring and deployment data.

Related event: Frontier Labs Spend 10–20% of Compute on Training Monitoring and Inference(2 posts)→

Original post →

More from Infra

Infra channel →