Estimated split between inference and training

Stacked Bar chart. Source: Patterson et al. (2022), The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink; approximately 60% inference/40% training in Google's measured ML workloads. More recent estimates for large AI developers commonly place inference around 60–80%, including higher splits such as 78/22 in some measurements. See Epoch AI's OpenAI compute-spend analysis and Meta's Wu et al. sustainability analysis.. Data table with 3 rows and 2 columns follows.

Source: Patterson et al. (2022), The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink; approximately 60% inference/40% training in Google's measured ML workloads. More recent estimates for large AI developers commonly place inference around 60–80%, including higher splits such as 78/22 in some measurements. See Epoch AI's OpenAI compute-spend analysis and Meta's Wu et al. sustainability analysis.