Where frontier AI compute is concentrated
This chart shows how estimated frontier-model training compute is distributed among the United States, China, and the rest of the world. It uses Epoch AI’s notable-model database and allocates multinational models across associated countries; it describes the selected database, not all national AI investment or capability.
What does it show?
The United States and China now account for nearly all estimated frontier-model training compute in the selected database.
Methodology
For each publication year, estimated training compute is summed across models in Epoch AI's notable-model database and shown as a 100% geographic share: United States, China, and all other countries. Models associated with multiple distinct countries are allocated equally across them so their compute is counted once; repeated country names within a row are deduplicated. Only models with a numeric training-compute estimate and organization-country field are included. This is the composition of Epoch's selected notable-model dataset, not a census of national AI investment. Estimates and the partial current year can revise. Fetched from https://epoch.ai/data/notable_ai_models.csv.