Most large-scale models are developed by US companies

The pace of compute-intensive model development has accelerated in recent years, with the United States and China leading the world by number of large-scale models developed. The European Union and United Kingdom are also substantial contributors.

The timeline for this graph begins in 2020, as the majority of large-scale models were developed after this.

Data

This data is from Epoch’s AI Models database, which contains data on models and their developers. The models are annotated with the country of origin of the developers.

Analysis

The dataset is filtered to models that have known or estimated training compute values greater than 10^23 FLOP. The models are sorted in descending order by release date, and a cumulative count of large-scale models is taken for each date from January 2020 onwards, summing by models within the same country or category as defined below.

Models with a single developer located in a single country are attributed to that country. Models with multiple developers all from the same country are attributed to that country. Models with multiple developers from different countries are classified as “Multinational”.

“Other” refers to models developed in any single country not listed on the legend.

Limitations

Some organizations are international, and have staff and facilities located across different countries. Models developed by these organizations are attributed to the country in which the organization is headquartered. This may underestimate the contributions of developers in countries that have not produced many large-scale models.