Epoch AI

Epoch AI is a non-profit research institute that uses data to study how AI is developing. It began in 2021 as a group of volunteers collecting data on machine-learning models, and became a formal organisation after their 2022 paper analysing trends in the compute (the amount of calculation done by computers) used to train AI drew attention. Its co-founders include Jaime Sevilla, its CEO as of October 2026, Tamay Besiroglu and Marius Hobbhahn. US tax records list the legal entity, Epoch Artificial Intelligence, Inc., at an address in San Francisco, California, and since early 2025 it has operated independently as a 501(c)(3) non-profit, a public-benefit organisation under US tax law.

It gathers the factors that drive AI progress, such as chips, training compute, investment and data centres, into databases that it publishes. As of October 2026, its AI models database lists more than 3,700 models. Its 2022 paper found that training compute had grown 10 billion times since 2010. In June 2024 it estimated that usable public text written by humans amounts to about 300 trillion tokens (the units in which AI reads text), which could be used up between 2026 and 2032 if trends continue. FrontierMath, a maths test it released in November 2024, is made of unpublished problems that take mathematicians hours or days to solve; at the time the best AI solved under 2%. However, in January 2025 it was criticised for not properly telling the problem writers that OpenAI had commissioned and funded the work, and Epoch AI admitted the mistake.

In October 2025 it launched the Epoch Capabilities Index (ECI), which combines scores from many benchmarks (tests of AI performance) into one. In September 2026 it released a test of 68 still-unsolved Erdős problems; in the first round only GPT-6 Astra solved any, getting 2 (3%). In October 2026 it gave AI models real tasks from its own work and reported that AI cannot yet do that work on its own. It says it raised $10.3 million in 2025 (company-reported). It runs mainly on donations and grants but also takes commissioned work from AI companies and government bodies, and it publishes lists of its funders and collaborators.