← Back to the dashboard

How we track AI models and chips

What the dashboard shows, where the numbers come from, how current they are, and what they can’t tell you.

The short answer

Epoch AI, an independent research institute, keeps open records of notable AI models and of the chips used to build them. This dashboard counts and charts those records: how much computing power the largest models are trained with, who builds them, how they are released, and how fast AI chips are.

How we know

The data. Epoch marks a model as notable if it has over 5,000 citations, cost over $1 million to train (in 2023 dollars), has over a million monthly users, set a state-of-the-art result, or is of similar importance. For each model it records the release date, the organisation and its country, how it was released, and the training compute: the number of mathematical operations used to train it. Few companies publish that number, so Epoch often estimates it from the chips, time and other details known, and labels each estimate confident (within about 3 times), likely (within about 10 times) or speculative (within about 30 times).

Our calculations. We count notable models by release date, by where the organisation is based (US, China, or elsewhere and joint efforts) and by how the model was released. For growth, we fit a straight line to the logarithm of training compute for Epoch’s frontier models (the top 5 by compute when released) since 2018. For chips we chart the peak 16-bit performance makers state, grouped by whether the maker is American or Chinese.

How current each part is

What the dashboard can’t tell you

How capable models are. Training compute measures scale; benchmarks and real use measure ability, and they are not shown here.

How many chips are made or sold, or whether they are scarce. The chip chart shows stated peak speed, not supply.

Every model. Epoch’s notable list is a selection, and its criteria involve judgement.

Further reading

Credits

Data: Epoch AI, ‘Data on AI models’ and ‘Machine learning hardware’, published at epoch.ai under the Creative Commons Attribution 4.0 licence. We count and chart the published records ourselves. Epoch AI has not endorsed this site.

← Back to the dashboard