Compute Curve

What it shows

Training a large AI model takes a very large amount of computing, and the amount has grown quickly. This chart plots every model in Epoch AI's public database that has a compute estimate, from 2010 on, and fits a trend line through them. In the data I retrieved on October 4, 2026, the trend line doubles about every 5.7 months.

Reading it

Each dot is a model placed at its publication date, and its height is the compute used to train it, measured in floating-point operations (FLOP) on a logarithmic scale where each gridline is a thousand times the one below. Color shows the kind of model (language, vision, multimodal, or other), the dashed line is the trend, and you can click a kind in the legend to isolate it, hover over or tap a dot to see its details, or switch to the table to see every model, largest first.

Where the data comes from

The data is Epoch AI's Notable AI Models, released under a Creative Commons Attribution 4.0 license. Epoch counts a model as notable if it set a state-of-the-art result, drew many citations, or came into wide use. I included only models that have a compute estimate and were published in 2010 or later, and for models listed under several domains I used the first. The trend is a least-squares line through the logarithm of compute plotted against date, and it describes the models in this dataset without predicting future growth.

What to keep in mind

• Epoch estimates compute from papers, technical reports, and hardware details, and its confidence in those estimates varies, so the largest recent models, whose developers publish little, have the least certain numbers.
• Only about half of the notable models have a compute estimate, and leaving out the rest may bias the chart toward better-documented work.
• "Notable" is a judgment, so the chart does not include every AI model.

To refresh the numbers, run npm run data:compute in the site's repository.