Frontier training compute
This chart follows the running maximum estimated training compute used by notable AI models. It relies on Epoch AI’s open model database and displays base-10 logarithmic FLOP, where each vertical unit represents a tenfold increase; recent and undisclosed model estimates remain uncertain and can be revised.
What does it show?
Frontier-model training compute continues to grow exponentially rather than linearly.
Methodology
Running maximum of estimated training compute across models in Epoch AI's AI Models database, by publication month, shown since 2010 as the base-10 logarithm of FLOP on a linear axis (each unit is a 10× increase; a straight line is exponential growth). Compute figures are Epoch's published estimates and carry uncertainty, especially for unreleased details. Data © Epoch AI, 'Data on AI Models', epoch.ai, CC BY 4.0; the dataset is updated as new models are added, so recent months can revise.