Largest training runs

Epoch AI maintains a running database of notable AI models and estimates the computing power — measured in floating-point operations, or FLOP — used to train each one. This chart shows the current top 10 by that estimate, refreshed daily from Epoch’s published dataset. Training compute is a proxy for the scale of a model’s development, not for its capability or benchmark performance; a model trained with less compute can still outperform a larger one through better data or architecture. At the time of writing (September 2026), all ten models on the list were trained by US-based organisations — that reflects where large-scale frontier training has concentrated to date, not a gap in Epoch’s coverage. Figures are Epoch’s own estimates, drawn from public disclosures, papers and their own analysis where a vendor has not stated a number directly; treat them as informed estimates rather than official specifications.