Quantifies that frontier training spend grows 2.4x per year, with the largest models costing hundreds of millions of dollars. This frames who can afford to train frontier models and how fast costs rise.
Key takeaways · AI-distilled
Epoch estimates training compute costs by allocating cluster costs over the training period. Its figures describe the training run's share of infrastructure costs, rather than the price of buying the whole cluster.
Epoch estimates that GPUs account for about half of large-model training spending. Other hardware and energy account for the rest, so GPU spending alone understates its cost measure.
Epoch's historical estimate puts large-scale model training costs on a 2.4-fold annual growth trend. The accompanying table reports costs in 2023 dollars and identifies the dataset update as November 24, 2025.
