Epoch AI's decomposition of 63 frontier AI models built since 2018 finds the computing power used to train them has grown at a median multiplicative rate of 4.17x per year, split roughly evenly across more hardware, longer training runs, and faster chips AI training compute grows 4.17x per year

Epoch AI analyzed 63 "frontier models" developed since 2018, defined as models not fine-tuned from a base model and in the top 10 by training compute at their release. The study decomposed the growth in computing power used to train these models into three factors: hardware quantity, training time, and hardware performance (FLOP/s). The median annual growth rate of 4.17x was found by fitting exponential trends to these factors.